Networks of Spiking Neurons Learn to Learn and Remember
summary
New networks of spiking neurons add an adaptive firing thresholds term (simple decay-plus-spike increment) and use rewiring during training plus backpropagation-through-time with a damped pseudo-derivative, producing LSNNs that match LSTM performance on one-pixel-per-ms sequential MNIST after standard training. Meta-learning (outer loop) optimizes all synapses; a ~200-neuron LSNN trained over ~200–250k episodes learns a 4×5 maze (100-step episodes) to near-optimal reward and, in a supervised demo, learns an unknown input–output mapping within ~15 timesteps. Preprint is on arXiv.
AI-written summary · about
chapters (7)
- 0:00Introduction to LSNNs
- 1:50Limitations of spiking models
- 6:56Adaptive neuron dynamics
- 13:45Performance benchmarks
- 18:06Learning to learn paradigms
- 27:48Navigation and supervised tasks
- 47:48Discussion and Q&A
related talks
50mFlexible Neural Networks and the Frontiers of Meta-Learning
27mWhat can artificial neural networks learn from biological neuromodulatory systems?
29mDeep Learning, Neuroscience and the future of AI
47mInterpreting Deep Neural Networks (DNNs)
49mML: Teaching machine to learn from its own mistakes (...)
01h 50mNeural Networks: Walkthrough
15mSMT-Based Modeling and Verification of Spiking Neural Networks: A Case Study
32mHow Neuroscience Makes AI Intelligent