- Abstract: This research paper describes a simplistic architecture named as AANN: Absolute Artificial Neural Network, which can be used to create highly interpretable representations of the input data. These representations are generated by penalizing the learning of the network in such a way that those learned representations correspond to the respective labels present in the labelled dataset used for supervised training; thereby, simultaneously giving the network the ability to classify the input data. The network can be used in the reverse direction to generate data that closely resembles the input by feeding in representation vectors as required. This research paper also explores the use of mathematical abs (absolute valued) functions as activation functions which constitutes the core part of this neural network architecture. Finally the results obtained on the MNIST dataset by using this technique are presented and discussed in brief.
- TL;DR: Tied weights auto-encoder with abs function as activation function, learns to do classification in the forward direction and regression in the backward direction due to specially defined cost function.
- Keywords: Neural Network architecture, Learned representation space, absolute valued function, bidirectional neuron