A framework for classification based on Characteristics of human brain using machine Learning approach

Abstract

One remarkable ability of humans is learning from few examples. This is in contrast to current newlinemachine learning approaches which require large datasets for training to achieve human level newlineperformance. Also, for novel learning tasks, like learning to play a new game or acquiring a new newlinelanguage, humans perform better. newlineIn this thesis, we explore the above issues in two complimentary directions. One is based on the newlinestudies on visual cortex which highlights that visual information processing in cortex occurs newlinethrough successive layers of filtering and pooling operation. This leads to learning a newlinerepresentation invariant to scale, position and view point transformation of images that result in newlinelow sample complexity during visual recognition. We implement this approach and observe that newlinethe features obtained reduces the number of examples required for training in comparison with newlineraw features. newlineAnother direction for few-shot classification is based on learning. For humans, learning is a model newlinebuilding process which utilises the compositional structure of objects. Inductive priors are learned newlineover time which guides inference and application of learned concepts to new domains. Taking a newlinecue from this, we implement the classification of handwritten digits based on the process in which newlinedigits are generated. The parts in the digit and relationship between strokes is modelled using a newlinemultinomial generative model and evaluated for one shot classification. newlineHumans remember best examples during recognition and use them as templates for classifying newlinenew inputs. This is conceptualised using a Gaussian regression neural network (GRNN) newlinearchitecture. The prediction is enhanced by using a Probabilistic neural network (PNN) and the newlineensemble approach gives promising results in one shot classification. newlineThere is an issue of uncertainty associated with the prediction of Neural networks. This is newlineaddressed through a mechanism based on probabilistic neural network which handles out of newlinesample examples effectively. newline