The Brain as an Efficient, Robust Adaptive Learner
summary
Claim: networks of integrate-and-fire neurons learn and implement low‑dimensional dynamical systems using only local signals, by combining a fast inhibitory cancellation loop with slower predictive recurrent weights trained by a presynaptic × postsynaptic error feedback rule and an anti-Hebbian recurrent rule. Tools: spiking simulators, adaptive control proofs with a Lyapunov guarantee. Results: leaky integrator and nonlinear dynamics learned, speech spectrogram compressed with ~60 neurons, a walking controller still works after 70% inactivation, and estimation error scales like 1/N with spike budget.
AI-written summary · about
chapters (9)
- 0:00Start
- 6:18Problem 2: The assignment problem
- 7:26A pure top-down approach
- 10:03Objective
- 20:10Learning the recurrent connections
- 20:37From the point of view of a single neuron
- 28:55Why so variable?
- 35:22Folded network
- 36:17Microcircuit implementation
related talks
27mWhat can artificial neural networks learn from biological neuromodulatory systems?
32mHow Neuroscience Makes AI Intelligent
01h 13mComputation in Networks of Neurons in the Brain II
41mArtificial Intelligence: Let's Think About Learning
31mNeuromorphic computing in spiking neuron-astrocyte networks
29mDeep Learning, Neuroscience and the future of AI
01h 07mTheory of Computation II
41mBrain-inspired computational primitives for reliable neuromorphic intelligence