Formation of Representations in Neural Networks

Authors: Liu Ziyin, Isaac Chuang, Tomer Galanti, Tomaso Poggio

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present experimental evidence that supports predictions resulting from the CRH. We also perform experiments to test mechanisms that break the CRH.
Researcher Affiliation Collaboration 1Massachusetts Institute of Technology 2Texas A&M University 3NTT Research
Pseudocode No The paper describes theoretical frameworks and experimental results but does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statement regarding the availability of source code or links to a code repository for the described methodology.
Open Datasets Yes res1: Res Net-18 (11M parameters) for the image classification; res2: Res Net-18 self-supervised learning tasks with the CIFAR-10/100 datasets. llm: a six-layer eight-head transformer (100M parameters) trained on the Open Web Text (OWT) dataset (Gokaslan & Cohen, 2019);
Dataset Splits Yes res1: Res Net-18 (11M parameters) for the image classification; ... We measure the covariances matrices with data points from the test set. res2: Res Net-18 for self-supervised learning tasks with the CIFAR-10/100 datasets.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No The paper mentions optimizers like SGD and Adam, and activation functions like ReLU, but does not provide specific software dependencies with version numbers (e.g., PyTorch 1.9, Python 3.8).
Experiment Setup Yes fc1: ...depth of the network (D = 4), the width of the network (d = 100), weight decay strength (γ = 2 10 5), minibatch size (B = 100). fc2: ...SGD with a learning rate of 0.1 with momentum 0.9 and γ = 10 4 for 105 steps... batch size of 100. res1: ...train with SGD with a learning rate 0.01, momentum 0.9, cosine annealing for 200 epochs, and batch size 128.