data-science-and-ml

[gritty vessel]

https://pastebin.com/BhSJLuPM

[lofty thorn]

can anyone tell me how much time it will take if i have to complete this project?

[dusty forge]

I might have missed it along the way, but the 'error', the distance between X(ᶦ) and the regression, does it have it's own letter?(ᶦ)

[dim crane]

In scikit-learn, linear regression models are typically trained using the least squares method by default. When you train a LinearRegression model in scikit-learn, it automatically fits the model using the least squares method to minimize the residual sum of squares (RSS). Afterwards, you can calculate and check this error metric: https://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_absolute_error.html

[dim crane]

Sorry, forgot to respond with my answer attached, just check my upper message

[gritty vessel]

wow dude

[dusty forge]

I learned about MAE, but there is not a letter designated to it I guess? What I mean is, when a formula has a SUM function, instead of typing SUM, you can type ∑

[dim crane]

The letter for the error is epsilon

So if the quoefficient has letter A and the intercept has letter B

They would have epsilon(A) or epsilon(B)

And as for MAE, it would be epsilon(Y)

[dusty forge]

ahhh yes, I knew I saw something somewhere haha thank you

[dim crane]

Np

[versed pilot]

The NIST handbook is a great resource

[gritty vessel]

yo please look into this would really appretiate your help #1208723301967470673

[mint palm]

In my job ,
I have worked on data asset management system with java, mongodb and open source framework: Nuxeo .
I have also worked on event streaming plattform using kafka, mongodb, java

how to write this in resume so that it looks better. Preferably in one line

[stable grove]

I want to make a database for training an AI model to predict the value of a house

Im thinking of doing the following

Year the house was built ---date
Age of the house --integer
safety scores -- double
Total square footage of the house --Double
Number of bedrooms --integer
Number of bathrooms --integer
Lot size --double
Number of floors --integer
Presence of a garage (yes/no) and its size(double) ---unsure
Presence of a basement or attic(yes/no)
Presence of amenities(Text) as theyre could be multilpe ammenities like a fireplace, swimming pool

[dusty forge]

Based on the tutorials I've done so far, it seems that features (columns) with format other than numeric should be avoided. My knowledge is limited so this might not be true at all. But in this case you could replace the garage to 0 for no garage, 1 for small, 2 for large (this way you combine the boolean with the double), the attic with 1 and 0, and the amenities with numbers. The amenities might cause issues because the model could interpret the number as a weight, for example 1 is fireplace is more valuable than 5 is swimming pool because 1 is earlier than 5, or 5 is more valuable because it's a larger number. I had a similar situation in which I wanted to use simple numbers as my unique identifier, and someone explained that I need to convert it into vectors so the model doesn't get mixed up.

Actually for the garage it can be either a 0 or anything larger than 0

[agile cobalt]

in general models literally only work with numbers, so yes you have to encode text and potentially other categorical features

it can either be Ordinal (one column with numerical values for sequential categories like no = 0, small = 1, large = 2 for the garage) or One-Hot (one column for each category, when they are not necessarily downwards compatible with each other, e.g. a boolean column for has fireplace True/False, a boolean column for has swimming pool True/False)

[stable grove]

thank you

[agile cobalt]

garage size could be just one double though, with no feature for "has garage", no = 0, yes = >0

also we call them "floats" in python rather than doubles

[dusty forge]

I wonder what will happen when text is fed as a feature into the model, will that simply not run the code, or worse, it runs but skewes the results?

[agile cobalt]

usually it should just error

[dusty forge]

ok great, I rather have that then results that I can't trust

[agile cobalt]

the models literally can only operate on numbers, if you had some very weird setup it might be possible for something to automatically cast them as numbers by using something like ord(), but that is not the sort of thing that just happens automatically

[dusty forge]

@stable grove are you by any chance doing the Coursera course? Because Andrew indeed introduced us to the housing example 😉

[stable grove]

No, its a project idea I came up with.

[stable grove]

could i convert all my text fields to ASCII values?

with a delimeter?

69,100,101,110,118,97,108,101 --Edenvale

[agile cobalt]

technically it is possible but it is a pretty bad way of converting text to numbers your model can use - you want for the data you give to be as meaningful as possible, and ascii is pretty arbitrary

I would strongly recommend just using one-hot encoding first to get used to it before trying anything fancier

btw, dates also have to be encoded
technically you could use a unix timestamp, or just separate year/month/day into their own numerical columns (or do more operations like iso calendar week, summer/winter/fall/spring etc.)

but for this case, you might as well not include it whatsoever considering it would be 'overlapping' with age

[stable grove]

thank you for the help

[dim crane]

Hello, I'm wondering in what cases is it needed to use the following techniques on my RandomForestClassifier script: MinMaxScaler, CalibratedClassifierCV and GridSearchCV (without it slowing too much the load time of the script).

My first guess is that we don't need MinMaxScaler; we just need to use GridSearchCV once for each dataset and use CalibratedClassifierCV to check where the log loss is lower (either on not calibrated, sigmoid or isotonic)

Also if you have any more suggestions to improve the results of my predict_proba output, I'll be really thankful

[dusty forge]

will 20240218 work as an index, for example a transactional datestamp?

by the way, my math is rusty as hell super embarassing, but what does this sign means? (tried to find it in the Greek alphabet, wasn't there?)

[agile cobalt]

it could work, specially for a tree based model, but it would be pretty weird for a regression task given the way it distort the distances between things like 20231230, 20231231 and 20240000
like... you might as well just use a timestamp if you want to turn it into a single number

[dusty forge]

yeah oke, I guess Time Series is a bit more on the advanced side of things, will come to that eventually 😄

[agile cobalt]

in which context?

[dusty forge]

This one

[wooden sail]

this usually denotes either partial derivative or the boundary of a set

yeah, there it's a partial derivative

[dusty forge]

Ah alright, does this symbol have a name?

[wooden sail]

i'm actually not sure, looks like some sort of calligraphic d

in most places you'll find it as \partial

[versed pilot]

It must be a very widely used example, we did something similar in Matt Harrison's Statistics course

[wooden sail]

.latex $\partial$ $\del$ $\delta$ $\nabla$

[tidal bough]

!charinfo ∂

[arctic wedge]

Character Info

\u2202 : PARTIAL DIFFERENTIAL - ∂

[tidal bough]

^

[wooden sail]

i think del is the impostor

.latex $del$

[wooden sail]

oof

[dusty forge]

Andrew's are the housing prices in California, where is Matt's? 😄

Ok I made a note on that partial derivative/ partial differential, thank you.

[agile cobalt]

I'll go back to another one of Andrew's courses I'm watching now lol
edit; 2.5/10 wouldn't recommend that mini course
ping me if you have a follow up question on what I said earlier

[dusty forge]

I absolutely love his course and the way he teaches, by drawing, which works really well for me. The last time I had math, algebra, calculus was ages ago, always loved it, but didn't keep it up since there wasn't anything that required me to do so. Now starting with ML, I'm slowly getting that feeling back of joy seeing and understanding formulas. I'm sure that at some point I'll want to bang my head into the wall, perhaps when I arrive at matrices. Seems to be a rough part based on some Reddit posts I saw.

[versed pilot]

https://github.com/LinkedInLearning/python-statistics-essential-training-4433355 that's the repo

Using the Ames, Iowa Housing Data https://jse.amstat.org/v19n3/decock.pdf

[dusty forge]

Oh cool. Well let us know if it all worked out with the database and properly feeding it into the model.

[final kiln]

I'm gonna try this thing https://www.prefect.io/opensource

Works extremely well locally

But it says it can automate the infra, which would be really really good rn

[dusty forge]

You read my mind, I was looking at platforms to make ML pipelines instead of just the experiment.

[final kiln]

It's really good, worrying about the infra has been slowing me down a lot, I think this lets you get a pipeline going, test it locally and then push it to a serverless compute worker

I already have that, but coded manually by me, which is cool but also a lot to manage for one person

[molten acorn]

What do y'all think of this roadmap?
https://roadmap.sh/ai-data-scientist

[final kiln]

Looks nice, I'd try to do some coding in parallel as you learn the math

A lot of math can be learned much better with the help of python, where you can just do your own experiments

Probability in particular can get super abstract if you go into the details of prob spaces and whatnot, code can keep you grounded, which is important when self studying

[molten acorn]

I'm in an odd place because I have over a decade of programming experience, in particular game dev, so I have exposure to linear algebra, I wonder if there's a topic or project that can fast forward my "onboarding" into data/ai

[final kiln]

Linear Algebra and calculus already put you in a good spot for ML imo, there's tools in stats that are important but if you know the others you can quickly pickup stats as you go

[molten acorn]

Or if there's any value to learning through training LLMs

[molten acorn]

I just got a job in data, starting out more on the "analyses" side and getting introduced to more stats and what not

[final kiln]

Training LLM will get you in direct contact with the thing and it will make you make all sort of mistakes

[molten acorn]

But when it comes to AI, I have a lot of unknown unknowns and how they related to my current skillset

[final kiln]

Which is good experience, but also probly the hard route

If you know math I'd go directly for doing a project that interests you and use it to learn

[molten acorn]

Ok, I view training LLMs as a top-down approach to learning AI, my other gut instinct is that learning neural networks will be a bottom-up approach to learning AI. Do you think this is accurate?

Does it make sense what I'm saying?

[final kiln]

I think it's good to first have an understanding of the basics, like the gradient descent algorithm, mini batching, motivation for why NNs work in the first place

[molten acorn]

I have a book

[final kiln]

Like a lot of it after the theory will be a sort of dark art where you kinda just experiment til you get it right

[molten acorn]

https://natureofcode.com/

[final kiln]

And the way you do it, the tools you use, etc, it's all important stuff you only learn by making mistakes

[molten acorn]

Since my previous background was in games and graphics, I'm way more inclined to learn from this book, but only the latter chapters are dedicated to the topic, but maybe it's a good enough intro

I also like that it would bridge concepts for me

[final kiln]

Looks really cool

I also come from a sort of graphics-y background

[molten acorn]

I have one project in mind, but because I have a lot of unknown unknowns, I have no idea how feasible by just me

[final kiln]

I used a lot of computer graphics concepts to do physics simulations

[molten acorn]

Oh dude, you're my new best friend

[final kiln]

Hehe, I just had an interview where the first questions were all computer graphics

[molten acorn]

Nice, which graphic libs are you familiar with?

