neuralduino's Introduction

neuralDuino

extendable perceptron library to build dynamic and reconfigurable Artificial Neural networks with back-propagation for arduino

Note : This library sways away from the conventional layered architecture that is taught in most youtube videos but it definitely supports it. Now you have the power to make you arduino actually LEARN from the environment.

Pre-Requisites

Tested using Arduino v1.6.5(windows) on UNO and Mega but should work on any board really and since it is based on C++ it can be ported to any platform easily

Background

this library imagines neural networks as connected individual neurons, obvious?, not so much because conventional libraries think of neural networks as layers of nodes.
Due to the above ideology the library creates a neuron object that has synapses incoming to it and a single output going from it
This design method gives the flexibility of making the most designer and complex ANN networks and have the control of designing and connecting the nodes to each other on the fly as well (Possibility of cascaded Neural Nets) with a very simple API
Read the header file comments for general usage of the library and examples to use it

Instructions

Start with the examples, read the comments and move from there
for any problems post an issue

Features

supports back-propagation , by calling backpropagate on a neuron the library backpropagates the beta to all the neurons that this neuron is onnected to
Non layered architecture to support any ANN design required by the user
choose your own activation function from the main sketch for each neuron individually
connect the neurons to each other as you like
each neuron is an object and whatever network you draw on the paper can be realized with this library
Layer management (Construction and connections) given to the user with a default back-propagation and adjWeights implementation that needs to be called by the programmer as and when they wish.
Lots of examples with different neural net architectures ;-P

Read the story on hackaday - http://bit.ly/neuralDuino_H

neuralduino's People

Contributors

Stargazers

Watchers

neuralduino's Issues

Can we have the forward Forward algorithm and less lines for a like 100 neuron network

Cascade Neural networks

Read about cascade neural networks which provide a faster convergence, make some example perhaps demonstrating the same.

byte inCount in "neuralduino.h" is commented to count only the no. of nodes connected to the neuron and not the total no. of inputs to the neuron.
the name is therefore misleading
choose a better name

When we want to make a looping network it overflows in backward propagation

I use a esp32 and wit only 14 neurons it crashes because the propagation is looping and consuming memory (to fix such issues make it that the backwards propagation number is limited by like
Define x 1
X (number of neurons) = max backwards going for a propagation
X (number of neurons)² = max num of propagation for all the network

Recommend Projects

pranavgulati / neuralduino Goto Github PK