Backpropagation and Deep Learning in the Brain
summary
Argues that brains likely send detailed, neuron-specific feedback for credit assignment and that some nonstandard algorithms can approximate backprop: feedback alignment uses a fixed random backward matrix B and yields near-backprop updates on small tasks (MNIST, SVHN) as forward weights adapt, but difference target propagation (learned imperfect inverses) matches backprop on MNIST yet collapses on CIFAR and cannot scale to ImageNet in locally connected nets — overall empirical verdict: random or learned feedback helps on small datasets but ImageNet failure shows current biologically motivated replacements still fall short.
AI-written summary · about
chapters (23)
- 0:00Intro
- 0:48The credit assignment problem
- 1:46The solution in artificial networks: backprop
- 3:50Why Isn't Backprop "Biologically Plausible"?
- 6:51Neuroscience Evidence for Backprop in the Brain?
- 8:03A spectrum of credit assignment algorithms
- 13:49How to convince a neuroscientist that the cortex is learning via [something like] backprop
- 16:04What about reinforcement learning?
- 16:43A Single Trial of Reinforcement Learning
- 17:20Measuring Outcomes
- 17:47Update Parameters with the Policy Gradient
- 18:41Training Neural Networks with Policy Gradients
- 20:46The backpropagation solution (AKA 'Weight transport)
- 21:49Feedback alignment
- 31:03Energy based models.
- 32:30Question
- 32:56Constraints on learning rules.
- 34:13Target propagation
- 37:09Gradient free DTP variants
- 38:56Performance on ImageNet
- 41:05New Models of a Neuron
- 46:48Future Directions
- 49:04Difference target-propagation (DTP)
related talks
32mHow Neuroscience Makes AI Intelligent
01h 51mUnderstanding Neural Networks (Somewhat)
47mInterpreting Deep Neural Networks (DNNs)
12mLightning Talk: Backpropagation-Free Optimization in PyTorch
41mArtificial Intelligence: Let's Think About Learning
29mDeep Learning, Neuroscience and the future of AI
27mUnderstanding Deep Neural Networks
27mDemystifying Deep Neural Networks