[final kiln]

For MLE

[final kiln]

Oh, none, I've implemented stuff from scratch and came up with new algos

Ig I've used SDL once to do a ray tracer engine

[molten acorn]

Nice, SDL is nice to use

So here's my first project idea; I like watching basketball, but I don't have time to sit through a full televised game. It would be great if I could get AI to take a game recording and cut out all the boring parts like free throws and players walking the ball up the court

[final kiln]

Like usually the data I got back was not a visual output per say, it was physics data of energy deposition, prob distributions and whatnot

[final kiln]

So, I reckon you're gonna be using a pret trained model for that, unless you got a lot of GPU with you and a lot of data

No wait, you can build a classifier right

And apply it to every frame

[molten acorn]

Could I just set something up on together.ai?

[final kiln]

Is basketball, or is not basketball

[final kiln]

They seem to be gen AI models

[molten acorn]

Ok, what do I need? An M3 macbook?

[final kiln]

But I'd recommend getting an Encoder CNN with two output classes

[molten acorn]

Will the new laptop core ultra-9 work?

[final kiln]

You need GPU, and I'd avoid ARM

[molten acorn]

How many gigs of VRAM?

[final kiln]

Idk how the ARM support is rn, but last time it wasn't very good, might've changed

[molten acorn]

I have a 3070 that's not doing anything

But realistically, how long of a project is this if done on nights and weekends?

[final kiln]

The more the better, I've been able to train 100M param language models on 16Gb, but it's frustrating and text data is more forgiving

[final kiln]

I have no idea tbh, the thing that usually decides success is data quality

And compute power

I thought sentiment analysis was gonna be easy, been like a month and ton of infra

[molten acorn]

I'd have to really research into how to go about doing this. Somehow I have to train it to identify what "boring" moments are

Using audio and video

[final kiln]

I'd go for image data

Like, videos are just images

And multimodal modals are very likely harder to train

[molten acorn]

Right

[final kiln]

I assume that you want to cut out ads and such

[molten acorn]

Exactly

[final kiln]

Then image should workout fairly well

[molten acorn]

a 2 and a half hour recording should turn into a 40-45 min thing

[final kiln]

Like, if you require temporal understanding, it's harder

What I mean is like, if you want to detect the times when someone scores

It might be harder than just detect the ad bits

You might want to consider starting with a simpler classification problem and build up

I also highly recommend using this for pipelining https://www.prefect.io/opensource

Even locally works wonders

[molten acorn]

Ok, that's a good suggestion

[final kiln]

https://mlflow.org/

And also this, it's indispensable

First thing you gotta do is find a dataset and evaluate it. The larger the better. You then split it into three:

• train
• validate
• test

Train dataset is the one that the network learns from, the network will perform forward pass, backwards pass and then gradient optimization

The test dataset contains data that the network has not seen during training

This is crucial because neural networks are functions with high capacity for memorization. So you want your network to perform generalization, or perhaps even better, you can call it compression.

To test that your network is not memorizing that dataset you test it against another dataset it had no chance of memorizing from

[molten acorn]

https://www.udemy.com/course/ai-video-editing-and-production-one-click-video-generation/

This course is interesting, but I guess I wouldnt necessarily be learning about AI

[final kiln]

If you accidentally use data from test to train, that's called a data leakeage

The validation dataset is a bit more subtle, but it's the same as the test data set

When you select a model, you have all these "hyper parameters", which decide the shape of the model, the shape of the batches, and a bunch other stuff

So like, in order to see which hyper parameters work best, you do a sort of search, you train the model on different combinations of the hyper parameters

And this is the subtle part

If you use your test data set to choose the best hyper parameters

That still constitutes a data leakeage, because in a way, information is still flowing from the test dataset, but through you

So you get this validation dataset to choose your hyper parameters. Then you train the thing for longer periods of time, and then you test it against the test dataset

The test dataset is then used to measure how well the model will perform in the real world

[molten acorn]

Interesting

[final kiln]

There's a ton of stuff, but this is the gist of it. The challange, at least for me, has been the GPU shortage and finding the hyper parameters that dont overfit my model

In terms of intuition for why models work at all

The reason is that they are universal function approximators. Same way you can build up any function with tailor series, or any image with Fourier components.

Everything in the world can be described with functions

And models are like these machines that you can tune to approximate any function. So they can do all these amazing things.

[molten acorn]

Nice

I got my work cut out, I guess I'll start going through that book I showed you tomorrow and see how far that gets me

[final kiln]

Good luck

Also do check the pinned messages here

[dusty forge]

can you drop some screenshots when you have it running, curious how clean and clear it is

[final kiln]

still work in progress, but this is how my code is turning out, I'll run it in a sec

@flow
def main() -> None:
    model = load_model()
    loss_function = nn.CrossEntropyLoss()
    mlflow.set_tracking_uri("sqlite:///mlruns.db")
    mlflow.set_experiment("asa")

    eval_interval = 1
    save_interval = 1

    for epoch_step in range(100):
        cleanup_memory()
        for training_loop_step, model, loss in training_loop(model, loss_function, accumulation_steps=1):

            # observability into the training loop:

            mlflow.log_metric("training/loss", loss.item(), step=training_loop_step, synchronous=False)

            if training_loop_step % eval_interval == 0:
                metrics = perform_evaluation(model, loss_function, total_count=5)
                log_metrics(metrics, training_loop_step)

            if training_loop_step % save_interval == 0:
                mlflow.pytorch.log_model(model, "models")


if __name__ == "__main__":
    main.serve("train")

with the serve it runs a sort of worker that awaits for me to trigger, now I run the UI

it appears right here

[dusty forge]

oh this looks really great

how much mem does it use?

[final kiln]

I still haven't benchmark performance, but haven't noticed any sgnificant differnece

[dusty forge]

around 55% max it seems

[final kiln]

let me see if I can catch the process

[dusty forge]

not bad, I've also been looking at KNIME, which is a low-code pipeline platform, also open source and local, but the nodes can be fully SQL and Python

[final kiln]

no like, the max is 15gb

[dusty forge]

@final kiln this is how KNIME looks like

[final kiln]

also looks good, does it deploy to spot ?

[final kiln]

oh it watches the layers themselves ?

[dusty forge]

if you want to model the pipeline like that, sure

you can also just flunk everything in sequence like a notebook

[final kiln]

my main worry is infra tbh

if it manages aws for me im a happy person

[dusty forge]

https://aws.amazon.com/blogs/apn/how-knime-users-can-build-intelligent-workflows-by-accessing-aws-services-through-boto3-sdk-integration/

[final kiln]

@task
def forward_pass(model, loss_function, rating_b, text_bw) -> Tensor:
    """ Forward pass through the model. """

    predicted_logits_b5 = model(text_bw.int())
    loss = loss_function(predicted_logits_b5, rating_b)
    return predicted_logits_b5, loss

@task
def backward_pass(loss_train, multiplier: float = 1.) -> None:
    """ Backward pass through the model. """

    (loss_train * multiplier).backward()

@task(log_prints=False)
def gradient_descent(optimizer: torch.optim.Optimizer) -> None:
    """ Perform a gradient descent step. """

    optimizer.step()
    optimizer.zero_grad()

im a bit wary of having these in their own task

[dusty forge]

seems it can

[final kiln]

but code looks better for some reason

[dusty forge]

I think Prefect is a smaller install? KNIME is almost 1Gb

[final kiln]

oh damn

yeah it's a lot smaller

bit knime is looking good

its like labview

[dusty forge]

I think it doesn't get much attention because it markets itself as low-code solution, even though you can put full code into every node

[final kiln]

visual programming languages are still code imo, labview can get pretty complicated, I think visual is better for pipelines

[dusty forge]

yep I agree, it makes learning ML much easier if I can see how it fits into the whole story. I understand that most courses focus on experiments, simply to avoid you getting lost, luckily I work in data so I know how to be lost properly 🤣

[vernal quartz]

hi guys, what is the expectation from a applicat applying for a a junior NLP developer?

[final kiln]

Eh, it's easy to get lost yeah

[final kiln]

Uhm, what do you mean by NLP developer

[vernal quartz]

naure language processing dveloper

for gold text processing and prediction

so what is the base line for a data scientist junior role or deep learning engineer role?
I got the take home assignment for code filtering a large text talk aobut gold, and they asked you need calculate the probabilities about the imcoming text, but i realize i don't know shit about how to code a linear regression, so what is the expecation for a junior role?

[final kiln]

def training_loop(model: SentimentAnalysisModel, loss_function: callable, accumulation_steps = 10):
    optimizer = create_optimizer(model)
    training_data = load_dataset_from_npz("data/asa/test.npz")
    for step, rating_batch_b, text_batch_bw in yield_batches(training_data[0], training_data[1], 15):
        model.train()
        _, loss = forward_pass(model, loss_function, rating_batch_b, text_batch_bw)
        backward_pass(loss, multiplier=1/accumulation_steps)
        if step % accumulation_steps == 0:
            gradient_descent(optimizer)
            yield step, model, loss

def evaluation_loop(model: SentimentAnalysisModel, loss_function: callable) -> Iterator[int, float, float]:
    model.eval()
    test_data = load_dataset_from_npz("data/asa/test.npz")
    with torch.no_grad():
        for step, rating_batch_b, text_batch_bw in yield_batches(test_data[0], test_data[1], 15):
            predicted_logits_b5, loss = forward_pass(model, loss_function, rating_batch_b, text_batch_bw)
            accuracy = calculate_accuracy(predicted_logits_b5, rating_batch_b)
            yield step, loss, accuracy

I think I got myself a nice little pattern

[final kiln]

my understanding

is that those roles don't exist, there are no entry level positions for MLE and data scientists

Or at least, they are not positions that lend themselves very well to being filled by entry level people

[dusty forge]

@final kiln Prefect is orchestration in general right, not specifically for ML? I'm turning the website inside out and don't see anything about experiment tracking or how the model behaves, etc?

