Predictive Coding Models of Perception
summary
Argues that training convolutional autoencoders with an LSTM bottleneck to do future-frame prediction yields internal codes that decode identity, motion and steering better with less data (outperforming a Comma reference CNN), can predict up to five frames at ~10 fps, and reproduce V1-like dynamics (transient on/off responses, end‑stopping, surround suppression) and the flash‑lag illusion. Implements a recurrent, layerwise PredNet that sends error maps forward; V1 recordings in darkness decode head pose and go away after M2 muscimol.
AI-written summary · about
chapters (7)
- 0:00Start
- 6:33Temporal Learning
- 16:09Examples
- 17:18Face-Recognition
- 25:21N Stopping
- 25:51Surround Depression
- 35:57There Is an Idea that We Could Go and Be Getting You Know Hundreds of Cells That Are Thousands of Cells at a Time You Know and Have a Prayer of Actually Doing those those Decoding Experiments and that's Work That We Have that's the Song Going but but You Know Right Now We'Re Just in the Stage Where We Said Can We Decode any Information and the Answer Seems To Be Yes I Agree There's a Lot of Questions That Come Downstream of that like Well You Know Predictive Coding Implies Certain Kinds of Correlational Structures That We Can Start To Look at We Also Have the Ability To Image Synapses in the Two-Photon
related talks
06mContrastive Predictive Coding for Human Activity Recognition
59mDeep Neural Networks for Video Applications
17mContent Aware Encoding for low latency live streaming encoders using deep learning
44mLearning from Machines
47mInterpreting Deep Neural Networks (DNNs)
37mLearning From Machines
41mLearning from multiple modalities, Predicting on unseen tasks
26mComputer Vision in the Era of Deep Learning