Abstracts
Talk 1: On Acceleration of CNN Training using Diffgrad
by Prof. Bidyut Baran Chaudhuri
In recent times Deep Neural Networks (CNNs) are showing magical performance for solving various complex problems of practical importance. However, the training of CNN is a data intensive and time consuming process. Several approaches have been proposed from different angles to tackle this problem, which will be briefly described in this talk. One of these angles is the learning rate optimization, for which a moment based adaptive method named ADAM has been extremely successful. We have examined the work and came up with a new method named Diffgrad optimization. The talk will end with a discussion on DiffGrad, which showed better performance on several benchmark deep net problems.
Talk 2: Revisit to the Question of Deep Learning or Cheap Learning
by Prof. Dipti Prasad Mukherjee
The talk will introduce basic deep learning architecture and potential research challenges. A major focus would be to understand the explain-ability of the deep learning system.
Talk 3: SAR Image Analysis: Physics in Deep Learning
by Dr. Avik Bhattacharya
Radar remote sensing has made significant technological and scientific advances in the past few years. Sensors and constellations can now acquire high resolution, polarimetric, wide-swath data with high temporal repetivity. This has led to an exponential increase in the volume and complexity of the available data. Machine learning algorithms have been developed that can take advantage of the acquired data and convert latent information to actionable knowledge. Advanced neural network techniques, collectively called, "Deep Learning Algorithms" have demonstrated the ability to self-learn features from data, significantly reducing the need for time-consuming feature tuning, while delivering a robust performance in classification and regression. This talk will unfold the fundamental representation of SAR data and the inclusion of the underlined physics of scattering in deep learning techniques for various earth observation applications.
Talk 4: Machine Vision in Industrial Automation
by Mr. Jayavardhana Rama Gubbi Lakshminarasimha
The first part of the talk covers multiple machine vision applications developed in TCS Research Labs at Bangalore in the past three to four years. In the second part, the focus will be on video analysis. Video analysis is gaining importance due to its usefulness in a wide variety of applications. The efficiency of a video analytics engine primarily depends on its ability to extract the spatio-temporal features, which has enough discriminative and representative power. In the recent past, it has been shown that spatio-temporal data analytics is often based on spatial data analysis and temporal information is not effectively represented and used. Inspired by the way the human visual system operates, the talk discusses a hierarchical architecture to capture the spatio-temporal information from a given input video. Multiple video based applications are discussed.
Talk 5: A Brief Introduction to Deep Learning and Emerging Research Trends for Natural Language Processing (NLP)
by Dr. Kuntal Dey
From the era of rule based NLP, to machine learning (ML) based, the journey of NLP has seen significant evolution of techniques. The emergence of deep learning has strongly limited the use of more traditional techniques. Most of the science and practice have moved over to the deep learning approaches, using novel techniques. In this talk, I shall introduce a few artifacts of deep learning for NLP, discuss problems that have been solved or are being approached by researchers using such artifacts, and present the general trajectory of evolution of this area of research has been moving into over the past few recent years.
Talk 6: Deep Learning for Computer Vision
by Dr. Arijit Sur
In recent past, deep-learning-based methods achieved state-of-the-art performance on challenging computer vision problems such as image classification, object detection, semantic segmentation and many more. In general, these deep computational models consist of multiple processing layers to learn and represent data with multiple levels of abstraction which enables them to understand the intricate distributions visual data such as image, video, etc. In this talk, firstly, we try to understand the applicability of the deep networks over the conventional machine learning approaches to model the visual inputs. In the second part of this talk, we will see an example of deep architecture to handle an emerging computer vision problem.
Hands on 1: Introduction to PyTorch with an Application to Image Classification and Object Detection
by Mr. Bikash Santra
In the recent past, the performances of various computer vision tasks like image classification and object detection have been significantly improved with the advent of deep learning algorithms. The reasons behind the progress include the availability of enormous labeled data, advanced hardware support for high-performance computing, and deep learning programming frameworks like PyTorch.
This talk has been divided into two parts. The first part will introduce the deep learning library, PyTorch, by explaining the basics of python programming language, while the latter part will present the art of implementing computer vision algorithms for image classification and object detection with PyTorch in python.
Hands on 2: Image De-hazing: A Deep Learning Approach to Remove Haze from Image using Keras (TensorFlow) Library
by Mr. Ranjan Mondal
Image dehazing is one of the trending research problems. The goal is to restore the visibility in images with haze present in the atmosphere. We can model that haze using the physical haze model. We will estimate the parameters of the haze model using the Convolutional Neural Network(CNN). In the first part of the talk, we will start introducing Keras library. The second part of the talk will be more focused on image dehazing using Keras library.