Observable Propagation: Uncovering Feature Vectors in Transformers
Authors: Jacob Dunefsky, Arman Cohan
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform a quantitative comparison between OBPROP and probing methods for finding feature vectors on diverse tasks (subject pronoun prediction, programming language detection, political party prediction). We find that OBPROP is able to achieve superior performance to these traditional data-heavy approaches in low-data regimes ( 4.3). |
| Researcher Affiliation | Academia | 1Department of Computer Science, Yale University, New Haven, CT, United States. Correspondence to: Jacob Dunefsky <jacob.dunefsky@yale.edu>, Arman Cohan <arman.cohan@yale.edu>. |
| Pseudocode | Yes | Then for a given observable n, the feature vector corresponding to sublayer l in P can be computed according to Algorithm 1. |
| Open Source Code | Yes | All code used in this paper is provided at https://github.com/jacobdunefsky/ Observable Propagation. |
| Open Datasets | Yes | The natural dataset was processed by taking the first 1,000,111 tokens of The Pile (Gao et al., 2020)...These names were obtained from the Gender by Name dataset from (UCI Machine Learning Repository, 2020) |
| Dataset Splits | No | The paper mentions training set sizes and implied splits (e.g., 3/4 of dataset for political parties) but does not explicitly specify separate training, validation, and test splits with percentages or counts for all datasets. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments (e.g., GPU models, CPU types, or memory specifications). |
| Software Dependencies | No | The paper mentions 'MDMM (Crowson, 2021)' as a Python library used but does not specify its version number or versions for other key software components like Python itself. |
| Experiment Setup | No | The paper describes the model architecture and general aspects of the method's application, but it lacks specific hyperparameters (e.g., learning rate, batch size, epochs) or detailed training configurations for the comparative experiments or the overall setup. |