Enhancing Siamese Neural Network performance using Artificial Immune Systems

What is it about?

Machine learning (ML) is able to perform well with complex tasks such as computer vision, speech recognition, text understanding, etc. However, it generally requires an enormous amount of data to "learn". Siamese Neural Network (SNN), a type of ML approach are able to deliver good performance with limited data. However, SNNs are only able to quantify the similarity between given data and cannot classify them, at least in their original form. SNNs have been modified and equipped further with additional techniques to allow classification through them. Nevertheless, existing approaches for SNN-based classification leave substantial room for improvement. In the proposed method, we utilised concepts drawn from Artificial Immune Systems (AIS) to enhance the classification performance of SNNs. The method has been tested on three different benchmark high-dimensional image datasets. We obtain an increment both in the accuracy and the inference time of the SNN.

Featured Image

Photo by Google DeepMind on Unsplash

Photo by Google DeepMind on Unsplash

Why is it important?

The work bridges two domains of ML, namely, AIS and SNN, thus also opening up a new area for further exploration. Due to the stark similarity between AIS and SNN, the concepts from both domains are highly compatible and can be interchangeably leveraged into one another. The existing works using SNN for classification hamper the scalability of SNN, add substantially to the computational overhead, increase the inference time and lose the inter-class information while making an inference. Usage of AIS in the proposed manner avoids all these loopholes, thereby increasing the accuracy and the inference time.

Perspectives