Summary - Artificial Intelligence and Machine Learning

The Diamond Jubilee Faculty Alumni Network (FAN) Symposium (US)

 

The AI/ML session at FAN 2018 was organised as two 2-hour long sessions. It was attended by approximately 40 people, of whom about 25 directly contributed to the programme. The audience included a mix of graduate students, post-doctoral researchers, early-career faculty, senior faculty, as well as a number of entrepreneurs and research scientists from Silicon Valley.

The primary goal of the session was for attendees to learn about each others’ work in the area of AI/ML, thereby forming the basis for future collaborations and partnerships. A total of 19 short, technical presentations were made, covering a variety of topics in AI/ML and computer science. These talks represented the state-of-the-art in areas such as natural language processing, computer vision, optimisation, decision making, supervised and unsupervised learning, neural networks, and large-scale data analysis. A small number of talks also addressed the design of systems infrastructure--networks, computing, and hardware--to support modern AI/ML applications.

The highlight of the session was a panel discussion on “AI and Education”: clearly a very relevant topic for the FAN audience. To begin, the panelists  identified advantages that an automated tutoring system would have over human instructors. Examples were provided of data-driven systems that can predict the final grades of students with remarkable accuracy after monitoring only a week of their on-line activity (and without access to other external signals). If students’ proficiency can be gauged early on in a course, surely one can also devise corresponding interventions where needed. A second surprising observation was that automated instructional interfaces may enjoy a unique advantage in being able to elicit certain types of information for students--such as their confusions and insecurities--that students would usually not be comfortable sharing with human instructors. Having thus established the case for augmenting human teachers with “super-human” tools facilitated by AI, the discussion specifically zeroed in on personalisation as a key focus area of AI-driven learning support systems. Given the need to scale to large and diverse groups of students, it would be necessary to devise specialised interactions for different groups of students, or even at the granularity of individual students. A human teacher would not have the time and resources for personalisation and repetition--but these are precisely the aspects that AI is especially adept at handling.

A second discussion held in the session contemplated the possible ill-effects of AI-driven growth, examples of which are beginning to arise in various sectors. Widespread deployment of cameras and the use of facial recognition for surveillance was cited as an illustrative example of the possibility that technology developed with one application in mind (security) could unintentionally trigger undesired offshoots (intrusion of privacy). Other examples of the potential risks of AI related to bias in data and the tendency of recommender systems to polarise. In all, the discussion group unanimously agreed that the future course of AI had to be planned with much care, after accounting for unintended side-effects. One guiding principle for so doing would be to stop viewing AI as purely a technical discipline, but as one that overlapped with the humanities and social sciences. It was also suggested that protocols and standards be evolved across different applications domains, so as to proactively contain unintended consequences. Such an approach has been adopted in traditional branches of engineering, which have reached a favorable and relatively stable trajectory of growth.

In short, the session achieved two objectives: (1) reinforcing a professional network and facilitating research collaborations among 20+ outstanding young researchers in the area, and (2) drawing their attention two topics (the role of AI in education, the risks of AI-driven growth) that are very relevant to their unfolding careers.

 

Programme

Welcome and Introduction
 Uday Khedker (IIT Bombay)

Panel Discussion on “AI and Education”
 Moderator: Pushpak Bhattacharyya (IIT Bombay and IIT Patna)
 Panelists: Vibhu Mittal (Edmodo), Pradeep Ravikumar (Carnegie Mellon
University), Arun Kumar Giri (Paatham)

Group Discussion on “Risks of AI-driven Growth”
 Moderator: Shivaram Kalyanakrishnan (IIT Bombay)

Technical Presentations
 Manas Joglekar (Stanford University)
 “Generalization in Deep Learning”

 Parikshit Gopalan (VMWare Research)
 “Interactively visualizing large datasets”

 Ganesh Ramakrishnan (IIT Bombay)
 “Human Assisted Machine Learning: Consensus Driven Data Curation,
Domain Knowledge and Performance Measures”

 Tarun Yellamraju (Purdue University)
 “Clustering in High Dimensions with n-TARP”

 Sujay Sanghavi (University of Texas at Austin)
 “Bad Training Data”

 Sujata Banerjee (VMWare Research)
 “Smart Programmable Networks”

 Kartik Nagar (Purdue University)
 “Verification of Transactional Applications under Weak Consistency”

 Arjun Radhakrishna (Microsoft)
 “Programming by Example: Automating the Mundane”

 Rajhans Samdani (Spoke)
 “Breaking the Silos of Knowledge Search”

 Ankur Taly (Google)
 “Analysis of Deep Neural Networks”

 Manmohan Chandraker (University of California San Diego)
 “Perception and Prediction in 3D Scenes for ADAS and Self-Driving”

 Amod Jog (Johns Hopkins University)
 “Pulse Sequence Resilient Fast Brain MRI Segmentation”

 Rishabh Iyer (Microsoft)
 “A Unified Video Summarization Framework”

 Ankur Sharma (Iowa State University)
 “Fast Gate Sizing Using Lagrangian Relaxation on Multi-core Processor”

 Aranyak Mehta (Google)
 “Online Matching and Allocations in Advertising Markets”

 Aditya Bhaskara (University of Utah)
 “Low Rank Approximation in the Presence of Outliers”

 Bhumesh Kumar (University of Wisconsin-Madison)
 “Catch me if you can: A tale of Non-Stationary Stochastic
Approximation”

 Arun Sai Suggala (Carnegie Mellon University)
 “Robust Estimation via Robust Gradient Estimation”

 Shivaram Kalyanakrishnan (IIT Bombay)
 “Improved Bounds for Policy Iteration in Markov Decision Problems”

Participants (other than those listed in programme)

 Pandu Nayak (Google)
 Shantanu Thakoor (Stanford University)
 Anchit Gupta (Stanford University)
 Anand Dhoot (Stanford University)
 Raunak Bhattacharyya (Stanford University)
 Pradyot Prakash (University of Wisconsin-Madison)