Dynamic Neural Network Structures Through Stochastic Rewiring
summary
Argues that constant, largely stochastic synaptic rewiring implements a sampler over network parameters so circuits sample from an objective distribution rather than converge to fixed weights; supports this with a stochastic differential equation/Fokker–Planck analysis and links to an Ornstein–Uhlenbeck fit to spine fluctuations and the empirical estimate that >50% of synaptic change is unexplained by determinism (pairwise synapse correlation 0.23). Builds a reward‑modulated spiking model that reproduces spine addition/elimination and rapid compensation only with rewiring, and a DeepR backprop variant that matches full networks while using 15% connectivity (near‑full accuracy).
AI-written summary · about
chapters (12)
- 0:00Introduction
- 3:08Implications
- 4:41plasticity
- 7:27noise
- 10:07Rewiring
- 15:03Analysis
- 18:57Rewardbased learning
- 21:41Experiments
- 24:31Dynamics of Parameters
- 27:36Initial Learning
- 37:21Conclusions
- 38:44Questions
related talks
01h 13mComputation in Networks of Neurons in the Brain II
27mWhat can artificial neural networks learn from biological neuromodulatory systems?
43mCoin-Flipping in the Brain: Probabilistic Computation and Learning in the Assembly Model
12mComputation with sequences of neuronal assemblies
32mHow Neuroscience Makes AI Intelligent
31mNeuromorphic computing in spiking neuron-astrocyte networks
36mBrain's Neural Network as a service
48mDynamic Neural Networks with DyNet