[final kiln]

It's general yeah, it takes care of logging and observability

[vernal quartz]

you code just did a baseline copy from any website that do a trainning, and its only for machine learning model i think, so i have n idea where it is impressive.

[final kiln]

And my hope is that it helps with the infra part, they mention it a lot in one section

[final kiln]

Yeah I guess we are all implementing the same Algo isn't it

But the part that I like is the yield statement

Which neatly seperates my observability from the ML stuff

[vernal quartz]

okay, move away

[final kiln]

sorry ?

[hollow furnace]

hey guys, anyone know what models to pick for object detection live camera feed on my mac from hugging face and how to set it up locally? im a newbie with setting up models

i know and download yolo models but dont know how to test it on the front mac camera

[final kiln]

Check media pipe, they have good pre trained stuff

[vernal quartz]

first you need to make sure you read the data correctly from your camera, check openCV topcis

then feed the live data to your models

for supervised leanring: you first prepare a large image data set from your camera, and lable them when the object is in the frame
for unsupervides learning: i have no idea

[final kiln]

I think dropout has done the trick, but can only be sure after more 3 or 4h, this gonna be one of my last runs before revamping all the infra

Batch size/LR combination give a lot of instability but I don't really see a lot of trouble in the eval so I'm just gonna let it be

[vernal quartz]

So what is the baseline expectation for a junior data scientist role? such for mechien learning and deep learning? what is the project make you outstand from this position? I know there are no junior role for it, but there are less expeirenced data scientist know basic stuffs

what you need to do and know?

[final kiln]

Like I said, these roles are usually not entry level, I think people usually start with data analyst or junior software

But knowing what a regression is oughta be a good start

[vernal quartz]

I know there are no junior role for it, but there are experience limit roles

[final kiln]

I've seen internships for ML Research scientist, but theyre usually for PhDs and stuff like that

[final kiln]

In terms of projects, theres a lot you can do to stand out

Just pick something that you are passionate about and make it impressive

[final kiln]

Tbh, I have this hunch. I'm gonna pre train the embedder module and pre compute the positional encoding. I recall during early testing that this was the part that would cause the model to overfit more rapidly

There's also the possibility of training it on next token prediction and use the in between signals to train a classifier.

An advantage of that approach is that I can reuse the model for the other tasks I have to do to replicate the MetaFormer study

[vernal quartz]

I lack of structured learning and i don't the water level

[final kiln]

Personally I find that learning outside a structured environment is way more fun

But it very much depends on each person's learning style

There's a ton of resources online, I like to recommend Khan academy for building up the math knowledge

3blue1brown has a great primer on ML too

[vernal quartz]

just for a understanding, i understand all machine learning concerpts, from linear to K-nearest to VSM, to linear algebra to CNN, RNN and TTS model design, as well spectrogram transforming (audio processing for TTS), is this deep ?

what is the water level of this

i think the only thing i don't understand is what the water level is a machine learning and deep learning engineer

what required for the job

[final kiln]

How come you understand linear algebra and CNNs but you don't know how to code a linear regression ?

[vernal quartz]

because i skipped the machine learning project, i learned all in 2 months, ML by nature is worse than deep learning

but i undertand now he many company maybe want a ML project as even worse

[final kiln]

People usually learn regression very early in college, which is why I thought it was weird

Ig you can bunch it with ML, but first time most people hear of it it's just an optimization in stats class or wtv

[vernal quartz]

i understand regression just like all other concerpts, but i don't understand how to code them

....

[final kiln]

"What I can't create I don't understand"

[vernal quartz]

that's ture

i have hands on experience in hot encoding to CNN stack, very limited, but i don't understand how to code a liner regression

LOL

[final kiln]

Idk, you can just apply gradient descent to any linear regression problem

Tho there's a ton other optimization techniques

My fav is the monte Carlo one + local refinement with gradient estimation

There's ones with Jacobian and whatnot

It's a whole thing

[vernal quartz]

we need a undersanding of this (we don't)

[dry geyser]

would pydantic questions fit here?

[vernal quartz]

ye, that's where my 2 months college lack of , remember bunch cauasions guys' name

[vernal quartz]

ye, i believe you

[teal lance]

Well this is my script I’m writing to create a strategy that’s built to trade strict criteria and focused on exiting the first signs of market reversals or the conditions not being true anymore

[final kiln]

Consider doing a degree if you can, it's worth it for many reasons

[teal lance]

Locking in profit and minimizing risk

[vernal quartz]

what degree?

[final kiln]

Ah, that's a good question actually. I don't know. I come from Physics for example. I'd wait for other people to answer that question

But I reckon a lot of people do Computer Science, others do an actual ML degree

[vernal quartz]

that's a good question, how did you get the job and how did you learned?

[final kiln]

Uhm, I got a lot of hands on experience during my masters thesis, which ended up being a very successful project that produced research and even got me to do a talk and a couple poster presentations and published abstracts.

I got a job in software fairly easily, it was during the peak hiring frenzy at the end of COVID. Did a bit of everything, frontend, backend, infra, etc etc. My last role in particular, started as a fullstack but quickly grew into an MLE role, I was training models, etc.

Right now I'm in between jobs more or less, I'm interviewing for companies in Switzerland, MLE roles only, I got all this experience behind me which makes me a safer hire, ML is very expensive.

[vernal quartz]

bro from full stack to MEL and thinking ML is experiensive

[final kiln]

And I'm still learning quite a lot. This last one I thought was gonna be easy. Just simple text classification. But it turns out it ain't easy

[vernal quartz]

great

[final kiln]

ML is super expensive, since there's a huge GPU shortage and it requires a lot of experimentation

[vernal quartz]

What is your master degree, i was wanted to do one

[final kiln]

Actually haven't defended my thesis yet, but it's more of a formality by now

Master is super worth because you get to do a thesis

If you choose an hands on thing, your time is being used very well rite

You're both getting XP in software, making all the mistakes you need to make to learn

But you're also getting a degree and possibly other cool stuff

[hollow furnace]

is media pipe a model? looks like its not on hugging face.

[vernal quartz]

I think you don't understand what you saying, apprantly you from full stack to MLE with a different degree

[final kiln]

Mediapipe is a python package with a bunch of ML stuff like object detection

[vernal quartz]

what you learned monte Carlo one + local refinement, in bachelor?

[final kiln]

I'm not from fullstack, I just worked several roles throughout the last two years, on startups this is natural

[hollow furnace]

ah ah, well my boss for this project wants me to look multiple models on hugging face and find the best models and put all the metrics on an excel spreadsheet if that makes sense. like specifically on hugging face.

[vernal quartz]

Uhm, I got a lot of hands on experience during my masters thesis, which ended up being a very successful project that produced research and even got me to do a talk and a couple poster presentations and published abstracts.

I got a job in software fairly easily, it was during the peak hiring frenzy at the end of COVID. Did a bit of everything, frontend, backend, infra, etc etc. My last role in particular, started as a fullstack but quickly grew into an MLE role, I was training models, etc.

?

[final kiln]

I used it during my thesis to fit data to be used during particle transport simulations, I had a lot of data on various materials which had to be used effectively during the simulation

[final kiln]

?

Medical Physics, I specialized in monte Carlo simulations of particle transport for predicting energy deposition of radiation treatments

[vernal quartz]

and doing full stack

[final kiln]

What about it ?

[vernal quartz]

then you aske people with a degree

i don't much believe you but good luck

[final kiln]

lol

[vernal quartz]

i guess i will being a wonderfull ML engineer

[final kiln]

You don't believe which part exactly

[vernal quartz]

full stack

=> MLE

[final kiln]

It's a pretty natural path I'd say

[vernal quartz]

and ask people for a degree

[final kiln]

MLE requires you to have understanding of software

[vernal quartz]

and saying MLE is expensive

[final kiln]

I'm confused, please make a point

Are you saying ML is not expensive ?

[vernal quartz]

you transfer from full stack to MLE in same company, means ML is not expensive

becuase its cheap

[final kiln]

What

[vernal quartz]

they don't matter cost

[final kiln]

Have you ever worked in the industry ?

[vernal quartz]

nope

[final kiln]

Makes sense

There's a lot of role fluidity in startups

You're expected to do a bit of everything

[vernal quartz]

so startups affored expensive ML

[final kiln]

Well they try

Some have funding for it too

But less hands on deck usually

[vernal quartz]

your logic didn't make sense from frontend to MLE as physic grad
=> ML is expensive
=> do a degree

[final kiln]

Less people means generalists are preferred

That's not a logic, that's my employment and educational history

Which do not say much about ML being expensive or not

What determines the cost of ML is the hardware and the nature of the task

Not if a random person has transitioned from full stack to MLE within a startup

What kind of argument is that even >.>

[vernal quartz]

while you should stufy pholosphy first, i thought smart people like i am doing physics

[final kiln]

Okay, you are now just saying random stuff, I'm gonna attend to my training loop

[vernal quartz]

i mean, i guess the MLE dosn't need much experiences, as i guess from this conversation, so thanks

at least i know the water here

[final kiln]

You strike me as a pretty hostile person. I'd work on that.

You have no idea how far you can get in life by just striving to be a good person

[vernal quartz]

because i got all the easy interviews from data scientist or NLP or DL, i guess you really don't need experience on this one

i mean i will learn the optimizing tho, even for deep learning there is no much small tweaks, for machine leanring i guess there little bit?

like your fav My fav is the monte Carlo one + local refinement with gradient estimation as your only did project in your school

[hollow furnace]

so do u think i go use openCV or media pipe like suggested above?

