[Pycon] [new paper] "Prakhar Srivastava" - Generative models - Building algorithms capable of imagination

[info a pycon.it]

Title: Generative models - Building algorithms capable of imagination
Duration: 60 (includes Q&A)
Q&A Session: 15
Language: en
Type: Talk

Abstract: Tl; Dr Generative models are making all the big headlines, understand what, how and most importantly why behind these esoteric algorithms. How can you start building one and why are these models the dark horse of artificial intelligence.

Yann LeCun, the man behind Convolution Neural Networks and the frontier of computer vision research, proclaimed that the future of artificial intelligence lies in unsupervised learning. Another such claim comes from, Yousha Benjio the head of MILA labs. According to him, clever feature engineering and better methods of unsupervised learning is the right way to progress.

What are Generative models? These are the machine learning models that can learn the hidden representation of a given dataset and then can extrapolate an accurate data point given certain conditions. If all that was jargon for you - here's a better intuitive understanding, imagine if you were given a white canvas to paint a picture of a bird, what would you do? First, you would imagine a bird and then fill in the colors. Well, Generative models do the same thing! Just with a lot of mathematics and sweet python libraries

Why are generative models important?

Simply put, any intelligent being can create a random scenario based on some constraint, let's try that now, think of a hummingbird sucking nectar from a flower. Simple, right? But a computer would drastically fail in such a trivial task! It would go bonkers! Now let's escalate this, try imagining your name but in the voice of your best friend. Easy, right! Well, this is exactly the same thing done by generative models. Google's duplex, rings any bell?

What to expect from this talk?

You'll understand everything from scratch about generative models. Most of the talk would concentrate upon the code of cycleGAN and demystifying the fundamentals of GANs. We'll talk about modern generative architectures like BigGan and AttnGAN. This talk would be beginner friendly but would not let experienced one fall off.

Pre-req: Understanding of fundamental machine learning models like convolution neural network, LSTMs etc., A knowledge of any modern deep learning framework 

Tags: [u'Deep-Learning', u'gans', u'mathematical-modelling', u'machine-learning', u'generative-models', u'Artificial Intelligence']