Unit Neurons

Unit Neurons is a repository for development of a C++ neural network library where each neuron is expressed using object instances embedded with its own states and functionalities, in hopes of gaining more understanding of neural nets through the perspective of complex systems.

We call for contributions on further development of the library by adding more functionalities, fixing bugs, etc.

Vision

Artificial neural networks are models that attempt to imitate features and functionalities of biological neural networks. Preexisting neural network libraries such as TensorFlow, PyTorch, and Keras strictly use matrix multiplication and neurons expressed in layers to execute feedforward and feedback loops.

However, in reality, a biological neural network is a network of mutually interacting neurons with their own function of calculating the signal output given the signal input. In our library, we treat each neuron as an object instance that includes a state, a feedforward method, and a feedback method.

Current Features

• Unit neurons for floating point values

• Classes necessary for training neural networks with the following learning algorithms:

  • Kohonen's Self Organizing Map (SOM)
  • Stochastic Gradient Descent

• Examples on the following operations:

  • Neural network fit to uniform distribution via Kohonen's SOM
  • Neural network fit to sine curve via gradient descent
  • Rough implementation of proximal policy optimization (PPO)

Documentation

Please refer to the documentation for instructions on installations and usage of the components in this library. Documentation

Open Problems

• Implementation of ADAM

• Multithreaded feedback operations

• Implementation of other activation functions (such as Leaky ReLU, etc).

• Implementations of other neural network models

  • Spiking neural networks
    • Boltzmann machine
    • Hopfield network

Major Contributors

• johnlime