#How to identify gaps in research?

9 messages · Page 1 of 1 (latest)

haughty bolt
#

You could make a situation model and apply link prediction...

#

How high the fruit is a matter of precision in a measure of barrier to entry

chilly pike
#

Lots of topics are left. The problem is that you cannot publish failed approaches so you have to choose a suit of experiments with which you can game the result and get published.
Also, many of the glamourous topics require big compute for experimentation so it's not feasible.
Personally I would like to know if you could taken an image to image transformer, and make it multimodal by simple finetuning.
Basically, you would take a normal LLM dataset but instead of plaintext you'd convert them to white background images with that text written over them then finetune the image models on this naively. I fully expect this naive method won't be actually useful but if it works on some cherrypicked examples then there's something to show.

#

Recently, in chat someone showed that Gemini can now colour manga. You could try experimenting and look for gaps. Then try finetuning gemini to fix that small gap. And that could be a paper as well.
LLM is a monster. You can use it for automation in many sectors and because of that there are lots of low hanging fruits, namely finetuning experients that can be done. Blocker for this is that you'd have to get your hands dirty and prepare the dataset on your own because all the public datasets have already been consumed.

atomic jasper
#

There are so many gaps?
Can you imagine the current state of models we have developed vs the gap in hardware

There is still a long way to go for something efficient and useful to run on a smaller compute

oblique frigate
#

Maybe conformal prediction or automl

obsidian gazelle
#

In my field, most papers specifically reference "ideas for future research"

rocky warren
#

I think there are plenty of gaps in RL. It was only until recently that it really became viable for real-world problems

coral sable
#
  1. Do a systematic mapping using PICO methodology on the topic, or 2. read systematic literature reviews on the topic, they always have a section like "future works" that talk about gaps in research