Using Artificial Populations to Study Psychological Phenomena in Neural Models
Authors: Jesse Roberts, Kyle Moore, Drew Wilenzick, Douglas Fisher
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through population based experimentation we find that language models exhibit behavior consistent with typicality effects among categories highly represented in training. However, we find that language models don t tend to exhibit structural priming effects. Generally, our results show that single models tend to over estimate the presence of cognitive behaviors in neural models. This paper contributes by drawing connections between social and behavioral experimental design and neural model uncertainty estimation resulting in a (1) tool called Population LM for the creation of populations of neural models via stratified MC dropout. We harvest novel metrics and explore population best practices by applying artificial populations to the (2) replication and extension of (Misra, Ettinger, and Rayz 2021) (correlation analysis) and (3) (Sinclair et al. 2022) (difference analysis). We present novel results regarding the presence of typicality and structural priming effects in language models. |
| Researcher Affiliation | Academia | Jesse Roberts1, Kyle Moore1, Drew Wilenzick2, Douglas Fisher1 1Vanderbilt University 2Cornell University jesse.roberts@vanderbilt.edu, kyle.a.moore@vanderbilt.edu, ddw77@cornell.edu, douglas.h.fisher@vanderbilt.edu |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The resultant tool, Population LM, has been made open source. This paper addresses a current need in the study of cognitive behavior in neural models by introducing Population LM1, a system built on MC dropout for the creation of efficient populations of neural models. 1https://github.com/Jesse TNRoberts/Population LM |
| Open Datasets | Yes | We use typicality data from (Rosch 1975) which gives a typicality rank, ri, for each item, i, in category C. We run a similar experiment using sentence data from their work. We used the BERT family training data frequencies from (Zhou et al. 2022) to assess training data frequency correlations. |
| Dataset Splits | Yes | We split 3000 examples into two groups and conduct all 3 treatments on all 50 population members per species. The results for the first group of 1500 are reported and the results for the second set of 1500 are used for cross validation. The cross validation showed all results repeated within 0.02 (p<0.01) of our reported results. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., specific GPU/CPU models, memory, or cloud instance types). |
| Software Dependencies | No | The paper mentions "Py Torch compatible network" and "Py Torch" but does not specify version numbers for PyTorch or any other software dependencies, which are necessary for reproducibility. |
| Experiment Setup | Yes | We use Monte Carlo (MC) dropout (Gal and Ghahramani 2016) to form populations from base models. Therefore, we recommend that statistical studies adopt a 0.1 nominal dropout rate. We empirically find that a population of 50 is an acceptable compromise... We reproduce and extend the experiment conducted in (Misra, Ettinger, and Rayz 2021) which assessed the base model total correlation between probability and typicality. Our base model probabilities agree with past results, and we contribute novel tests using dropout populations and within category analysis which shed light on the factors that support the emergence of typicality effects in language models. We adopt 3 treatment conditions: the control (CT) is the probability of a sentence, Sx, without any priming P(Sx); the primed treatment (PT) is the probability of that sentence when the language model is first prompted with a sentence, πx, of similar structure P(Sx|πx); and the alternative treatment (AT) is the probability of Sx when prompted with a sentence, πy, of differing structure P(Sx|πy) but identical semantic meaning. |