[vernal quartz]

i don't have experience in company wise project.

[final kiln]

Open CV works if I'm not mistaken, but mediapipe comes with the stuff all setup, easier for you just starting out I assume

[vernal quartz]

for the porject you describled, use openCV read the camera images, store the image data set, lable the images has targets, you need to deal with with image transforming, such as what size of the image you want, and then you need a model for trainning

[hollow furnace]

lol its not really a company project. more like a professor trying to test me before I can join his lab, but ig u can say he is my boss later on.

[hollow furnace]

and the model u get from hugging face?

[final kiln]

https://developers.google.com/mediapipe/solutions/vision/object_detector

[vernal quartz]

for what purpose? jus object dection? what objects

[final kiln]

https://stackoverflow.com/questions/70650357/python-opencv-external-camera-live-stream

Just mix both

Mediapipe is good, I've used it for a PoC recently

It fails here and there, if you want better performance you'll have to train your own CNN

[vernal quartz]

if you don't know what is a openCV, quit the topic

[hollow furnace]

actually all objects really. the overall project is to get the camera feed. postprocess it and pass the feed running multithreaded through different models so different models can focus on detect their "purposed" objects better than just using one model.

[final kiln]

Could you please explain what you mean ?

Anyway

[vernal quartz]

and how come openCV and mediapipe can just "just mix both"?

[final kiln]

I think you should be a bit more humble

Pleas open the link

@hollow furnace I just sent you the two links you need, the first one is the package that will do object detection for you, the second is the code that you can use to read from your camera

@vernal quartz Please stop being hostile towards me.

[vernal quartz]

you should be more humble, this is the harmer to humble you

[vernal quartz]

I guess YOLO is a overall better chocie, but you can try different models. I guess if you only use the APIs for model trainning, the main problem will be adjust the camera image so you can have a standarlized image database
The more important part is how you capture the images and then standardlize them such as size to feed all the models, also multithreading coding stuff

[final kiln]

This is not impressive or humbling

[vernal quartz]

i guess you didn't have an IQ to look through it

[final kiln]

Sigh

[vernal quartz]

i give you 1 million for using image detection techniqes for these kind of CAPTCHA recoginition

try and do it

this is called Lemin CAPTCHA

this is my own created algorithm for image recognition, it doesn't even used ML and deep learning

[final kiln]

I'm sure your technical skills are good for someone without experience, but people are also looking for someone they can work with, it's horrible to work with people who are hostile and think they know more than the experienced folks

[vernal quartz]

i guess you can't even see how many anti-recognition layers on there, this is not your common life image, and i guess you even don't have the skill for nature world images processing and recoginition

from skills good for someone without experience, then half sentence drop the skill talking about personality,

again, this is a multi-layer anti recognition CAPTCHA

this is not a common even real world image

i guess someone from physic transfer to MLE don't understand it

[final kiln]

It's time for my coffee I suppose.

Yesterday I found out my model was overfitting, which I thought wouldn't be possible on such a. Large dataset >.>

[vernal quartz]

my hammer is coming

billyboby

[left tartan]

That’s not really the way we do things in this server. The whole point is the welcome any question, not gatekeep who can ask. The less you know about a topic, the more reason to ask. “Everyone you meet knows something you don’t”

[wicked osprey]

Hiya, fella's. I'm kind of eager to figure out why data science is interesting and what benefits it provides.

Does anyone mind sharing their insight?

[left tartan]

And, if someone asks something that’s wrong, either take the time to explain… or just don’t respond. It’s all good.

[vernal quartz]

GOT IT

[left tartan]

Can you be more specific? Data science is a broad field

It’s like asking ‘why math is interesting’, very hard to answer succinctly

[wicked osprey]

Would you be so kind to elaborate the most common topics within data science, and what purpose they might serve?

[vernal quartz]

Data Science is only two parts, more tradiontional data anaylzing part that most replaced by GPT, and other great data scienctist doing there job i think, but mostly 300K+ plus just my future, others reporting their basic insights to the captlislim companies that don't know how to or too lazy to use GPT. The second part the only fun part is deep learning, no gurantee income, company don't understand it, but they want it

[vernal quartz]

answer good?

[wicked osprey]

So the process of analyzing sets of data and extracting conclusions from them in a nutshell?

[left tartan]

Data science is hard to define the edges of. On one end, it starts with statistical techniques for the analysis of data and on the other, machine learning and AI topics, which include NLP, LLMs (like GPT), LSTM and other techniques.

[left tartan]

Not necessarily conclusions: sometimes just an understanding of the data

This def seems pretty good:

https://en.m.wikipedia.org/wiki/Data_science

[wicked osprey]

Could you summarize practical use-cases that are most common in the industry?

[vernal quartz]

heavy or easy process part maybe given to data engineer, data scientist has two parts as i mentioned, really depends on what level of them, most them = GPT and others are great, these excluded DL engineers

[final kiln]

Uhm, within ML you get the fancy stuff like semantic segmentation, text summarization, all the generative AI stuff

[left tartan]

Are you writing a paper? Feels like these are paper prompts.

[final kiln]

I've heard someone say Data Science can be thought of as an evolution of statistics

[vernal quartz]

oh gush

please documented what i said and write on your paper

that will be so ture and funny as hell

[left tartan]

I like to think of it as a mashup of CS, Stats and often some domain (ie; finance or physics or whatever)

[final kiln]

Yeah makes sense, I always wonder if I should learn more stats, so far I haven't needed a lot of the advanced stuff when doing ML

[vernal quartz]

do you know what is a good data scientist?

and where to learn?

[left tartan]

#data-science-and-ml message is where to start

[left tartan]

You can never learn too much stats (half joking): it’s completely worth learning a little stats every week. You’ll be amazed how often stats problems come in real world

In particular, the subfield of EDA is very useful.

https://en.m.wikipedia.org/wiki/Exploratory_data_analysis

[wicked osprey]

No, I'm genuinely curious on why people think it's an interesting field.

[vernal quartz]

its creating GPT!!! Like i do i will create terminator!!!

[final kiln]

Oh I think Ive always done it in one form or another, like I just know what to look for when I get a bunch of data in front of me.

I wonder how close what I do is to what is supposed to be done tho >.>

[vernal quartz]

The only way you can get free labor is from GPT or

can i say it?

[left tartan]

Why DS? Because you can discover new information from old information, and new information makes you smarter. Eli5 version

[vernal quartz]

or from when America created

[final kiln]

I think ML is fascinating because it touches on stuff I like, math, philosophy and computer science. On top of it, it'w usually applied to some domain, so I can actually mix all the stuff I like into one job

[wicked osprey]

What could the results of ML contribute towards to? Like more efficient ways of doing generic tasks?

[final kiln]

That does sound like an assignment question

[vernal quartz]

ML is trash, its based on data analyzing

[left tartan]

Perhaps it makes an airplane more efficient. Or prevents a country from an agricultural disaster. Or answers your homework question. Anything, I guess

[wicked osprey]

Hmm, how could it prevent a country from an agricultural disaster?

Comparing previous data with existing current data and forecasting some sort of scenario?

[final kiln]

But it has the potential to disrupt entire industries. I'm not convinced about the singularity actually, recently changed my mind about it

[left tartan]

Sure. Predicting weather perhaps? Or crop cycles?

Or finding more efficient planting methods, etc

[final kiln]

My next project is gonna be an image classifier that tells me when my pizza is ready, cuz I need that in my life

[wicked osprey]

Are there any dynamic changes it could bring towards the development of an individual. Ex; would you be able to use your skillset as a way to enhance your lifestyle?

[wicked osprey]

Interesting.

[final kiln]

Depends on what you mean, but there's a lot that can be said about that

[wicked osprey]

Could you define a few examples?

[vernal quartz]

farmers give us better food

[wicked osprey]

And is forecasting the only type of byproduct of ML/DS or are there more?

[left tartan]

Iirc, the very first web order site was for pizza.

[vernal quartz]

you can anaylzing how many nicotine you inhalred, and predicte when you will die based on your speices

[wicked osprey]

LOL, that's quite an example.

[vernal quartz]

i guess that's what called death are coming

[final kiln]

Just look into what open AI is doing, they're doing so much cool stuff rn

[left tartan]

This topic is also https://en.m.wikipedia.org/wiki/Actuarial_science

[wicked osprey]

Was ML or DS the major contributing factor of OpenAI's new tools?

[left tartan]

What’s the difference?

[wicked osprey]

Zero clue.

[final kiln]

I believe ML is a subset of DS, or something like that

[left tartan]

I think you should probably start with those Wikipedia pages, and find a good explanation and read it thoroughly. These are broad topics that need a more thorough explanation

I’m not sure we can adequately explain the entire industry effectively here

[wicked osprey]

Ah, well.. I appreciate the insight(s) and guidance.

[vernal quartz]

depends on new tools, openAI GPT model is using DL creating and analazing, sora is another transoforming model, in the nxet few years we can transfer babies to food

[wicked osprey]

[left tartan]

Theyre good questions, it just sounds like you should read a basic ‘what is DS’ page to get the terminology

[wicked osprey]

Perfect, HBS. Source (https://online.hbs.edu/blog/post/what-is-data-science) for learners that want to tune along.

[final kiln]

Yeah Medical is huge

[wicked osprey]

Thanks, just a bit dazzled as an full stack dev.

This entire industry seems like a new digital world opening up for me.

[vernal quartz]

please remember data science only two parts, one part is GPT stuff, the other part is DL

[left tartan]

The DS terminology gets very confusing quickly, since there’s so many branches and techniques.

[serene scaffold]

ChatGPT depends on DL, so that isn't a coherent partitioning of DS.

[final kiln]

Actually one of my collegues was working on similar stuff, detecting cancer in PET scans or something of the sort

[vernal quartz]

no, deep leanring developer is a different role, but most industry don't have enough courage or people to hire as a deep learning developer

[wicked osprey]

Are DS'ers involved in building the algorithms, i.e. developing the software? Or are they limited in some sort of way? (Had a quick skim, couldn't get to the bottom of it) So far, I've seen most is data gathering.

