Robust Feature Learning for Multi-Index Models in High Dimensions

Authors: Alireza Mousavi-Hosseini, Adel Javanmard, Murat A Erdogdu

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental As a proof of concept, we also provide small-scale numerical studies on Gaussian data to support intuitions derived from our theory. Additional experiments on real datasets are provided in Appendix E. [...] The first row of Figure 1 compares the performance of three different approaches. [...] As can be seen from Table 1, the STD + ADV training approach achieves a higher test accuracy compared to the ADV approach across all model architectures considered here.
Researcher Affiliation Academia 1University of Toronto, 2Vector Insitute, 3University of Southern California EMAIL,edu, EMAIL
Pseudocode Yes Algorithm 1 Adversarially robust learning with two-layer NNs. [...] Algorithm 2 Gradient-Based Feature Learner for Single-Index Polynomials (Oko et al., 2024, Algorithm 1, Phase I). [...] Algorithm 3 Gradient-Based Feature Learner for Multi-Index Polynomials (Damian et al., 2022, Algorithm 1, Adapted)
Open Source Code Yes The code to reproduce the results of Figure 1 and Table 1 is provided at: https://github.com/mousavih/robust-feature-learning.
Open Datasets Yes Additional experiments on real datasets are provided in Appendix E. [...] on the MNIST dataset (Le Cun et al., 1998)
Dataset Splits No To estimate the robust test risk, we fix a test set of 10,000 i.i.d. samples, and use 20 iterations to estimate the adversarial perturbation. [...] For both approaches, we use the corss entropy loss, a batch size of 64, a learning rate of 0.01 for both PGD and SGD updates
Hardware Specification No The paper does not provide specific hardware details used for running its experiments.
Software Dependencies No The paper mentions 'Py Torch initialization' but does not specify a version number or other key software components with their versions.
Experiment Setup Yes We implement adversarial training in the following manner. At each iteration, we sample a new batch of i.i.d. training examples. We estimate the adversarial perturbations on this batch by performing 5 steps of signed projected gradient ascent, with a stepsize of 0.1. We then perform a gradient descent step on the perturbed batch. [...] The student network has N = 100 neurons, and the input is sampled from x N(0, Id) with d = 100. [...] For both approaches, we use the corss entropy loss, a batch size of 64, a learning rate of 0.01 for both PGD and SGD updates, and we use ℓ norm to constrain perturbations, where pixels are normalized between 0 and 1.