Human-like Few-Shot Learning via Bayesian Reasoning over Natural Language

Authors: Kevin Ellis

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

Reproducibility Variable Result LLM Response
Research Type Experimental We contribute (1) a model of symbolic concept learning that supports efficient inference over a flexible hypothesis class; (2) an evaluation on human data from two different concept learning experiments; and (3) a simple recipe for extracting a humanlike prior over concepts, given raw behavioral data.
Researcher Affiliation Academia Kevin Ellis Cornell University kellis@cornell.edu
Pseudocode No The paper includes Python code snippets for translating natural language concepts but does not present a formal pseudocode block for the overall algorithm or any of its main components.
Open Source Code Yes Code and data available at: https://github.com/ellisk42/humanlike_fewshot_learning
Open Datasets Yes We take human data from [43]. ... We obtain this human data from [55], which covers 112 concepts, collecting judgements from 1,596 human participants as they attempt to learn each concept over 25 batches of examples.
Dataset Splits Yes For the number game we do 10-fold cross validation to calculate holdout predictions.
Hardware Specification Yes All model were trained on a laptop using no GPUs.
Software Dependencies No The paper mentions software like “Codex code-davinci-002”, “GPT-4”, “Code Gen 350M”, “all-Mini LM-L6”, and “Adam”, but does not specify their version numbers to allow for reproducible setup.
Experiment Setup Yes We use Adam [46] to perform maximum likelihood estimation of the parameters, following Eq. 5. ... We perform 1000 epochs of training for the Number Game, and 100 epochs for logical concepts. ... We also place a learnable temperature parameter on the posterior. ... We fit these parameters using Adam with a learning rate of 0.001.