[serene scaffold]

There is no consistency in what a "data scientist" actually does.

[final kiln]

Uhm, a lot of the data scientists I worked with seemed confined to jupyter notebooks and such, but I also heard that there's a trend towards requiring software skills from data scientists

[left tartan]

I laughed at how the HBS link above separated DS and DA (the distinction is very muddy in real world)

[wicked osprey]

So, pretty transformative. You just get to dabble in everything that's necessary? Quite chill, imo.

[serene scaffold]

what I mean is, companies can ascribe whatever job responsibilities they want to a data scientist role. it doesn't mean that data scientists get to "dabble in everything" in the context of their empolyment.

[wicked osprey]

This question is for the employed and or fascinating hobbyists, what responsibilities do you get to adhere to?

[final kiln]

I wish there was enough time in a human life to do it, there's a ton of cool stuff to do in this industry

[wicked osprey]

I wish, these time constraints on our lifespan as humans are quite tedious atleast.

[final kiln]

My last role was at a small startup, I did a bit of everything from frontend, to platforming and to ML and even research (customer was keen on one of the projects)

I think at larger companies you'll be a lot more confined to your lane, but I've been in small places only so far

[hollow furnace]

im tryna run this code on vs code but it seems like cv2 didnt detect any feed: cv2.error: OpenCV(4.9.0) /Users/xperience/GHA-OpenCV-Python2/_work/opencv-python/opencv-python/opencv/modules/highgui/src/window.cpp:971: error: (-215:Assertion failed) size.width>0 && size.height>0 in function 'imshow'

[final kiln]

Looks like a validation error, did you notice anything in your camera

[vernal quartz]

I lost interests when i read here

[wicked osprey]

I'd say in general, it's normal to not have interest or a positive bias in abstract knowledge. Say you were to make it tangible and realized it into a project, you'd have more pleasure.

[hollow furnace]

oh and im using the front mac camera btw, not some plugin external camera to the side

i dont think so, didnt see any windows pop up or any green light flashing

[left tartan]

That’s why school is important: to ‘force’ you through these foundational topics

[final kiln]

Wait did you check the answer ? I think there was a bug in that code

[wicked osprey]

Quite rigurous, you could use some positive reinforcement or a different mental model to tackle your problem. Gets you through the boring stuff.

[final kiln]

Matrices are just crazy and keep getting crazier the more you study them

[vernal quartz]

no i understadn what is matric and what is linear, i just don't have courage to read the mathatics

i just lost the interest when i read N

[final kiln]

Try this one

import cv2 

  

  
# define a video capture object 

vid = cv2.VideoCapture(0) 

  

while(True): 

      

    # Capture the video frame 

    # by frame 

    ret, frame = vid.read() 

  

    # Display the resulting frame 

    cv2.imshow('frame', frame) 

      

    # the 'q' button is set as the 

    # quitting button you may use any 

    # desired button of your choice 

    if cv2.waitKey(1) & 0xFF == ord('q'): 

        break

  
# After the loop release the cap object 
vid.release() 
# Destroy all the windows 
cv2.destroyAllWindows() 

[hollow furnace]

yeah. hmm maybe its a user password one at the bottom but 1) i dont know if my camera has credentials and 2) mac password has an @ symbol so dont know if it messes with the '@192..."

[left tartan]

Yah, academic writing like this takes a lot of practice and familiarity with the boilerplate stuff. If you’re out of practice, it’s painful.

[final kiln]

I just searched online

[vernal quartz]

i mean after read it i know its talking about metrix, but what is EN?

what is this shit?

[final kiln]

https://docs.opencv.org/4.x/dd/d43/tutorial_py_video_display.html

[wicked osprey]

We're on the same boat.

[hollow furnace]

in the Natural number field

[final kiln]

This one is the official docs

[left tartan]

Like, I know the terms, etc here so I can just skim and understand what it’s saying. But take a paper on a topic you’re not familiar with, you’ll get stuck on the definitional stuff

[hollow furnace]

n is an element in the set of natural numbers

[left tartan]

The e means; an element of

[final kiln]

It's like any programming language really

Except people keep adding features to it

There's actually an isomorphism https://en.m.wikipedia.org/wiki/Curry–Howard_correspondence

[hollow furnace]

dang, its asking me permission for camera. i think that might work. ill wait until it runs completely

[vernal quartz]

so it means MN is in nature number

Fuck

[hollow furnace]

@final kiln wow, ur a wizard. it really works.

[final kiln]

Sets are actually a good start to study math. Everything is sets. The number zero is the empty tuple, which is a set with the empty set, something crazy like that

[vernal quartz]

why we don't just say it MN is in nature numbers?

[final kiln]

https://en.m.wikipedia.org/wiki/Surreal_number

[hollow furnace]

like say it in english like what ur saying here?

[final kiln]

This is the coolest number system ever

[tidal bough]

because that takes 5 times more characters to write

[hollow furnace]

cuz math needs precise definitions and should leave no doubt for the reader. at least thats what my math professor said.

[vernal quartz]

this entire paper is bsting about rows and columns

and then i realize its matrx adding

fk

help

what is this

[hollow furnace]

i think its also bc u want math to transcend beyond the language barrier. so a french math nerd can still understand some papers written in english from uc berkeley folks for example.

[vernal quartz]

i know its talking about a matrix, but i felt discouraged

????????????

????????????

[wooden sail]

do you have a particular question about this?

here they're just introducing notation

[vernal quartz]

what is this mean?

how to read this?

[hollow furnace]

is this saying ur multiplying elements diagonally and then summing them up?

[vernal quartz]

waht is the reverse A?

[wooden sail]

it says in english: for all matrices A, B, C containing real numbers and with shapes m x n, n x p, and p x q, respectively, multiplication is associative

[serene scaffold]

this is what ML scientists have been demanding our respect for?
we've been played for absolute fools.

[vernal quartz]

the angle?

[serene scaffold]

"all". and upside down E is "exists"

[tidal bough]

forall

[serene scaffold]

backwards E, I should say

[hollow furnace]

for all A in Real number field of size m times n

m times n are dimensions if i remember correctly

[wooden sail]

.latex $\forall$ means for all'' and $\exists$ means there exists''

[wooden sail]

.latex $\in$ means ``is an element of''

[vernal quartz]

so "all" means A or all means A,B,C, and why they formated like A,B,C:, does they connected?

[tidal bough]

i always get jumpscared when I remember this is how one writes quotes in latex

i'd much rather csquotes, thank you.

[wooden sail]

not sure i understood your question

[hollow furnace]

they are connected. but the for all means for each of the elements specifically

they're connected tgt by the last part

[wooden sail]

ah it means all A, B, and C with the given properties

[hollow furnace]

(AB)C MUST be equal to A(BC)

[tidal bough]

forall ..., it's true that (A B)C = A (B C).

[vernal quartz]

so for A in R, B in R, C in R, (AB)C=A(BC)

[wooden sail]

with the given shapes

[serene scaffold]

not just "in R". the dimensions matter.

[wooden sail]

since matrix multiplication is not defined for arbitrarily shaped matrices, the shapes are important

[serene scaffold]

R^(a x b) means "an array of real numbers of shape a-by-b"

[tidal bough]

well, this specific statement is true as long as it even makes sense. matrix multiplication is associative for all shapes for which it's even defined.

[vernal quartz]

∀A, B ∈ Rm×n what is A,B means in R 2 Darray?

evey element?

[serene scaffold]

.latex Have you ever seen $\mathbb{Z}^{p \times q}$?

[serene scaffold]

it would mean that both of A and B are m-by-n arrays.

[wooden sail]

.latex formally, the $\times$ symbol denotes the cartesian product. consider the tuple $(a, b)$ where $a$ and $b$ are both arbitrary real numbers. we can describe this as $\real \times \real$, or equivalently $\real^2$

[serene scaffold]

(that is, they are both elements of the set of all m-by-n arrays)

[wooden sail]

hmm do we have \real in the header

i guess not

[serene scaffold]

is that just an alias for \mathbb{R}?

[wooden sail]

.latex formally, the $\times$ symbol denotes the cartesian product. consider the tuple $(a, b)$ where $a$ and $b$ are both arbitrary real numbers. we can describe this as $\mathbb{R} \times \mathbb{R}$, or equivalently $\mathbb{R}^2$

[wooden sail]

yeah. also some modulo stuff, like Z/nZ

[vernal quartz]

but isn't A,B are elements in m-by-n arrays? why they became the m-by-n array?

[wooden sail]

.latex $A \in \mathbb{R}$ and $a \in A$ mean different things

[wooden sail]

oops, forgot the m xn

anyway, one says A is a multidimensional array

the other says that lowercase a is an item inside of the matrix A

[vernal quartz]

so waht is

[wooden sail]

same thing

for all A and B of size m x n, and for all C and D of size n x p

[vernal quartz]

A is array or element

[wooden sail]

it's telling you right there

that's exactly what that says

you might wanna invest some time in familiarizing yourself with the symbols

you already know "for all" and "is an element of", so you're able to read this

[final kiln]

I found my issue

the first layer is overfiting hte data, fo' sure

[vernal quartz]

so A is an array and a is element?

[wooden sail]

in what you shared, yes, but symbols are not universal. different authors will use different symbols

the symbols for "is an element of", "for all" and "exists" are usually kept standard, but anything else will vary depending on where you read it from, so it's in your best interest to learn how to read it, not memorize the symbols

[vernal quartz]

i will use a shit symbol when i publish articals

[wooden sail]

you can use whatever you like as long as the notation is made clear

some of the more cursed stuff uses musical symbols

[tidal bough]

oooh, algebra flashbacks

i don't even remember what that means anymore

[wooden sail]

some cursed commutative diagrams among eldritch objects

https://en.wikipedia.org/wiki/Musical_isomorphism one example

[vernal quartz]

but, just want to confirm, what is actually A,B in Rmxn means, because it should be mean element right??

