Temporal Context in Brains and AI
summary
Argues that sequences and hours-long working memory arise from cortical traveling waves coupled to spike-timing plasticity: gamma packets (~100 ms) produce STDP windows (~±10 ms) that raise synapse strength 20–50% depending on pairing frequency, using mostly the 90% labile synapses. Engineers’ state space models (Toeplitz / Hippo-like matrices) produce traveling-wave dynamics, solve copy/add tasks perfectly and cut error by ~10^8 after ~600 training steps; large arrays (~100k neurons) and Mexican hat connectivity are needed and the models reproduce 0.1 m/s wave speeds.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:57Temporal context in AI
- 5:51Transformers and attention
- 9:07State space models
- 12:32Traveling wave simulations
- 17:11Waves in the brain
- 20:07Sleep spindles and memory
- 22:48Connectivity and waves
- 25:42Hippocampal modeling
- 29:30Learning without backprop
- 34:05Spacetime representations
- 35:43Memory mechanisms
- 42:56Concluding predictions
- 45:48Q&A session
related talks
01h 13mComputation in Networks of Neurons in the Brain II
46mThe Anatomy of Memory in Humans and AI Agents
01h 06mOn cognitive maps, LLMs, world models, and understanding
15mMemory-Augmented AI Agents: Enhancing Long-Term Reasoning and Context Awareness
27mMemory Engineering: Going Beyond Context Engineering
21mAhead of the Plasticity Curve: AI, Biology and Technology
09mSmale’s 18th Problem and Intelligence Science 2
47mWhat is Consciousness and Can Machines have it? Part 1