Skip to content

Code for paper "Using natural language and program abstractions to instill human inductive biases in machines" (NeurIPS 2022)

Notifications You must be signed in to change notification settings

sreejank/language_and_programs

Repository files navigation

Using natural language and program abstractions to instill human inductive biases in machines

Link to preprint: https://arxiv.org/abs/2205.11558

Description of files:

Data

  • data/500_gsp_samples_text_human_encoded.npy: Human language embeddings

  • data/500_gsp_samples_text_synth_encoded.npy: Synthetic language embeddings

  • data/500_gsp_samples_text_synth.npy: Synthetic language descriptions

  • data/500_gsp_samples_text_human.npy: Human language descriptions

  • data/gsp_samples-recognition_activations-structurePenalty2.npz: Program embeddings

  • data/500_gsp_samples.npy: Training boards

  • data/gsp_4x4_full.npy: Boards from the GSP chain

  • data/gsp_4x4_full_probs.npy: Frequencies of each board in the GSP chain.

  • data/data_grulrep.csv: Raw chain data for GSP experiment [network_id: indexes the different GSP chains, active_index: indicates which cell is being changed, degree: indicates which iteration in the chain you’re looking at (0 initial random seed), definition: grid being changed]

  • data/gsp_4x4_sample.npy: Test set GSP boards

  • data/gsp_4x4_sample_starts.npy: Start tiles for test set GSP boards

  • data/gsp_4x4_null_sample.npy: Test set control boards

  • data/gsp_4x4_null_sample-starts.npy: Start tiles for test set control boards

  • data/hyperparams_grounding.pkl: Hyperparams for grounding agents.

  • data/hyperparams_nogrounding.pkl: Hyperparams for non-grounding agents.

Meta-RL Agent Code

  • auxillary_model.py: Training code for agents w/ auxillary loss.
  • auxillary_polcy.py: Setup for agent training code w/ auxillary loss.
  • small_env_lang_4x4.py: Modified training enviornment for grounded meta-rl agent for the GSP task distribution.
  • small_env_4x4.py: Training enviornment for baseline meta-rl agent for the GSP task distribution.
  • task_performance_zscore.py: Code to calculate task performance metric of paper.

Program Induction with DreamCoder

  • ec-master/: Folder with program induction code. This is a fork from: https://github.com/ellisk42/ec The vast majority of the code here is from the public repository of DreamCoder (Ellis et al. 2021) here: https://github.com/ellisk42/ec. Our additions are the implementation of DreamCoder in our enviornment (mostly in: dreamcoder/domains/grid, which has its own readme).

About

Code for paper "Using natural language and program abstractions to instill human inductive biases in machines" (NeurIPS 2022)

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published