Capacity of Neural Networks for Lifelong Learning of Composable Tasks
summary
Proposes a concrete neural circuit model using a random graph spike-timing model with parameters N (neurons), D (fan‑out), K (synaptic strength) and R (neurons per concept), runs simulations showing simple association capacity falls to ~3,000 before degradation but reaches ~250,000 if strong synapses or a stabilizing hippocampal allocator (feedforward allocator) are added, and proposes optogenetic/in‑circuit stimulation experiments to test one‑shot composable association learning.
AI-written summary · about
chapters (18)
- 0:00Intro
- 0:50Computational Systems have Many Levels of Description.
- 2:03Light Bulb Problem
- 4:11Symbolic Processing
- 6:07The Four Primitives
- 9:46Neuroidal Model as Programming Language . Directed graph, pikes for communication
- 13:09Neuroidal Model as Resource Model
- 17:01Representations
- 18:00Type II Add associations between represented concepts
- 18:46Some Large Scale Simulations
- 22:40Seven Requirements For Associations
- 24:29Basic Mechanism for Association
- 26:19Complexity analysis for realizing Association
- 29:38Alternatives to Strong Synapses
- 31:20Hippocampus
- 32:08Some Telltale Predictions
- 33:28Possible Direct Validation of Particular Systems Level Primitives
- 36:16Novel Opportunities for Experimentation
related talks
43mCoin-Flipping in the Brain: Probabilistic Computation and Learning in the Assembly Model
57mWhat Should a Computational Theory of Cortex Explain?
01h 13mComputation in Networks of Neurons in the Brain II
12mComputation with sequences of neuronal assemblies
43mCan Cognitive Neuroscience Provide a Theory of Deep Learning...
58mStrong Generalization from Small Brains and No Training Data
27mWhat can artificial neural networks learn from biological neuromodulatory systems?
41mBrain-inspired computational primitives for reliable neuromorphic intelligence