[wooden sail]

all you have to do is scroll up to where i told you exactly

but really, you're not reading

what you're asking me is what the image is telling you

sit down calmly and translate the expressions into english or whatever language you like

[final kiln]

if I use like, a super agressive dropout on the first layer, would that work to prevent overfitting ?

[vernal quartz]

but you said it can be element or an array

i don't know if that make seens in the notation it means array

∀A, B ∈ Rm×n, C, D ∈ Rn×p : (A + B)C = AC + BC (2.19a)
A(C + D) = AC + AD

[wooden sail]

start with ∀A, B ∈ Rm×n

what does that say?

[vernal quartz]

for all A,B in Rmxn array?

but what is A,B?

[wooden sail]

let's do this another way

if i write for i in range(10), what is i?

[vernal quartz]

you mean code?

[wooden sail]

that's python code

[vernal quartz]

it will create a list, from 0 to 9, and read from it

[wooden sail]

and what is i?

[vernal quartz]

this is off topic

[wooden sail]

no, it's an example of the same notation

that line of code tells you i is a number in the range from 0 to 10

[iron basalt]

Let A = {apple, strawberry, grape, 7}, then a∈A is either an apple, a strawberry, a grape, or 7. @vernal quartz

[wooden sail]

the line you shared tells you to consider some object A, doesn't matter what it is, that is in the set R^mxn

[vernal quartz]

i is a memory, bytes, when it read it be assingend with bytes, represent each 0 to 9, its never been a number

[wooden sail]

the variable i will have different values in different iterations, but the value will always be in the range from 0 to 9

[iron basalt]

This notation is set notation, dealing with sets of things. E.g. the set of all real numbers (the fancy R).

[odd meteor]

How aggressive? Something in the range of 0.7 - 0.9? Well, I guess the only way to know is to try it out. Maybe throw in BatchNorm + Dropout

[vernal quartz]

i said its off topic beacause its a bytes, nothing realted to the question

[final kiln]

I'm gonna give it a try, but tbh, the solution will likely be to steal lamma's embedder, or the one from uncased bard from hugging face

[wooden sail]

even if you want to look at bytes, they're saying exactly the same thing

you're failing to see the connection

look also at the example squiggle gave you

[final kiln]

positional encoding I can just pre-compute with the sine's fucntion thing from the 2017 paper

[vernal quartz]

math is a representation, it is not existing, while bytes and coding stuff is machanical stuff, it should never be reprenseted as non-existing meanings, such as number

[wooden sail]

i have bad news for you if you intend on working with AI, then

[vernal quartz]

a bytes is a physical stuff

[wooden sail]

it's also definitely not mechanical

[vernal quartz]

and number is not

[wooden sail]

well, good luck

[final kiln]

wasn't penrose who published a book about all math being real and we just live in the chunk of it that allows for sentience to exist

[hollow furnace]

wait so for this portion set up in google mediapipe, do u want this to run while the camera is on (in the while loop)? and what goes in the detector initialization part at the bottom?

[iron basalt]

Math is a way to communicate ideas (including to oneself, by augmenting ones own memory via external devices such as paper). It also happens to be the best known way to communicate these kinds of ideas as efficiently as possible. Prior to this kind of notation proofs were in plain words, and they were long and really hard to follow. Trust me, you don't want that.

[final kiln]

I think you need to code the while true (also called event loop) inside the with

[vernal quartz]

if i write for i in range(10), what is i?
jjjj — Today at 1:14 AM
you mean code?
Edd — Today at 1:14 AM
that's python code --- i just saying this is totally two questions, as one is coding behending machanical, such as logical gate and can't simplified as fictional - numbers

you are describing a design machinal and designed physical situation within memories compare with a pure mathmatical (fictional) situation

back again, so A,B is array in R or A,B is element in R?

[wooden sail]

you're gonna have a really bad time if you don't learn to abstract real world phenomena

[vernal quartz]

array or element or both

[wooden sail]

scroll up to where i already answered the question

[vernal quartz]

because if A,B means array i don't understand the notation

[final kiln]

wait til you get to whatever the heck category theory is

[wooden sail]

that does not mean array and has nothing to do with that

[vernal quartz]

so you mean everything in this R

[wooden sail]

no

[vernal quartz]

what is it

[left tartan]

Just to drop a Wikipedia link: https://en.m.wikipedia.org/wiki/Universal_quantification

[wooden sail]

why don't you just read the line and replace the symbols we discussed with their meaning in english?

[agile owl]

i think the hardest part about math is the notation is not as descriptive as code

[vernal quartz]

A is element in R or array?

[agile owl]

if you don't know what you're reading it might as well be hieroglyphics

[left tartan]

In this conversation, I think the notation and code are interchangeable.

[wooden sail]

it's the same

and you mean math notation, because maths is independent of the notation

you could just as easily do it in english if you use precise language

[agile owl]

right, but everyone uses math notation

[wooden sail]

yes, but not the same one

[final kiln]

Category theory is a general theory of mathematical structures and their relations

I thought study of mathematical structures and their relations was maths, so like category theory is meta math ?

[left tartan]

Where are you coming up with ‘array’?

[wooden sail]

every single book and paper starts with a description of what the symbols will mean throughout their work

[iron basalt]

Well, actually, it depends on which notation you use. Some is just straight up code, and it's also one of the preferred ways to provide a rigorous proof (since it can be checked by a machine).

(See formal methods)

[agile owl]

I'm not talking about math papers I'm talking about papers that use math

[wooden sail]

same thing

[agile owl]

most of the time people aren't proving their results or anything in model papers

[wooden sail]

they will start with a definition of the symbols, and they will also have text and figures and pseudo-code

be it pure math papers or engineering ones

because a mixture of all those things is perfectly valid, and the symbols are not universal

[wooden sail]

AI papers are generally bad is my 2 cents for you

[left tartan]

For any A,B where each are elements of Rmxn

[vernal quartz]

Thanks

That PRO

[wooden sail]

all you had to do was scroll up to where i had already told you that

[vernal quartz]

i just past your anwser, you told me you are not sure

[wooden sail]

what?

[vernal quartz]

you said one says its an array, one says its an element

[wooden sail]

yes

[vernal quartz]

????

[iron basalt]

The math for this kind of stuff does not really fully exist yet, nor do the AI/ML papers often make any sense. The notation / meaning tends to fall apart once they start using notation involving probability, often due to a lack of analysis (the mathematical branch).

[wooden sail]

you shared an image with an element of a matrix, and with a matrix

[vernal quartz]

it doesn't make sense if it is an array

[wooden sail]

you're not reading and you're refusing to understand

sit down calmly for a second and read carefully

.latex the image you shared spoke of a scalar element $a \in A$ and a matrix element $A \in \mathbb{R}^{m \times n}$

[agile owl]

I thought it was me having a problem good to know that other people think a lot of the papers don't make any sense

[wooden sail]

and i told you to read the symbols carefully so you can distinguish them, because authors will use different symbols

nowhere in what i wrote was there any "i'm not sure", there was only "learn to read the notation, don't memorize"

.latex going back to the for loop, you can write it in the form $i \in [0, 10)$ which is of the same kind of what we're discussing here

[left tartan]

I think I understand part of the confusion: the element type depends on the what the set contains. Could be anything. Could be the set of all adults in Kansas. Could be the set of all Real numbers. Could be a set of all sets containing at least one prime number, etc

The element would thus be any single member of the set

[wooden sail]

i also explained the notation R^{m x n} above, but they skipped that, it seems

[iron basalt]

Papers are not perfect, it's a messy thing, lack of knowledge for the math, or even trying to force it in to somehow make the paper feel more valid (instead of just "IDK, it works in practice").

[wooden sail]

i am bereft of patience now so i'll be on my way

[left tartan]

‘Gotta publish something this semester’

[agile owl]

Trying to remember how to index into my numpy tensor

ndarray whatever

3d array

I ended up fixing my polars so I only collect once but now I have this big honking numpy array I have to index manually for different columns

I guess I could just use the eager frame

[iron basalt]

The reality is that a lot of things start out as "IDK, it works in practice," and stay that way for a long time. This is not a bad thing. It's how all inventions (even in math) tend to start. But academia does not like that one bit, because it requires admitting a lack of knowledge (that does not exist yet), that you are on the bleeding edge.

[agile owl]

does someone have a one liner to get the upper triangle only in a flattened numpy covariance matrix

im about to ask chatgpt

[vernal quartz]

i will listen to Bobby

[final kiln]

In a flattened array ?

[agile owl]

I didn't even know np.triu existed

I feel dumb now

[left tartan]

That’s a bad idea. Edd actually knows this field, I’m just a visitor.

[wooden sail]

triu would've been my suggestion

[vernal quartz]

so the single member does contain an array, because array at least 2 elements, wait, array can be 1 element, but when we talking about the elements in array, there are only elments, there are no array

[left tartan]

The universe is the set of all (m,n) dimensional arrays. M and N could be whatever you want.

[vernal quartz]

but in a array, it only containes elements

[left tartan]

You’re confused.

[agile owl]

        self.observation_space = spaces.Box(
            low=-np.inf, high=np.inf,
            shape=(n_lags, 17 * self.no_symbols + 39 + self.no_symbols*(self.no_symbols+1)//2)
        )

I need to write a comment on the shape of this thing at this point it's gotten monstrous

[left tartan]

You’re using the term element in an ambiguous way, which is your confusion, I think

X element of Y means any member of the set Y

[vernal quartz]

let's say this array, it can only contain elements. ∀A, B ∈ Rm×n, C, D ∈ Rn×p : when we use this A and B means anything in side this array as a single element, so it will be element

[iron basalt]

Apple, strawberry, grape, and 7 are all elements of A. @vernal quartz

[vernal quartz]

yes, i know

[left tartan]

No.

