The Fifth Elephant 2015

Deep Learning results looks very fascinating but it seems to require a huge infra to start using it. In this talk, we present how to approach it in incremental manner to make real use of Deep Learning.

Outline

Various problems related to text, image and voice have state-of-the-art results from Deep Learning. But all the groups who are giving these results have massive infrastructure to run Deep Learning algorithms. But same is not needed to start consuming goodness of Deep Learning.

In this talk, we present the reasons why Deep Learning is effective and present the ways to start consuming it using existing available models for solving the real problems. And then we talk about building the unsupervised pretraining models from scratch and introduce tricks and tools for same. We will cover various DL frameworks like theano, caffe, DL4j, Pylearn2 etc.

Speaker bio

Vivek Mehta is Data Scientist at Flipkart. He is currently interested in product discovery, personalization and deep learning. He has worked on various problems and domain including fraud detection, inventory planning, online ad optimization, handwriting recongnition, machine transalation, etc.

Devashish is a software developer at Flipkart. His primary work is in the area of deriving insights from unstructured data. His previous works include Review Summarization at Flipkart, User Insights Platform - a system which enables us to dig into 100s of TB of data to find user insights. He is currently working on sentiment analysis and aspect extraction from Social Media.

Links

• https://atlas.mindmup.com/2015/06/a5d62d20fb1c0132aa0b1a2e2c3822b9/how_to_stop_admiring_and_start_using_dee/index.html
• https://www.youtube.com/watch?v=zb98EWAUBxc&feature=youtu.be&t=225 (Vivek talking about outlier detection. Audio is poor, can work with headphones)
• https://youtu.be/sLDXm_53-mM?t=281 (Devashish talking about user insights engine)