NoticeThis site demonstrates one possible use of this domain. For acquisition, partnership, or investment inquiries, please use our contact form.

Cognition & Neuroscience

Learning and memory

How biological learning works — and what carries over to machine learning.

Brains and neural networks both learn from experience, but the mechanisms are different enough that the analogy can mislead. Here is what biology actually does, and where the machine learning comparison holds and breaks.

Hebbian plasticity and beyond

The classic slogan — 'cells that fire together, wire together' — is Donald Hebb's 1949 proposal. Modern neuroscience has refined it into spike-timing-dependent plasticity (STDP) and many neuromodulator-gated rules: dopamine, acetylcholine, noradrenaline, and serotonin each gate plasticity in different ways depending on context, reward, novelty, and arousal.

These rules are local: each synapse updates based on signals available at it. There is no biologically realistic mechanism for the global backpropagation that trains artificial networks.

Working memory and long-term memory

Working memory holds a small amount of information actively, typically for seconds. It is implemented partly by persistent activity in prefrontal cortex and partly by short-term synaptic facilitation. Capacity is famously small — four items, give or take.

Long-term memory is consolidated more slowly. Episodic memory (specific events) is bound by the hippocampus, then gradually re-encoded in cortex during sleep — a process called systems consolidation. Semantic memory (facts), procedural memory (skills), and emotional memory each have their own circuits.

What ML borrows — and what it doesn't

Artificial neural networks borrow the metaphor of weighted connections, but the learning algorithm — gradient descent via backpropagation — is mathematical, not biological. There is no global error signal travelling back through synapses in the brain.

Recent work on biologically plausible learning rules (equilibrium propagation, predictive coding networks, target propagation) tries to close that gap, but none has yet matched gradient descent on hard tasks. The honest answer is that we still do not know how the brain assigns credit.

Sample efficiency

Children learn the meaning of a new word from one or two exposures. Frontier language models need orders of magnitude more text than any human will ever read. The gap is the deepest open problem in AI: nobody has a clean story for why biological learning is so sample-efficient.

Hypotheses include rich inductive priors from evolution, embodied multi-sensory grounding, active curiosity-driven exploration, and the use of episodic memory for fast one-shot binding. AGI will likely need some combination.

Key terms

Hebbian learning
Synaptic strengthening when pre- and post-synaptic neurons co-activate.
STDP
Spike-timing-dependent plasticity; relative spike timing decides strengthen vs weaken.
Consolidation
Gradual transfer of memories from hippocampus to cortex, mostly during sleep.
Backpropagation
Gradient-based credit assignment used to train artificial networks.
Sample efficiency
How much data is needed to learn a task.

Connects to AGI

Closing the sample-efficiency gap is widely seen as a prerequisite for AGI. Whether the fix comes from architecture, pre-training tricks, embodiment, or new learning rules is one of the live arguments of the field.

Continue reading