Alignment with human representations supports robust few-shot learning

Authors: Ilia Sucholutsky, Tom Griffiths

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We confirm this prediction empirically, finding such a relationship in an analysis of the performance of 491 computer vision models. We also show that highly-aligned models are more robust to both natural adversarial attacks and domain shifts. Our results suggest that human alignment is often a sufficient, but not necessary, condition for models to make effective use of limited data, be robust, and generalize well.
Researcher Affiliation Academia Ilia Sucholutsky Department of Computer Science Princeton University is2961@princeton.edu Thomas L. Griffiths Departments of Psychology and Computer Science Princeton University tomg@princeton.edu
Pseudocode No The paper contains mathematical definitions and proofs (Lemma 3.7, Proposition 3.8, Theorem 3.9, etc.) and discusses models and experiments, but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes All code and full results data are provided as part of the supplemental information. We will share them publicly after the anonymity period is over.
Open Datasets Yes All of the models used in this paper were pre-trained on Image Net-1k [31] and had their performance evaluated on the Image Net-1k validation set. Few-shot transfer learning performance was evaluated on the CIFAR100 [18] test set with n {1, 5, 10, 20, 40, 80} examples per class used for few-shot learning and the remaining examples used for evaluation. We measure alignment by computing the three metrics described below on the six image datasets from Peterson et al. [29]... We now test few-shot transfer learning performance on the MNIST [21] and FMNIST [42] datasets
Dataset Splits No All of the models used in this paper were pre-trained on Image Net-1k [31] and had their performance evaluated on the Image Net-1k validation set. Few-shot transfer learning performance was evaluated on the CIFAR100 [18] test set with n {1, 5, 10, 20, 40, 80} examples per class used for few-shot learning and the remaining examples used for evaluation.
Hardware Specification Yes All experiments were conducted on an AWS x1.16xlarge instance (no GPUs).
Software Dependencies No For linear probing, we take the embeddings coming from the penultimate layer of a model and fit a logistic regression using scikit-learn [28]. For the classifier heads, we use a oneand two-hidden layer neural network implemented in Py Torch [27]... compute partial correlation with Image Net-1k Top-1 validation accuracy as a covariate using the Pingouin Python package [39].
Experiment Setup Yes For the classifier heads, we use a oneand two-hidden layer neural network implemented in Py Torch [27], both with an input dropout rate of 0.8 and the latter with a Re LU hidden layer. The heads are provided in the supplement and fixed hyperparameter values (selected based on exploratory experiments to confirm stable convergence) were used for every dataset/model for consistency.