By Aishwarya Menon
An Introduction to the World of Deep Learning
With several advances made in computer science-related domains like machine learning, artificial intelligence, and deep learning, each day we see a new solution cropping up to deal with the ever-changing demands of the present world. From home assistance systems to solutions for various organizations, the use of such advanced technology has become indispensable in modern times. In the following article, we will explore how the latest models of neural networks are deployed in the field of Image Recognition.
Siamese…Doesn’t the word sound familiar?
So the first thing that comes to mind when we hear the word “Siamese” is conjoined twins, being given this moniker by virtue of being bodily attached to each other and identical in most respects. Ever thought this terminology could be applied to a class of neural networks? That’s exactly what we are going to talk about here. Many of us would have heard of Convolutional Neural Networks, their various layers, their working, etc. These traditional neural networks have worked just fine to train data for a wide variety of purposes and arrive at suitable models. But one drawback we observe here is that it requires a large amount of labeled data, whose training would be a challenging task.
Delving Deeper into Siamese Neural Networks
The basic SNN architecture consists of twin sub-networks that are used to map similarities between input data rather than extract the differences. Thus data for which limited supervision is available can be trained and modeled with relative ease. It is also called one-shot classification as only one training example is needed for each class. A very basic outline of how this works:
1. Feed the input images to the individual Convolutional Neural Networks.
2. Compute the absolute difference of the encodings (containing features) of the images so produced.
3. The difference is fed to the sigmoid function within the SNN which will then output a similarity the score for the images, ranging between 0 and 1.
4. 0 indicates dissimilar inputs and 1 indicates similar inputs.
Moving Forward: Research Work and Real Life Use-Cases
A very illuminating example is the extensive research work done by Gregory Koch and his team, using N-way one-shot classification on an Omniglot data set (containing 1623 characters over 50 different alphabets). N-way simply means a single Test Image will be compared with a Support Set containing N images (in this case characters). This work can be extended to various applications like signature verification where various characters in the signatures of different individuals vary in style, spacing, etc., and seem to have an individuality of their own. The seemingly difficult task can be handled effectively through SNNs. Another example is access control for big corporations through facial recognition. In the case of employees joining and leaving corporate firms, re-collection and re-training the data would prove to be another task, which stands eliminated with SNN.
Several tech companies have employed this concept in their algorithms and moving forward one can hope for greater innovations and even better implementations of such machine learning models.