[vernal quartz]

how come this became a array for A?

[left tartan]

(Someone correct me if I’m wrong)

A and B are both m,n dimensional arrays

[vernal quartz]

Associativity:
∀A ∈ Rm×n, B ∈ Rn×p, C ∈ Rp×q : (AB)C = A(BC) (2.18)
distributivity
Distributivity:
∀A, B ∈ Rm×n, C, D ∈ Rn×p : (A + B)C = AC + BC (2.19a)
A(C + D) = AC + AD (2.19 ----- so you saying A,B can be an array in nxp shape R array?

[iron basalt]

What is "this" in this sentence?

[final kiln]

I'm not understanding the confusion, this is pretty simple to interpret I feel

[vernal quartz]

the question is A,B can be an array in this notation?

[final kiln]

They're matrices

[vernal quartz]

as an element in nxp shape R array?

[left tartan]

The confusion is whether the ‘element of’ is an element of an array or is it an member of the universe of all arrays

I don’t know how to explain differently

[final kiln]

A e R2x2 means A is a 2x2 matrix

With real value elements

[iron basalt]

If you had something like: Let T = {[1, 2, 3], [4, 5, 6]}, and a∈T, then "a" is either [1, 2, 3] or [4, 5, 6].

[tidal bough]

consider reading ∈ as "belongs to the set", for less confusion. "element" is an ambigous term.

[vernal quartz]

∀A, B ∈ Rm×n, my question is, as Rmxn is an array, so array only contains single digital numbers or anything in this array, so can A,B be an array

[left tartan]

Rmxn is not an array

It’s the set of all m,n arrays

(Real valued arrays)

[vernal quartz]

∀A, B ∈ Rm×n, C, D ∈ Rn×p : (A + B)C = AC + BC ===> because if you think A,B can be an array, i don't understand the following: (A + B)C = AC + BC, so this can aslo equaltion to array additiaon?

[vernal quartz]

(A + B)C = AC + BC so this is array addition?

[final kiln]

That's a matrix equation

[tidal bough]

... uh, yes? this identity involves matrix multiplication and matrix addition.

[final kiln]

It's stating that the equation holds for all such matrices

Every time you write a math sentence you're implicitly stating it to be true

[vernal quartz]

(A + B)C = AC + BC, if A,B are matixs, is this notation right? or equation

[left tartan]

If a and b have the same shape

That’s why they’re both members of Rmxn, to say: they both have same dimensions.

[agile owl]

how do I flatten the upper triangle only though

that's still the issue

np.triu gives you nans elsewhere but if you call .ravel() or something it still gives you the nan values

[tidal bough]

you could make a boolean mask with ones in the upper triangle, and do arr[mask]. that'd be a 1d array.

[left tartan]

You could just mask it with … ^

I guess you could invert it and triu?

Oh, tril duh

[tidal bough]

a oneliner would be arr[np.triu(np.ones_like(arr))!=0], I guess

[agile owl]

covariance = np.triu(np.eye(self.no_symbols))[np.triu_indices(self.no_symbols)]

[final kiln]

"It is true that for all A in Rmn and for all B in Rnp, (A+B)C = AC + etc ... "

[tidal bough]

triu(eye)? isn't that always just an eye?

[vernal quartz]

i got it here, the author called it matrix

[agile owl]

that's just the case where it doesn't have enough observations

also I need to remove the triu call

I just need the indices

but yeah there's a triu_indices method too that's handy

[tidal bough]

purpose of whom

[final kiln]

Do you need to keep any order when flattening ?

[agile owl]

no

as long as it's consistent every time

I am using np.triu_indices to index it

[final kiln]

Then just sort them and slice from the index of the first Nan

[agile owl]

nah that won't be consistent

[final kiln]

Wdym you'd be literally sorting them

[vernal quartz]

λ ∈ R. Let A ∈ Rm×n any idea of difference between R and Rmxn? this scalar λ in anything?

[agile owl]

the mapping from i, j of the covariance matrix to the flattened array can be in any order, but it has to be the same mapping every time

[final kiln]

Ah you can just kinda use index trickery for it I think

[left tartan]

But it says that A is a matrix, it’s just saying for any matrix of real numbers

Because the dimensions aren’t important to these properties

[vernal quartz]

λ ∈ R you mean this scalar is any real number?

[final kiln]

https://numpy.org/doc/stable/reference/generated/numpy.triu_indices.html

I used this not long ago

[left tartan]

You’re mixing multiple links here but that’s what that says yes

[vernal quartz]

i got it

[iron basalt]

This is all just describing the basic ways in which matrix algebra works.

You could list all this with real numbers, or whatever you are used to, and it will look similar (list of rules).

The first line here, I can show you something similar you are familiar with: a(1/a) = 1 = (1/a)a.

[vernal quartz]

thanks

is analytic gemotrey and matrix decomposition important?

[iron basalt]

Few things are more important.

[vernal quartz]

?

[iron basalt]

Very important.

[vernal quartz]

https://mml-book.github.io/book/mml-book.pdf can i change a book

this book feels off

i need a pre-education book

what s is this, its even write weird

why it needs to be twisted like that

[left tartan]

What’s twisted? Have you taken stats before?

Bayes theorem is in college stats, right?

[final kiln]

That's why I suggested you to consider a college degree earlier, a two month long course can't possibly have given you the tools to do ML properly

[iron basalt]

Yes, should be in undergraduate.

[vernal quartz]

i took

[left tartan]

I went to college decades ago. This stuff isn’t fresh in my mind, but it’s just part of my base knowledge.

[iron basalt]

In a review of probability before really getting to the actual stats part.

At the start of the first semester.

[vernal quartz]

thanks sharing, i did bussiness analytic degree

[vernal quartz]

i'm not sure that's ture, my first semester stats is 83

[final kiln]

And you haven't learned neither Bayes theorem nor linear regression ?

[left tartan]

We’re just trying to figure out what you need to review: as (mostly?) CS students, we take some of this stuff for granted.

[final kiln]

Business analytics (BA) is a set of disciplines and technologies for solving business problems using data analysis, statistical models and other quantitative methods.

[vernal quartz]

tell me everything

[agile owl]

the applications of bayes theorem aren't always immediately obvious

everyone should remember the basic idea though

[left tartan]

A good stats background is definitely important for DS

[vernal quartz]

where i should start

i guess i passed high school level

[left tartan]

The three important courses: Linear Algebra, Stats (calculus based/engineering stats), and hmm, discrete Id guess

[final kiln]

I may be missing something, but this looks like a good start

[vernal quartz]

linear Algenra i got it, so Stats

[iron basalt]

Discrete mathematics / proof writing / mathematical notation seems missing. Counting and probability.

[left tartan]

Also review the progression section here: https://en.m.wikipedia.org/wiki/Mathematical_maturity

And understand you’re just at the first step… it’s ok, just appreciate that there’s more ‘maturity’ to develop

[vernal quartz]

i remember i passed the exam with this

so where should i start

[final kiln]

With the basics

[left tartan]

Sounds like a business stats course (non calculus based)

Many colleges have two stats courses, one for engineers and one for non-engineers

[lapis sequoia]

I have a PPO implementation with Actor and Critic both having 2 Dense layers 256 each with RelU activation and an output Dense layer with size 2 I thik with softmax. Unfortunately the average score for Cartpole v0 peaks at 100 and I don't know why. Can anybody advise if my layers are good enough?

[vernal quartz]

i remember i stopped here, bootstrap

[final kiln]

Is this reinforcement learning ?

[lapis sequoia]

yeah PPO reinforcement learning

[vernal quartz]

do not discriminate business degree student, we are smart ass

[final kiln]

Ah, I haven't done any of it, if you could put in terms a analogous to supervised learning I might be able to help

But if it plateus it might be that the model is at capacity and you gotta make it bigger

[agile owl]

what's a good reward on cartpole

[lapis sequoia]

possibly, I'm very new to reinforcement learning, never made a model before

[agile owl]

I spend most of my time comparing rewards on my own custom environments not benchmarks

[final kiln]

I never touched it, gradient man might be able to help tho

[agile owl]

so what is a good reward on cartpole

what are you expecting

[vernal quartz]

thoughts?

[lapis sequoia]

i'm using gym, so assuming time survived, the higher the better

[final kiln]

If you can, I strongly suggest a college degree.

[agile owl]

are you using sb3 for the PPO

[lapis sequoia]

there's very sparse documentation of gym environments to give concrete answers

[iron basalt]

I think you seem to be struggling with the notation, so anything that gets you used to that. One such thing is a discrete mathematics book, often used to get people in CS used to it (and it's relevant to writing algorithms).

[agile owl]

I mean you can see benchmarks for cartpole on a lot of sites about RL

I just don't remember what a good score is

[vernal quartz]

stopped here, bootstrap

[agile owl]

There is going to obviously be a limit to how far you can get with a given environment

[lapis sequoia]

is this stable benchmarks 3?

[agile owl]

stable baselines

[lapis sequoia]

never used that library

[agile owl]

ah ok they have an implementation of PPO was just curious

[lapis sequoia]

i glued it together from https://www.youtube.com/watch?v=9DO63MSGeNA this video

[vernal quartz]

i think is my bussiness stats 108

[lapis sequoia]

also do you think stacking more dense layers in parallel into Actor and Critic would improve my score?

[iron basalt]

Being able to apply existing methods / tools is great, and what a business degree should focus on, but in this case you are trying to understand how those work fundamentally, at a mathematical level. To do so you need to be able to read something like the Bayes theorem notation. This can be learned by any number of mathematical books on probability. But also you want to be comfortable with the set notation earlier, and the linear algebra notation. In addition, ML makes heavy use of calculus.

[agile owl]

deeper networks don't help on cartpole

but wider ones do

[lapis sequoia]

any types in particular apart from Dense?

