Interpreting Deep Neural Networks (DNNs)
summary
Transfer-learning from ImageNet conv layers plus simple ridge/LASSO regression predicts single-neuron V4 firing rates at state-of-the-art levels across 18 model variants, so early CNN features generalize to macaque V4; maximizing fitted models with consensus smoothed DeepTune yields stable curvature-like receptive fields (patterns survive 50% filter pruning) and guides closed-loop stimulus design. A separate method, agglomerative contextual decomposition, clusters phrase/patch contributions for LSTMs and CNNs and wins a small human judgment study versus integrated gradients and occlusion, though adversarial attacks remain a concern.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 7:06Interpreting deep learning in neuroscience
- 28:18General interpretability framework
- 33:41Hierarchical interpretation for language and images
- 41:36Summary
related talks
45mMachine Learning, Deep Networks and Interpretability in Systems, Cognitive and...
54mIntroduction to Understanding Deep Learning
56mTutorial on Deep Learning II
52mTransfer Learning and Finetuning Deep Convolution Neural Network
35mDeconstructing Deep Learning
27mDemystifying Deep Neural Networks
27mUnderstanding Deep Neural Networks
37mWhat's new in Deep Learning?