Experiments to measure the behavior of deep neural networks performing a visual search task.
For some background and a summary of the results, please see this Jupyter notebook.
Experiments were run using Anaconda on Ubuntu 16.04. The following commands were used to create the environment:
tu@computi:~$ conda create -n searchnets python=3.6 numpy matplotlib imageio joblib tensorflow-gpu tu@computi:~$ source activate searchnets tu@computi:~$ git clone https://github.com/NickleDave/visual-search-nets.git tu@computi:~$ cd ./visual-search-nets tu@computi:~$ pip install .
Installing this package (by running
pip install . in the source directory) makes it
possible to run the experiments from the command line with the
searchnets command, like so:
tu@computi:~$ searchnets train config.ini
The command-line interface accepts arguments with the syntax
searchnets command config.ini,
command is some command to run, and
config.ini is the name of a configuration file
with options that specify how the command will be executed.
For details on the commands, see this page in the docs. For details on the
config.ini files, please see this other page.
Data is deposited here: https://figshare.com/articles/visual-search-nets/7688840
Makefile replicates the experiments.
tu@computi:~$ make all
- Research funded by the Lifelong Learning Machines program, DARPA/Microsystems Technology Office, DARPA cooperative agreement HR0011-18-2-0019
- David Nicholson was partially supported by the 2017 William K. and Katherine W. Estes Fund to F. Pestilli, R. Goldstone and L. Smith, Indiana University Bloomington.