gyyang / multitask

Code for Task representations in neural networks trained to perform many cognitive tasks

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

MultiTask Network

Dependencies

The code is tested in Tensorflow 1.8.0, Python 2.7 and Python 3.6, and on MacOS 10.13 and Ubuntu 16.04.

Scikit-learn (http://scikit-learn.org/stable/) is necessary for many analyses.

The seaborn package (https://seaborn.pydata.org/) is needed to correctly plot a few analysis results.

Reproducing results from the paper

All analysis results from the paper can be reproduced from paper.py

Simply go to paper.py, set the model_dir to be the directory of your model files, uncomment the analyses you want to run, and run the file.

Pretrained models

We provide 20 pretrained models and their auxillary data files for analyses. https://drive.google.com/drive/folders/1L8v-OZgYHVcKh1UKtCJl5QVlz8mkaRxr?usp=sharing

Get started with training

Train a default network with:

import train
train.train(model_dir='debug', hp={'learning_rate': 0.001}, ruleset='mante')

These lines will train a default network for the Mante task, and store the results in your_working_directory/debug/.

Get started with some simple analyses

After training (you can interrupt at any time), you can visualize the neural activity using

from analysis import standard_analysis
model_dir = 'debug'
rule = 'contextdm1'
standard_analysis.easy_activity_plot(model_dir, rule)

This will plot some neural activity. See the source code to know how to load hyperparameters, restore model, and run it for analysis.

About

Code for Task representations in neural networks trained to perform many cognitive tasks


Languages

Language:Python 100.0%