Deep Learning, Neuroscience and the future of AI
summary
Argues that spiking neural networks plus neuromorphic chips (Intel Loihi, SpiNNaker, BrainScaleS) already cut energy and latency by up to 100× on tasks like keyword spotting and image classification, and that combining them with an event-based camera enables drone stabilization in darkness and extreme spin where frame cameras fail. Describes three training routes: conversion from conventional nets, backprop through time on GPUs, and emerging biologically plausible, asynchronous learning rules that are still less accurate but promise scalable, low‑power online adaptation.
AI-written summary · about
chapters (16)
- 0:00Introduction to talk
- 0:28History of neural networks
- 2:15Deep learning explosion and challenges
- 4:24Von Neumann bottleneck in AI
- 6:31Neuroscience inspiration for AI
- 7:36Spiking neural networks explained
- 11:50Neuromorphic hardware progress
- 13:57Neuromorphic hardware performance
- 14:17Event-based sensors and drones
- 17:27Challenges and future of spiking networks
- 18:31Q&A introduction
- 19:18Practical implementation status
- 21:08Applications beyond drone stabilization
- 23:13Edge AI vs cloud AI
- 25:43Learning mechanisms in spiking models
- 28:08Summary and closing
related talks
32mHow Neuroscience Makes AI Intelligent
01h 16mRethinking Computing with Neuro-Inspired Learning: Algorithms to Hardware
24mFrom Neural Networks to Digital Brains: The Next Leap in AI
01h 57mIntroduction to Neuromorphic Computing
01h 14mTissue vs silicone musings on the future of deep learning hardware and software
03mRethinking software engineering in the AI era
41mDeep Learning to Solve Challenging Problems
43mThe Biological Path Towards Strong AI