A Spectral Theory of Neural Prediction and Alignment

Authors: Abdulkadir Canatar, Jenelle Feather, Albert Wakhloo, SueYeon Chung

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

Reproducibility Variable Result LLM Response
Research Type Experimental We test a large number of deep neural networks that predict visual cortical activity and show that there are multiple types of geometries that result in low neural prediction error as measured via regression.
Researcher Affiliation Collaboration 1Center for Computational Neuroscience, Flatiron Institute 2Center for Neural Science, New York University 3Zuckerman Institute, Columbia Univeristy
Pseudocode No The paper describes mathematical formulations and processes like ridge regression, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures.
Open Source Code Yes Code for analyzing the empirical and theoretical neural prediction error can be found at https://github.com/chung-neuroai-lab/SNAP.
Open Datasets Yes The neural datasets were previously collected on 135 texture stimuli for V1 and V2 [34], and on 3200 object stimuli with natural backgrounds for V4 and IT [35] (Fig. 1A). Both datasets were publicly available and obtained from [11].
Dataset Splits No The paper states 'we used a 60/40 train-test split (p = 0.6P) in the experiments below.' While a train-test split is specified, a separate validation split for hyperparameter tuning or early stopping is not mentioned.
Hardware Specification Yes All experiments were conducted on single Nvidia H100 or A100 GPUs with 80GB RAM using Py Torch 2 [65] and JAX 0.4 [66] libraries.
Software Dependencies Yes All experiments were conducted on single Nvidia H100 or A100 GPUs with 80GB RAM using Py Torch 2 [65] and JAX 0.4 [66] libraries.
Experiment Setup Yes We performed ridge regression... using a ridge parameter of αreg = 10−14, and also calculated the theoretical neural prediction error for each using Eq. 3. Given that Eq. 3 is only accurate in the large P limit in which both the number of training points and the number of test samples is large (see Sec. SI.1 and Fig. SI.4.2), we used a 60/40 train-test split (p = 0.6P) in the experiments below.