Studying Failure Modes of Deep Learning and Ways To Fix Them

on YouTube ↗59m

" summary

Shows that non-additive splitting noise (SSN) produces deterministic certificates and a 1/(2λ) L1 Lipschitz bound, doubling ImageNet-scale certified radii versus prior smoothing and improving clean accuracy; proves an additive-smoothing upper bound that collapses with dimension. Introduces a strong perceptual threat model (LPIPS-based) attack that explains why L∞ adversarial training (~50% robust on CIFAR-10) generalizes poorly to L2 (~20%), and releases Salient ImageNet (>1M images with core/spurious masks) to measure and reduce spurious-feature reliance.

AI-written summary · about

" chapters (23)

1. 0:00Intro
2. 0:29Recent Success of Deep Learning
3. 3:59Adversarial Attack Problem
4. 5:05Certifiable/Provable Defenses
5. 10:20Landscape of Provable Defenses A smoothed classifier
6. 17:36SSN Leads to Provable Robustness Theorem (Levine & F 20) f(x), a smoothed classifier with non additive splitting noise, is
7. 19:40Intuition behind SSN
8. 22:12Empirical Results on ImageNet . Our method establishes new state-of-the-art provable robustness with improved clean accuracy
9. 31:57Relationship Between Threat Models
10. 33:21Neural Perceptual Distances
11. 36:35Example Perceptual Attacks
12. 37:21Perceptual Evaluation
13. 41:52Some Thought Experiments
14. 43:17Foreground/Background Noise Sensitivity
15. 45:05Challenges & Prior Work • To study reliance of models on spurious features, we need large-scale datasets with rich feature annotations
16. 45:50Our Proposed Methodology
17. 47:00How to Visualize Neural Features?
18. 47:38Salient ImageNet
19. 48:25Core vs. Spurious Accuracy
20. 49:23Core/Spurious Accuracy of Pre-Trained Models
21. 50:57Summary
22. 54:44Model Reliance on Spurious Features We observe deep models heavily rely on spurious features in their predictions
23. 56:29Core Risk Minimization (CORM) Training models that rely more on core features

" related talks

• 47m1Provably Robust Deep Learning via Adversarially Trained Smoothed ClassifiersJerry Li · Frontiers of Deep Learning 2019
• 34mNeural networks: insane in the membraneCédric van Beijsterveldt, Chiel van de Steeg · J-Fall 2018
• 01h 24mAdversarial ExamplesSebastien Bubeck · Deep Learning Boot Camp 2019
• 44mAdversarial attacks on deep learning modelsKonda Reddy Mopuri · Anthill Inside 2017
• 48mProvable Robustness Beyond Bound PropagationZico Kolter · Frontiers of Deep Learning 2019
• 36mAdversarial Machine Learning: How Secure Is ML?Rudolf Mayer, Tanja Šarčević · sec4dev 2021
• 38mUsing adversarial samples to robustify your Neural Network ModelsIrina Vidal · Crunch Conference 2019
• 27mUnderstanding Deep Neural NetworksVladimir Osin · PyData Eindhoven 2020