Meet Idan Schwartz – a researcher in our research department. About a month ago, he gave a lecture in the NIPS convention in California, where only 20% of the articles sent were presented.
Hi Idan, congratulations on presenting on NIPS. How was it?
“I gave a lecture in NIPS as a representative of the Technion. The article was published as part of my doctoral studies. The convention, Neural information Processing Systems, is for researches in the academy and industry, and is all about Machine Learning and Computational neuroscience. It’s a very popular convention in the field of Deep Learning, and all the hot trends of machine learning. 10 thousand people participated in it this year.”
What was your article about?
“My article was about image Question Answering. Let’s say there’s an image and you have a constant question about it, such as “how many people are in the photo?”, “what’s the color of the clothing?”, “what kind of animals appear in the picture?” and so on. There are a few possible answers. These questions are trying to understand what appears in the image and the solution usually combines a system that know which parts of the image you have to look at, in order to answer the question. In my research, I have created a multidimensional listening system, so while the algorithm answers, it deduces which part of the picture to examine, which words in the question to examine and which answers to focus on. The innovation comes from listening not only to the image, but also to the answers and questions at the same time. This led to better results than other algorithms.”
What topics were particularly hot this year?
“The main hot topic was applications of Deep Learning, but there was a lot of discussion on Reinforcement Learning – training by reinforcing correct behaviors in algorithms. This is popular in robotics. This is unlike other algorithms that try to find a solution to minimize the error. Other interesting discussions were around GANs, a popular algorithm that enables creating new information and not just classifying an existing one. Another field was Bayesian inference, which is based on probability models, which allows for a better understanding as to why the learning algorithm operated in a certain way. Another topic was learning bias – networks are allowed to train on data that has bias. For example, if a bank wants to give loans to customers, and the race of the customer is part of the data, it may cause a bias in the decision making process. This is a warning to the community, because in the future more and more big organization will use learning algorithms.”