One-Shot-Learning-on-Omniglot-data

One shot learning using siamese twins architecture on Omniglot dataset

Table of Contents

Show/Hide

1. File Descriptions

• Dataset
• File Structure

2. Technologies Used
3. Structure
4. Executive Summary
   • [ 1. Why use One-Shot Learning ](#Why use One-Shot Learning)

File Descriptions

• Dataset : The Omniglot dataset is a collection of 1623 hand-drawn characters from 50 alphabets. For every character there are just 20 examples, each drawn by a different person at resolution 105x105. Each image is paired with stroke data, a sequence of [x,y,t] coordinates with time (t) in milliseconds. This data is split as 30 alphabets used for training and the other 20 used for validation of the model.

• File Structure :

  image_background

  Gujarati

  Character01

  20 images

Tecnologies Used:

• Python
• Pandas
• Numpy
• Matplotlib
• Open CV
• Scikit-Learn
• Keras
• Tensorflow

Executive Summary

Why use One-Shot Learning

In conventional image processing, an image is put through a CNN(Convolution Neural Network) to extract features from which an object/edge is detected or classified, but this is very computationally intense and requires a very large training set to alleviate all bias. Another issue is that when a new set of training data is added the model has to be re-trained. So one-shot learning proposes to extract features from a small set of training data and make a faster prediction based on similarity. Ex Face unlock in phones take only multiple images of and feature for your face and store it on device but when a new face is added it does not have re-train the entire model, it extracts the features from each image and compares for similarity thus reducing the time it takes to unlock while still being reliable.