Dr. Bapi Raju Surampudi
Professor, Cognitive Science Lab,
International Institute of Information Technology (IIIT)

Title: The Human Brain and Deep Learning: A Closer Look

Abstract:
In this talk current attempts at decoding the representations that the human brain might use to store information are reviewed. Interestingly, many of these investigations in turn use deep neural networks in order to decode functional magnetic resonance imaging (fMRI) responses of the human brain. Representational Dissimilarity Metric (RDM) is often used to compare brain responses and activation in the layers of suitably trained deep neural networks. We review these results with a view to highlighting what we are learning about the connections between human visual processing and deep learning based algorithms. We shall also highlight the implications of these insights for the design of future deep neural networks.

Arunkumar Balakrishnan
Director, ikval Softwares LLP

Title: Challenges and solutions, in developing Convolutional Neural Networks and Long Short Term Memory networks, for industry problems

Abstract:
Development of two classic Deep learning architectures, Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM) for two industry problems are described. CNN: The input to the application was a PDF of a scanned image of a document. The requirement was to identify  hand-marks, presence of a signature and handwriting. Further to the above challenges was the “Pose” problem. This was handled, through an innovative application of OpenCV functions. Hyper parameter values of dropout and normalization (to avoid overfitting), changes in activation functions, additional layers and different learning rates were experimented. The application correctly classified each document and improved efficiency of the department by Fifty percent. LSTM: The requirement was to predict the next month’s Equated Monthly Instalment (EMI) payment by a customer given data of history of EMI payments till the previous month. The challenge was, identifying correct time series from the possible “Multi Time Series”, in addition to making correct prediction for each time series. A set of data preprocessing functions were designed to identify categories of input. As LSTMs are prone to overfitting, dropouts were tried in initial phases. Surprisingly results were seen to improve when, application was tried without dropouts. Analysis of this behavior and other factors like changes in accuracy over training epochs are described. More than seventeen thousand customer’s data were handled. The results had Ninety five percent accuracy. Software engineering techniques, unique to building a Deep learning solution were designed and used to improve efficiency of development and quality of results. While an analysis of input data and its properties helps in choice of an initial architecture, a record of experimental values used and corresponding results was also maintained. This record was analyzed to design the next architectural setup.