mothnet 0.0.5

Neural network modeled after the olfactory system of the hawkmoth.

pymoth

This package contains a Python version of MothNet

Neural network modeled after the olfactory system of the hawkmoth, Manduca sexta (shown above).

This repository contains a Python version of the code used in:

• "Putting a bug in ML: The moth olfactory network learns to read MNIST", Neural Networks 2019

---

Docs (via Sphinx)

Installation

Built for use with Mac/Linux systems - not tested in Windows.

• Requires Python 3.6+

Via pip

pip install mothnet

From source

First, clone this repo and cd into it. Then run:

# Install dependencies:  
pip install -r pymoth/docs/requirements.txt
# Run sample experiment:
python pymoth/examples.py

Dependencies (also see requirements.txt)

• scipy
• matplotlib
• scikit-learn(for kNN and SVM models)
• pillow
• keras (for loading MNIST)
• tensorflow (also for loading MNIST)

---

Example experiment (also see examples.py)

import os
import pymoth

def experiment():

    # instantiate the MothNet object
    mothra = pymoth.MothNet({
        'screen_size': (1920, 1080), # screen size (width, height)
        'num_runs': 1, # how many runs you wish to do with this moth
        'goal': 15, # defines the moth's learning rates
        'tr_per_class': 1, # (try 3) the number of training samples per class
        'num_sniffs': 1, # (try 2) number of exposures each training sample
        'num_neighbors': 1, # optimization param for nearest neighbors
        'box_constraint': 1e1, # optimization parameter for svm
        'n_thumbnails': 1, # show N experiment inputs from each class
        'show_acc_plots': True, # True to plot, False to ignore
        'show_time_plots': True, # True to plot, False to ignore
        'show_roc_plots': True, # True to plot, False to ignore
        'results_folder': 'results', # string
        'results_filename': 'results', # will get the run number appended to it
        'data_folder': 'MNIST_all', # string
        'data_filename': 'MNIST_all', # string
                            })

    # loop through the number of simulations specified:
    for run in range(mothra.NUM_RUNS):

        # generate dataset
        feature_array = mothra.load_mnist()
        train_X, test_X, train_y, test_y = mothra.train_test_split(feature_array)

        # load parameters
        mothra.load_moth() # define moth model parameters
        mothra.load_exp() # define parameters of a time-evolution experiment

        # run simulation (SDE time-step evolution)
        sim_results = mothra.simulate(feature_array)
        # future: mothra.fit(X_train, y_train)

        # collect response statistics:
        # process the sim results to group EN responses by class and time
        EN_resp_trained = mothra.collect_stats(sim_results, mothra.experiment_params,
            mothra._class_labels, mothra.SHOW_TIME_PLOTS, mothra.SHOW_ACC_PLOTS,
            images_filename=mothra.RESULTS_FILENAME, images_folder=mothra.RESULTS_FOLDER,
            screen_size=mothra.SCREEN_SIZE)

        # reveal scores
        # score MothNet
        mothra.score_moth_on_MNIST(EN_resp_trained)
        # score KNN
        mothra.score_knn(train_X, train_y, test_X, test_y)
        # score SVM
        mothra.score_svm(train_X, train_y, test_X, test_y)

        # plot each model in a subplot of a single figure
        if mothra.SHOW_ROC_PLOTS:
            mothra.show_multi_roc(['MothNet', 'SVM', 'KNN'], mothra._class_labels,
            images_filename=mothra.RESULTS_FOLDER+os.sep+mothra.RESULTS_FILENAME+'_ROC_multi')

Sample results

Dataset

MNIST Data

Modules

• classify.py Classify output from MothNet model.
• generate.py Download (if absent) and prepare down-sampled MNIST dataset.
• params.py Experiment and model parameters.
• sde.py Run stochastic differential equation simulation.
• show_figs.py Figure generation module.
• MNIST_make_all.py Downloads and saves MNIST data to .npy file.

---

Questions, comments, criticisms? Feel free to drop us an e-mail!

Bug reports, suggestions, or requests are also welcome! Feel free to create an issue.