Mysore Park Workshop on Vision, Language and AI

27-29 December 2016

Summary

In this workshop we would like to broadly meet and discuss current recent developments in the areas of Vision, language and AI. We specifically intend to explore two main themes
  • Vision and Language
    In this area we intend to explore the interconnections between Vision and Language and the deep learning techniques that enable both these areas.
  • AI Techniques
    In this area we intend to explore further the techniques of deep reinforcement learning and other advances in deep learning. Some of these advances would have relevance in Vision and Language.

There have been a number of advances made in both these themes. The main advances have been made in terms of learning of deep representations  and advanced optimisation and machine learning techniques in learning representations. The implications of the representations learned using these techniques go beyond just one single application. For instance, a convolutional neural network architecture that can recognise objects in images, can also be used to understand the sentiment in a document. A deep reinforcement learning approach can compete in games like Go rivalling experts. These networks involve learning of millions of parameters with very large datasets. There are a number of practical challenges in terms of actual optimisation techniques that work.

In order to keep pace with the rapid advances happening in Vision, Language and AI, we would like to organize a workshop that would facilitate exchange of ideas between the Indian community of researchers with the best experts in all these areas. The key topic for discussion would be Deep Representation Learning for Scene Understanding in terms of Vision, Language and AI. This would enable a discussion on various aspects where Vision and Language would pertain to scene understanding and description, answering of questions and AI techniques that could be applied for this problem.

The workshop would allow students to be able to learn from the best experts on the latest cutting edge developments in this area. It will also aid establishing dialogues between researchers from India and abroad to form collaborations.

Scope

The scope for this workshop is based on the two main themes of the workshop.

Vision and Language

In Vision and Language we are interested in a number of areas such as:
  • Vision to language
    There has been a number of interesting works that aim at automatic association of language with an image such as image captioning and visual question answering . Further interesting work has also been done in terms of captioning videos
  • Language to vision
    - A fascinating interaction between vision and language is the generation of images based on language. In recent work generative models of images have been used generate images from captions.
  • Advanced architectures for vision and language
    There have been a number of methods that explore the different neural network topologies such as guided LSTMs \cite{Jia2015} and Spatial transformer networks.

AI

In AI, we are interested in a number of areas such as:
  • Advances in deep learning
    There have been rapid advances made in terms of networks such as Generative Adversarial Networks, Adversarial autoencoders, reduced parameter models for neural networks, deep contextual models.
  • Advances in Reinforcement learning
    A particularly exciting area of research in AI has been deep reinforcement learning that have yielded beating of humans at games such as Google Deep Mind's recent work on AlphaGo. There have been a number of works that have explored this area recently.
  • Advances in program induction
    Another area of recent interest is incorporating classical models of computing in neural network architectures yielding Neural Turing Machines, Neural random access networks and Memory networks.