Strong Generalization from Small Brains and No Training Data
summary
Argues small brains form low-power geometric models and demonstrates concrete algorithms: a resonator network that couples two Hopfield-style attractors to factor images into object and transformation, and complex phasor encodings that implement a residue number system for large-range position coding. Uses calcium imaging and the FlyWire connectome as motivation, cites 139,355 fly neurons and a 200 billion‑transistor GPU contrast, shows simulations (capacity near quadratic, arithmetic over a 105 range) that are robust to heavy phase noise and outperform naive gradient descent.
AI-written summary · about
chapters (14)
- 0:00Introduction to neural systems
- 2:53Animal intelligence and vision
- 6:05Navigation and path integration
- 8:59Brain mechanisms in flies
- 11:25Connectomics and brain models
- 16:05Brain structure and learning
- 19:32Physics of computation in AI
- 22:27Merging data and physics
- 30:55Perception as factorization
- 35:41Phasor neural networks
- 42:05Resonator networks
- 44:32Residue number systems
- 52:14Efficient spatial coding
- 54:10Summary and discussion
related talks
01h 14mTissue vs silicone musings on the future of deep learning hardware and software
43mThe Biological Path Towards Strong AI
01h 16mRethinking Computing with Neuro-Inspired Learning: Algorithms to Hardware
41mBrain-inspired computational primitives for reliable neuromorphic intelligence
30mCracking the code: neuronal networks in the brain
45mMachine Learning, Deep Networks and Interpretability in Systems, Cognitive and...
36mBrain's Neural Network as a service
25mBig Brains, Small Targets: Whole-Brain Image Analysis with Python