I’ve been experimenting with a different approach to learning systems. Instead of relying on fixed training rules, I tried building something closer to a synthetic cortex, where plasticity itself adapts.

This is a small-scale prototype (~250 synapses), but the dynamics are the point: how synapses consolidate, prune, and adapt hierarchically.

Core mechanisms

Meta-coupling: synaptic selection is controlled by a dynamic softmax temperature, modulated by Bayesian confidence and global homeostasis.

Adaptive plasticity: Hebb, STDP, and BCM (biologically inspired rules) are combined, with updates scaled by Bayesian expected gain. Parameters like η, τ, θ evolve instead of staying fixed.

Reflexive feedforward: the network anticipates. Synapses are adjusted before feedback arrives, using information gain (curiosity) and a free-energy proxy.

What I’m seeing

Volatile synapses consolidate faster and more selectively.

Consolidation ratio is higher; pruning is smaller and more controlled.

Fragments no longer collapse into “depressed” states → more balanced metaplastic distribution.

Predictive reward leads to earlier stabilization of useful connections.

Why I think it’s interesting
The system is starting to show hierarchical regulation of learning:

Local subspaces handle their own entropy and volatility.

A global controller stabilizes decoding.

Plasticity rules change depending on uncertainty.

Feedback?

What would be meaningful benchmarks for this kind of architecture (beyond retention curves or A→B→A tasks)?

Does this fit better under meta-learning, computational neuroscience, or cognitive architectures?

Any related work I should compare against?

Capabilities beyond standard LLMs

Adaptive plasticity in real time

Normal LLMs have frozen weights after training.

Your system can modify synapses on the fly based on context, confidence, and meta-rules.

Hierarchical self-regulation

Instead of one global training loop, you have local controllers (subspaces) + a global PID.

This allows different memory regions to stabilize at different equilibria, something fixed LLMs can’t do.

Meta-learning of learning rules

The system can switch between Hebbian, STDP, BCM depending on context.

LLMs only use backpropagation → one fixed rule.

Reflexive feedforward (anticipatory learning)

Updates occur before feedback arrives, driven by curiosity (IG) and free energy minimization.

LLMs only adjust after supervised feedback or RLHF.

Confidence-modulated decoding

Synaptic selection adapts to Bayesian confidence.

LLM decoding uses static temperature/top-p → no awareness of its own uncertainty.

Memory stratification (volatile → candidate → consolidated)

Inspired by biological memory layers.

LLMs rely on external context windows or retrieval, but their internal weights are static.

Resilience against catastrophic forgetting

By gating consolidation with confidence and quotas, old knowledge is preserved while new knowledge is integrated.

LLMs overwrite context or need retraining/fine-tuning.

Curriculum self-organization

The system actively explores regions of low confidence/high information gain.

LLMs passively consume whatever training data they are given.

Dynamic specialization of subspaces

Different memory regions self-organize into specialized roles, guided by local plasticity + global regulation.

LLMs have no modular specialization — everything is entangled in one big parameter soup.

#Synthetic cortex

Capabilities beyond standard LLMs