WeightlessNeuralNetwork.jl Package

Introduction

Implements the WiSARD Neural Network (Wilkes, Stonham and Aleksander Recognition Device). It is composed of a set of discriminators, often called as classifiers[1]. This specific implementation is based on the description provided by [2], and the zero_skip and padded_retina were inspired on the IAZero implementation.

2019,2020 (@) Diego Carvalho - [email protected]

[1]: I. Aleksander, W. Thomas, P. Bowden, WISARD a radical step forward in image recognition, Sens. Rev. 4 (1984) 120–124.

[2]: M. De Gregorio and M. Giordano, “Cloning DRASiW systems via memory transfer,” Neurocomputing, vol. 192, pp. 115–127, 2016.

Example

using WeightlessNeuralNetwork

X = [
             Retina([1,1,1,0,0,0,0,0]),
             Retina([1,1,1,1,0,0,0,0]),
             Retina([0,0,0,0,1,1,1,1]),
             Retina([0,0,0,0,0,1,1,1])
           ]
           
y = [
             "cold",
             "cold",
             "hot",
             "hot"
           ]

w = Wisard(3,9)

train!(w,X,y)

classify(w,X)

TODO

• New feature: parallel discriminators --- Implement parallel discriminators.
• Refactor Wisard to access data by indices --- The parallel version of Drasiw is super fast due to the indices accesses of data. So, the Wisard class should be rewritten to incorporate the advantages.
• New feature: a test module --- Implement the test module. It should scan addresses, size, classes, etc.
• New feature: random generation control --- Provide a way to control the random generation. Each wisard net should have a random generation stream and this information must propagate through layers since Discriminator must use on the retina mapping.
• New feature: record wisard state --- Should implement IO functions in order to record the current wisard training state. Perhaps, it should use JSON.
• New feature: confidence level --- Should calculate the confidence level for each wisard output.
• Bug fix (1) --- The code below can crash when b reaches m during bleaching, and winners is greater than one (line 83).
• Implement ClusWisard
• Implement MageNet