[agile owl]

not afaik

[lapis sequoia]

could an LSTM layer help? since you need to balance a cartpole and previous information could be useful

[agile owl]

there's people researching different network architectures for policy networks

[vernal quartz]

yes, i unsderstand, i want to figure out where i should get rerestarted, i already got linear algebra, so calculus then?

[agile owl]

there is a version of PPO with an LSTM layer but empirical results aren't that great in most applications compared to simple framestacking

[final kiln]

What do you mean when you say that you have linear algebra ?

[agile owl]

I definitely don't have linear algebra I need to get the linear algebra brought in

[left tartan]

Linear, Discrete and Stats are the holy trinity for CS… calculus is assumed, altho I don’t think is as important, but part of the assumption is that you’ve grinded through years of math and developed the math maturity to cope.

[vernal quartz]

what level stats is this?

[final kiln]

I'm probly biased, but I think calculus is the most important cuz of gradient descent

[lapis sequoia]

what effect does widening layers has as opposed to making them deeper?

[past meteor]

I'm honestly less hung up about what method belongs in what box

[agile owl]

I don't really have an interpretation for you besides the actual thing

[left tartan]

I’m conflicted on this point: I think it’s simplest to explain the eli5 of this, but maybe only because of a few years of calc? I dunno

[agile owl]

they lead to different results though

[iron basalt]

To what extent do you understand linear algebra? Have you done matrix factorization? Calculus, yes (including multivariate (cross over with linear algebra)). Set theory (set notation is everywhere). Probability (and counting) theory. Statistics (after probability and calculus review) with calculus in it (and set notation again). Then back to that ML book.

[lapis sequoia]

for example if I were to process images and stack conv2d layers, would making them wider help network learn better?

[agile owl]

do they teach people statistics without calculus?

[past meteor]

Yes

[agile owl]

yikea

[left tartan]

Business stats has no calc. Most colleges offer it to non engineering majors

[lapis sequoia]

calculus is assumed a lot of the time

[past meteor]

My SO took 7, yes seven stats classes

[agile owl]

I took statistics for economics and it had calculus

[past meteor]

And they only gave her math after like 4 or 5?

[left tartan]

AP stats, for instance, has no calc

[lapis sequoia]

we had to do real analysis in 1st year before diving deep into statistics

[past meteor]

I studied business engineering. My path was a lot more linear. Year 1 had lin alg in semester 1, with some programming course, semester 2 had calculus.

[lapis sequoia]

i argue real analysis is calculus on steroids

[agile owl]

I didn't take a real real analysis course I took a course called mathematics for economics that was about optimizing in convex spaces and hessians and all that jazz

[past meteor]

Year 2, semester 1 brought stats

Year 3 brought econometrics and so on

I think econometrics was actually the most valuable data course I took because it was the one that taught you the "finesse" of working with data

That's so missing in most

[lapis sequoia]

i miss the funny questions like can an infinite union of countably infinite sets be countably infinite

[iron basalt]

If i'm being honest, with regards to what you are trying to get into, level 0.

[agile owl]

I feel like my econometrics course was much ado about multiple linear regression

[left tartan]

That’s something I find myself repeating a lot: that ‘calculus’ is the beginning not end of the math journey. Students were built up to believe that calculus was the pinnacle, rather than the pit stop at the foot of the mountain.

[past meteor]

Yeah, it started slow but ramped up. By the end we were covering logit/probit and so on in detail

By my masters I was exclusively taking CS faculty ML courses and they felt half as rigorous. Many of my cohort were lacking that "finesse"

[lapis sequoia]

did anyone do Measure Theory?

[iron basalt]

Probability and stats don't really work without it.

[left tartan]

Hah, I took a stats class in my CS masters and it kicked my ass. All the Cs majors were dying.

Like, half the class was stats majors

[lapis sequoia]

measure theory was breaking my brain. It's like real analysis for stats

[vernal quartz]

just plane, multi planes and dot product, maybe more but i don't remember

[past meteor]

They were so complementary. I took a course called "data mining" and there we had to do logistic regression by hand

[iron basalt]

It's great, and then you can look at all the stats online and realize most of it is busted.

[agile owl]

I feel like all the courses I've taken have been for stupid people that's the only way I could have passed them

[vernal quartz]

Probability i remember i did this

[past meteor]

It was a nice eye opener to do SGD, chain rule etc. to really get it

[iron basalt]

Yeah, so linear algebra goes really deep, leading into / part of abstract algebra.

[vernal quartz]

where i should go

[past meteor]

The previous semester, econometrics, was really all about residual analysis and "modelling" moreso than "Yeah SGD is scalable!!!" (which was more the CS faculty approach)

[iron basalt]

It's probably the most applicable branch of math other than calculus (maybe even tied with it).

[lapis sequoia]

I did a lot of measure theory in terms of proving theorems and answering weird questions, but we never applied it in machine learning, so I forgot everything haha

[left tartan]

Doesn’t matter, pick any one topic you don’t know and learn it. Repeat until you know everything (joke).

[agile owl]

I learned backprop really well once and I have completely forgotten the actual math behind it tbh

[vernal quartz]

anyone give me a road map for math

[agile owl]

I forget everything

[left tartan]

You can learn in any order, linear, stats, discrete, calc

[lapis sequoia]

school->college->uni

[agile owl]

I must have done too many drugs or something

[past meteor]

I'm never going to forget, data mining had this one exam question that was basically

What happens to the weights of a logistic regression if you have no regularization and the classes are perferctly linearly seperable. Answer in 15 words or less

[vernal quartz]

i think i learned probabilty, and my stats i don't know what level is it

[lapis sequoia]

completely depends on your level

[past meteor]

I remember they go to infinity and I used L'hopital's rule to prove it

I don't remember why tho

[agile owl]

ah L'Hopital's rule

[lapis sequoia]

which areas of maths are you interested in?

[agile owl]

I remember using that

[left tartan]

They’re trying to build up to broad understanding of ML/DS

[vernal quartz]

i interested whatever the shit is for me to build a deep leanring model comfortably

[lapis sequoia]

coding alongside linear algebra

[past meteor]

My 2 cents, if you're like most people then you should iteratively go between math and coding

[agile owl]

I'm gonna just admit it it's a real problem that I don't do anything by hand because I learn it all and then I forget how it works and at the end i'm no different than someone using a library they don't understand

[vernal quartz]

deep learning is not coding

[past meteor]

I'd not cope with doing math for the sake of it

[lapis sequoia]

calculus is important

[iron basalt]

Calculus, linear algebra, multivariate calculus, discrete, probability, statistics. Maybe this order is alright. Thoughts anyone?

[agile owl]

I did multivariate calculus before linear algebra believe it or not

[lapis sequoia]

would real analysis be useful?

[iron basalt]

Sounds rough.

[agile owl]

I just went calculus -> multivariate calculus -> linear algebra

[past meteor]

I don't think we should overstate the amount of math you need tho

[iron basalt]

Yes, but at the end, it's like level 2. You don't really need it, but it will take your understanding to the next level.

[agile owl]

Actually I found multivariate calculus relatively easy compared to Linear

[past meteor]

What it mostly gives you is certainty that you're not doing stuff incorrectly

[agile owl]

I think Linear algebra requires a completely different part of your brain or something

[past meteor]

I think the part you should really get as a practice focused ML professional is how to evaluate models correctly

[vernal quartz]

so for math, should i calculate everything by myself, after leanred, or just watching it

[agile owl]

because I was fine at calculus but I suck at linear algebra

[past meteor]

But I always say this 😄

[lapis sequoia]

i crammed all the exams with energy drinks and 7 days of revision so i probably did multivatiate calculus, cant remember xD

[left tartan]

Yah. But that was the third in your series of calc courses too

[vernal quartz]

for math, should i calculate everything by myself, after leanred, or just watching it

[iron basalt]

Implement stuff in code. For example, can you do Gauss-Jordan elimination in Python?

[past meteor]

If you know how to evaluate models correctly, that means you could technically do no harm in using the models incorrectly if you don't know how they work

[lapis sequoia]

you need to attempt as many problems as physically possible

[vernal quartz]

i don't know what that shit is.

[lapis sequoia]

dpeends if you are applied, visual, or whatever learner

[vernal quartz]

i only know gussian

[lapis sequoia]

even I dont know what that is lmao

[past meteor]

I couldn't code guass-jordan elimination 🤷

I know what it is

But idt it's particularly relevant to be honest

[iron basalt]

Being able to go from math to code is a great skill and will also solidy your understanding of the math too.

[past meteor]

It's the same story with CS proper

Not every programmer needs to be a systems programmer

[vernal quartz]

i don't question my coding skills, i just don't understand math

[lapis sequoia]

oh is that where you subtract rows and columns from each other?

[vernal quartz]

how to calculate them

[iron basalt]

Not that gauss-jordan in particular is important to implement, but being able to do stuff like that makes you very valuable labor.

[past meteor]

In that vein, not every ML professional needs a deep understanding of math

[vernal quartz]

valualbe what?

[past meteor]

A working knowledge is enough for the vast majority of roles

[iron basalt]

Everyone will want to employ you.

[agile owl]

in my experience people care more about your experience doing commercial things

[vernal quartz]

i will let every recruiter know i know guass-jordan elimination

[iron basalt]

Yes, but with enough experience, you become the number one target for hiring. People with these math + coding skills are highly valuable (and rare).

[past meteor]

I disagree tbh

This is only true for some roles

[lapis sequoia]

what area of maths is Kullback–Leibler divergence in?

[past meteor]

This is handy to know, intuitively

Information theory

[iron basalt]

Yeah. I'm kind of talking within CS roles here.

[left tartan]

If you’re aiming for a DS role, you’re certainly expected to ‘know’, not just ‘do’

[past meteor]

We have several CS/Physics PhDs at work

[vernal quartz]

Calculus, linear algebra, multivariate calculus, discrete, probability, statistics. can you order it from high scholl to uni? i want to know here to start

[past meteor]

And I'm sure if I roll in tomorrow and ask everyone to code up Gauss-Jordan elimination