Parth Gupta (Indian Institute of Information Technology, Vadodara)
Manoj Chinnakotla (Microsoft, India)
Date: 7th December (Afternoon Session)
Deep learning approaches have taken the machine learning community
by storm. It's been successfully applied for image processing, speech
recognition and natural language processing (NLP). The main advantage of
deep learning over conventional approaches is that it's completely data
driven with stacked layers of neural networks progressively "learning"
the data with increasing levels of abstraction, without the necessity of
manually hand-coded features. In the context of NLP, word embedding is
the starting point of transforming a categorical feature, e.g. a word
from a vocabulary, into a continuous representation of a real-valued
vector in the Cartesian space of p (a finite integer) dimensions. The
outcome of the embedding ensures that semantically close terms, e.g.
'sun' and 'solar' are placed in close proximity in comparison to terms
that are not semantically similar, and hence are placed farther apart,
e.g. 'computer' and 'sun'.
The objective of this tutorial is to
introduce the deep learning technologies including basics. The
state-of-the-art models for information retrieval based on deep-learning
will be covered and various applications will be discussed. The overall
goal is to familiarize the audience with the latest developments in the
field and explain them with details