Research map / Learning theory and algorithms / Prediction and decision losses
Computing Linear Restrictions of Neural Networks
Matthew Sotoudeh, Aditya V Thakur
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use exact line restrictions to compute integrated gradients instead of relying on sampled estimates | Integrated Gradients on CIFAR10 networks · CIFAR10 | mean relative IG error (%): 24.95; 24.05; 45.34 | Mixed |
| Use trapezoidal sampling instead of left-Riemann sampling | Integrated Gradients on CIFAR10 networks · CIFAR10, convsmall | samples to reach 5% relative IG error: 136.74 -> 91.67 | Improved |
| Use trapezoidal sampling instead of left-Riemann sampling | Integrated Gradients on CIFAR10 networks · CIFAR10, convmedium | samples to reach 5% relative IG error: 150.31 -> 91.57 | Improved |
| Use trapezoidal sampling instead of left-Riemann sampling | Integrated Gradients on CIFAR10 networks · CIFAR10, convbig | samples to reach 5% relative IG error: 269.43 -> 222.79 | Improved |
| Use trapezoidal sampling rather than left or right sums | Integrated Gradients on CIFAR10 networks · CIFAR10 | samples to reach 5% relative IG error | Mixed |
| Compare adversarial directions with random directions using exact line partitions | Neural networks trained on MNIST · MNIST | linear-partition density: 1.36; 0.98; 1.22 | Mixed |
| Compare adversarial directions with random directions using exact line partitions | Neural networks trained on CIFAR10 · CIFAR10 | linear-partition density: 1.78; 1.67; 1.84 | Improved |
| Train with DiffAI instead of normal training | MNIST network · MNIST, FGSM direction | linear-partition density: 3.05 | Improved |
| Train with DiffAI instead of normal training | MNIST network · MNIST, random direction | linear-partition density: 2.50 | Improved |
| Train with DiffAI instead of normal training | CIFAR10 network · CIFAR10, FGSM direction | linear-partition density: 3.37 | Improved |
| Train with DiffAI instead of normal training | CIFAR10 network · CIFAR10, random direction | linear-partition density: 3.43 | Improved |
| Train with PGD instead of normal training | MNIST network · MNIST, FGSM and random directions | linear-partition density: 0.88; 0.80 | Mixed |
Closest papers by meaning
- SplineNets: Continuous Neural Decision Graphs (NeurIPS 2018 · similarity 0.87)
- Convex Geometry and Duality of Over-parameterized Neural Networks (JMLR 2021 · similarity 0.85)
- Piecewise linear activations substantially shape the loss surfaces of neural networks (ICLR 2020 · similarity 0.85)
- An Infinite-Width Analysis on the Jacobian-Regularised Training of a Neural Network (ICML 2024 · similarity 0.85)
- No Wrong Turns: The Simple Geometry Of Neural Networks Optimization Paths (ICML 2024 · similarity 0.84)
- Sharper Guarantees for Learning Neural Network Classifiers with Gradient Methods (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
