Machine Learning, Deep Networks and Interpretability in Systems, Cognitive and...

on YouTube ↗45m

" summary

Argues the mammalian brain behaves like a dynamical deep network but neuroscience is severely data-limited, so two concrete methods are used: extract pretrained deep net layers as feature spaces and fit them with banded ridge regression to single voxels (example: 33 feature spaces, ~40k features regressed onto ~100k MRI voxels), or build an anatomically motivated model and train it on huge single-neuron datasets (300 V4 neurons, hours per neuron) — a 20,000-parameter architected model outperforms VGG and approaches usable prediction but still fails to reach the noise ceiling.

AI-written summary · about

" chapters (18)

1. 0:00Intro
2. 2:01Science and engineering have different priorities and these affect our choices (especially regarding deep learning)
3. 4:39The brain is a highly interconnected, dynamic network organized hierarchically, in parallel, and at multiple scales
4. 6:30A central problem in systems and cognitive neuroscience is to model the representation of information across the brain
5. 12:59The best classical way to model representation in the brain is to use some form of the GLM
6. 13:53Because brain data have low SNR, regularization must be carefully managed across different feature spaces
7. 14:29We use an encoding model approach to fit multiple navigation-related feature spaces to each voxel in each subject
8. 17:40We visualize the representation of navigation-related features by projecting voxelwise model weights onto the cortical surface
9. 18:32representations we use a low-dimensional embedding
10. 20:43PCA of these data reveals that navigation-related networks are organized into three main functional classes
11. 23:05Most of the variance in brain activity is explained by variables related to perceptual, motor and goal-directed behavior
12. 26:03Can we use deep networks to model brain data directly?
13. 27:29Can we use the learned weights of pre-trained deep network as a source of features for the GLM?
14. 29:41One problem with pre-trained deep networks is that their features are often correlated across layers
15. 30:29Can we develop a hybrid modeling scheme that uses deep learning to fit an explicit model?
16. 35:37To test this idea we analyzed and modeled long-term recordings from 302 area V4 neurons
17. 37:26To better understand visual representation in area V4 we analyzed the predicted optimal patterns (POPs) for each cell
18. 38:43Spatial, chromatic and temporal tuning of the V4 sample can be recovered from feature-specific embeddings

" related talks

• 47mInterpreting Deep Neural Networks (DNNs)Bin Yu · Frontiers of Deep Learning 2019
• 36mBrain's Neural Network as a serviceDhivya Raj · Converge 2022
• 01h 14mTissue vs silicone musings on the future of deep learning hardware and softwareNir Shavit · Hydra 2020
• 58mStrong Generalization from Small Brains and No Training DataBruno Olshausen · Unknown Futures of Generalization 2024
• 43mCan Cognitive Neuroscience Provide a Theory of Deep Learning...Ted Willke · MLconf 2016 Seattle
• 56mTutorial on Deep Learning IIRuslan Salakhutdinov · Foundations of Machine Learning Boot Camp 2017
• 48mNeural Networks and BrainsRaghav Subramaniam · Devoxx US 2017
• 29mDeep Learning, Neuroscience and the future of AIDarjan Salaj · Berlin Buzzwords 2021