Speaker
Description
Solving problems of mutually exclusive situations with arbitrary relations, such as the parity problem, poses a significant challenge for artificial neural networks (ANN). A possibility of improvement of ANN in the form of the so-called product neurons has been proposed a very long time ago, yet all the product neurons in such models had fixed synaptic weights. A layer of product neurons performs multiplication of neuron activation values instead of the sum, and the synaptic weights serve as exponents, which, in the case of weight adaptation using the classical error-backpropagation method, yields complex numbers. To this end, we introduce QuasiNet, a novel ANN architecture that realises trainable product neurons through a new mathematical operation we call quasi exponentiation. To the best of our knowledge, our approach offers the first-ever ANN with trainable product neurons. Our preliminary results indicate that QuasiNet performs superbly in the parity or bitshifting problems of any rank with the minimal number of neurons needed to solve the problem. We also observed that it can outperform any other neural networks in hard tabular data problems. We also evaluated ways of utilising QuasiNet in the domain of neurorobotics, an application domain that relies on efficient and timely computation to learn arbitrary relationships. Currently we explore possibilities of utilizing product neurons in the field of deep learning. Thus, we are seeking discussion, inspiration and prospective collaboration to make good use of this intriguing novel neural network architecture.
| Pracovisko fakulty (katedra)/ Department of Faculty | KAI |
|---|---|
| Tlač postru/ Print poster | Budem požadovať tlač /I hereby required to print the poster in faculty |