Vector-based navigation using grid-like representations in artificial agents

This package provides an implementation of the supervised learning experiments in Vector-based navigation using grid-like representations in artificial agents, as published in Nature

Any publication that discloses findings arising from using this source code must cite "Banino et al. "Vector-based navigation using grid-like representations in artificial agents." Nature 557.7705 (2018): 429."

Introduction

The grid-cell network is a recurrent deep neural network (LSTM). This network learns to path integrate within a square arena, using simulated trajectories modelled on those of foraging rodents. The network is required to update its estimate of location and head direction using translational and angular velocity signals which are provided as input. The output of the LSTM projects to place and head direction units via a linear layer which is subject to regularization. The vector of activities in the place and head direction units, corresponding to the current position, was provided as a supervised training signal at each time step.

The dataset needed to run this code can be downloaded from here.

The files contained in the repository are the following:

• train.py is where the training and logging loop happen; The file comes with the flags defined in Table 1 of the paper. In order to run this file you will need to specify where the dataset is stored and where you want to save the results. The results are saved in PDF format and they contains the ratemaps and the spatial autocorrelagram order by grid score. The units are ordered from higher to lower grid score. Only the last evaluation is saved. Please note that given random seeds results can vary between runs.

• data_reader.py read the TFRecord and returns a ready to use batch, which is already shuffled.

• model.py contains the grid-cells network

• scores.py contains all the function for calculating the grid scores and doing the plotting.

• ensembles.py contains the classes to generate the targets for training of the grid-cell networks.

Train

The implementation requires an installation of TensorFlow version 1.12, and Sonnet version 1.27.

$ virtualenv env
$ source env/bin/activate
$ pip install --upgrade numpy==1.13.3
$ pip install --upgrade tensorflow==1.12.0-rc0
$ pip install --upgrade dm-sonnet==1.27
$ pip install --upgrade scipy==1.0.0
$ pip install --upgrade matplotlib==1.5.2
$ pip install --upgrade tensorflow-probability==0.5.0
$ pip install --upgrade wrapt==1.9.0

An example training script can be executed from a python interpreter:

$ python train.py --task_root='path/to/datasets/root/folder' --saver_results_directory='path/to/results/folder'

Disclaimer: This is not an official Google product.