protobrain

This is an experimental project for biologically-inspired machine learning techniques.

Loosely based on Hierarchical Temporal Memory by Numenta.

Usage

The purpose of this library is to enable quick experimentation with different computation and learning models.

There are three steps to building any model:

1. Define an architecture This involves building a nested-layer architecture and setting up connections between them.

   You can build groups of neurons and individually manage the connections between them.

   n1 = neuron.Neurons(5)  # [x, x, x, x, x]
   n2 = neuron.Neurons(3)  #    [x, x, x]
   
   n2.input = n1  # Sets the 'MAIN' input of n2 to be n1

   Inputs have names, which can be referred to in the computation functions. You can access these through the set and get functions.

   n1.set('feedback', n2)

   You can treat neurons as layers and connect them with either neuron.FeedForward for forward connections, neuron.FeedBackward for backward connections, or neuron.LoopBack for connections between the same layer.

   layers = [neuron.Neurons(i) for i in [10, 10, 10]]
   # Not specifying an input name leads to setting the 'MAIN' input
   neuron.FeedForward(layers)
   neuron.FeedBackward(layers, 'feedback')
   neuron.LoopBack(layers, 'neighbors')

2. Define a computation function: Every neuron has to be able to compute an output based on its inputs.

   Some available computations are computation.StandardComputation and computation.SparseComputation.

   std = computation.StandardComputation(threshold=0.5)
   sparse = computation.SparseComputation(n=0.05)

   You can also implement your own function by creating a child class of computation.Computation.

3. Define a learning function: Synapses might be adjusted based on some logic.

   The only currently available learning function is learning.HebbianLearning.

   learn = learning.HebbianLearning(increase=0.005, decrease=0.001)

4. Define a sensor object that feeds inputs into the brain.

   The inputs are transformed from a raw value into a binary representation by the defined encoder.

   encoder = numerical.CyclicEncoder(min_value, max_value, encoding_length)
   sensor_ = sensor.Sensor(encoder=encoder)

5. Define a brain object that combines all of these:

   brain_ = brain.Brain(neurons, sensor_, computation=computation, learning=learning)

Once you've built a model, you want to run a benchmark to verify that it's doing what it's supposed to, and compare against different setups. Read more about benchmarks and metrics here.

An example of this can be seen in the benchmark.py script.

Building the protobuf files

Some functionality of this project depends on protobuf files. To build them, use the following command:

protoc -I=. --python_out=. protobrain/proto/*.proto

For more information on Protocol buffers, read the official documentation.

Testing

This project uses pytest and the tests can be run from the repository root.

python -m pytest

License

protobrain is a project licensed under GNU GPLv3.