Joint processing of linguistic properties in brains and language models

Authors: SUBBAREDDY OOTA, Manish Gupta, Mariya Toneva

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

Reproducibility Variable Result LLM Response
Research Type Experimental We investigate this correspondence via a direct approach, in which we eliminate information related to specific linguistic properties in the language model representations and observe how this intervention affects the alignment with f MRI brain recordings obtained while participants listened to a story.
Researcher Affiliation Collaboration Subba Reddy Oota1,2, Manish Gupta3, Mariya Toneva2 1Inria Bordeaux, France, 2MPI for Software Systems, Saarbrücken, Germany, 3Microsoft, India
Pseudocode No The paper includes a schematic diagram (Figure 1) to illustrate its approach but does not contain any sections or figures explicitly labeled "Pseudocode" or "Algorithm", nor structured steps formatted like code.
Open Source Code Yes We make the code publicly available1. 1https://github.com/subbareddy248/lingprop-brain-alignment
Open Datasets Yes We use a dataset of f MRI recordings that are openly available (Nastase et al., 2021) and correspond to 18 participants listening to a natural story. ... We analyze the Narratives-21st year dataset (Nastase et al., 2021), which is one of the largest publicly available f MRI datasets (in terms of number of samples per participant).
Dataset Splits Yes Cross-Validation The ridge regression parameters were fit using 4-fold cross-validation. All the data samples from K-1 folds were used for training, and the generalization was tested on samples from the left-out fold. ... the best λ was chosen by tuning on validation data
Hardware Specification Yes All experiments were conducted on a machine with 1 NVIDIA GEFORCE-GTX GPU with 16GB GPU RAM.
Software Dependencies No The paper mentions using specific software components like Hugging Face (for BERT and GPT-2) and Stanford CoreNLP stanza library, but it does not provide specific version numbers for these software dependencies, only stating general names.
Experiment Setup Yes The ridge regression parameters were fit using 4-fold cross-validation. All the data samples from K-1 folds were used for training, and the generalization was tested on samples from the left-out fold. ... We used banded ridge-regression with the following parameters: MSE loss function, and L2-decay (λ) varied from 101 to 103; the best λ was chosen by tuning on validation data; the number of cross-validation runs was 4.