Emergence of Sparse Representations from Noise

This is the codebase behind the paper Emergence of Sparse Representations from Noise and is sufficient to reproduce our core experiments.

Follow these steps:

1. Set up a new Conda environment using:

conda create --name NoiseSparsity python=3.9
conda activate NoiseSparsity
conda install pip
pip install --upgrade pip
pip install --upgrade setuptools

2. Clone this github repo and install its requirements:

git clone https://github.com/TrentBrick/SparsityFromNoise.git
cd AnonPaper2
pip install -r requirements.txt

You may be able to use other versions of the libraries found in requirements.txt but no promises.

3. Setup Wandb

Make an account here and put the details into test_runner.py under wandb_logger=.

Otherwise, set use_wanbd=False to run the models but limited outputs will be printed.

4. Run python test_runner.py that will by default run an ReLU model on CIFAR10 latents with 0.8 noise.

See exp_commands/ for all parameters used to run all of our experiments.

You can put these parameters into test_runner.py to run them and fill in load_path= with a trained model. To reproduce all of our results we recommend using a job parallelizer like Ray or SLURM to run each experiment as a different job. See ray_runner.py for our use of ray.

---

Code Base Overview

• models/ - has all models.
  • models/Base_Model.py - has the PytorchLightning backbone that all models inherit.
  • models/Diffusion_Models.py - has the Diffusion model backbone and two different models: Feedforward, SDM. SDM models that implement some form of Top-K and have built in L2 normalization of weights, ability to enforce all positive weights, etc. All models use the TrackedMLPLayer.py that tracks many different metrics.
  • models/InhibCircuit.py - Implements explaining away and facilitation.
  • models/TopK_Act.py - Top-K activation function that implements all of the variants considered.
  • models/TrackedActFunc.py - Tracks metrics such as the number of dead neurons and number of neurons active for an input.
• notebooks/ - ipynb notebooks to analyze the different models.
• py_scripts/ - handle model and data parameters, additional helper scripts for training.
• data/ - We provide the CIFAR10 ConvMixer latent embeddings.
• data/data_preprocessing/ - scripts to process and make the default CIFAR10 and other datasets load faster.
• ray_runner.py - runs a number of Ray jobs in parallel. Pulls experiment details from experiment_generator.py