Page 262 - AWSAR 2.0
P. 262

238 || AWSAR Awarded Popular Science Stories - 2019
Our research matters because we focus on addressing the fundamental issue of data hungriness of the most successful genre of learning systems called deep learning. By addressing this limitation, we
can apply the deep learning
methods to the domains where
we do not have much data. Our
research focusses on making
a machine learn about a novel
or a rare visual concept by
showing it only few samples.
For example, how a machine
can recognize red panda when
we show it just one example of
red panda. Though identifying
a species may seem a low
impact problem, a machine that
can learn from just a few visual
examples can have high impact applications. India has several such application domains where we do not have data and still want the machine learning system to learn. For example, our research can help when we wish to identify if some tumor is cancerous or not from just a handful of tumor scan images. Teaching machines the art of learning from justafewsamplespracticallyopensaplethora of opportunities to solve problems, which we could not solve earlier.
Why does it interest researchers?
Teaching machines to learn from just a few training samples is also called as few shot learning(FSL). FSL has been tried earlier in shallow machine learning methods. These methods fall under the umbrella term of transfer learning and had limited success. But these traditional methods lack the general advantage of deep learning where the machine learn in automated fashion with least human intervention. Deep learning also exploits the parallelism in computing so that the training is quick. Probability based approaches store
and use prior knowledge much similar to humans. Humans apply their learnt wisdom while learning something novel. For example, whenever humans see red panda for the
first time, they recollect their previous knowledge of all animal species that look or behave similar to red panda to learn about it.
Despite these several approaches, there is no clear mandate on effectiveness and, therefore, FSL is still an open research problem. Our proposed work investigates a generative approach for visual concept learning. The generative approach of learning is much similar
to human learning where we can generate several mental images of rare visual concept (for example, red panda) we learnt. We realized that even measuring how much, the machine has learnt (machine cognition) is an open problem. The research objectives of proposed work are: 1) To propose a framework for benchmarking machine cognition and to investigate its usefulness while comparing computer vision tasks, 2) To propose and investigate effectiveness of a generative method(s) for visual concept learning from fewer labelled samples. Pro posed approach is to use generative machine learning techniques to tap knowledge learnt during other concepts’ learning to learn a rare visual concept.
Research Problem
This research aims to understand how human-like visual concept learning can be achieved in machines. As a first step toward addressing this objective, we set out to answer the following research questions:
RQ.1. How much cognitive processing current machine learning methods do in
   Microalgae grows faster than plants and trees, captures harmful carbon dioxide from the environment, can be grown on sewage and marshy waters, helps in reducing the harmful chemical load from the wastewater.
  







































































   260   261   262   263   264