Digit Classification Based on Mechanisms Used by Human Brain

摘要

One of the obvious differences between present machine learning and performance of human brain lies in the number labelled data that needs to be provided to the model for classification. The ultimate target of machine learning algorithms should be to learn from very few examples as demonstrated by human brain. The work presents an instance of reducing sample complexity in handwritten digit classification using the concept followed in object recognition with reduced training set by humans. The characteristics of selectivity and invariance to transformations in visual cortex lead to a feature representation which can learn from few examples. The same has been implemented for digit classification, and promising results are obtained which can be extended to variety of object classification/recognition tasks involving images, strings, documents, etc.