The Fifth Elephant 2016

Deep Learning has made lot of progress in the last four years:

• Newer ideas in architectures (module like architectures of NiN, GoogLenet, VGG Net etc),
• Adverserial examples and its aftereffects in certain networks (impercetible but intentional changes to inputs causing highly confident wrong results),
• Newer results with theoretical narratives (from invariant scattering transforms, group theoretic perspective, probablistic perpective, theoretical proofs on learnability of some of these models and many more).

among other things.

This talk is a roundup of some of the interesting ideas published at top tier academic conferences and Arxiv from last two years covering above themes. The main idea is to collate the various narratives to help create a story of why some of these models work.

Given the current rate of publications and the large areas of Computer Science, Electrical Engineering and Mathematics (and even Physics) that it covers; the attempt is to create a map of coherent ideas so that a practitioner can get a holistic picture of current results and directions and use them to debug/develop their models. Due to the broad nature of the topic and even broader selection of papers/topics available and limited time, only a representative portion will be covered.

Outline

• What is new about CNN and what are some of the ingredients of the current revolution
• Interesting results in CNN Architectures: NiN, GoogLeNet, VGG Net, Residual Nets, Spatial Transformer Networks
• Why Theory ? Can’t we just take the hacker approach ?
• Why Hierarchy ?
• What is being learned ?
• Why non-linearity ?
• Can we make networks more interpretable ?

Speaker bio

Sumod Mohan holds an M.S degree and was purusing his Ph.D in Electrical Engineering from Clemson University with his broad interest in application of Graph Algorithms and Probablistic Graphical Models in Computer Vision. He decided to dropout and join HighlightCam (in California) where he led Computer Vision Algorithm Development. He currently leads the Computer Vision and Machine Learning Business Division (CVD) at Soliton Technologies. His experience spans Computer Vision, Machine Learning, 3D Vision, Deep Learning, Graph Algorithms, Probabilistic Graphical Models, Code Optimization and Parallelization and has worked in the Computer Vision and Machine Learning for past 10 years.