Week of 11-15 June
Day 1 – 11-June
10:00 AM – 11:30 AM : Welcome / Setting Context
Introduction to Microsoft Research India, the IDE project, the workshop format. Introduction of the teams, committee and Research Fellows working on the IDE project. Goals /expectations from the workshop – to build connections between algorithms and practice; to being together domain experts working in the embedded ML space.
2:00 PM – 4:30 PM: Priyan Vaithilingam, Microsoft Research – Azure Hands-on Tutorial
Get participants familiar with using Azure subscriptions and the services available on their Azure accounts.
Day 2 – 12-June
10:00 AM – 11:15 AM : Harsha Vardhan Simhadri (opens in new tab), Microsoft Research – The EdgeML library
In this talk, we will present two supervised learning algorithms Bonsai (tree-based) and ProtoNN (nearest-neighbors based) suited for resource-constrained devices such as IoT devices, sensors and embedded devices. These algorithms can generate models that are order(s) of magnitudes smaller, faster and power-efficient (for inference) than popular algorithms such as SVM, GBDT, etc. while still providing competitive accuracies. The talk will include the technical details of the algorithms, their availability and use.
2:00 PM – 3:15 PM : Prateek Jain (opens in new tab), Microsoft Research – Recent work on FastRNNs and multi-instance learning+early stopping in RNN.
In this talk, we discuss recent results on training RNNs with smaller memory and computation footprint for supervised learning tasks on time-series data such as audio wake-word detection, keyword spotting, and gesture recognition. The first part of the talk will discuss new RNN architectures.The second part of the talk will discuss new results on learning from noisy data in which the true time-series signature of a class is a small subsequence of a larger training example. This kind of training data is typical of time-series tasks. For example, the positive examples in a wake-word detection training data set might include occurrences of the wake-word in a noisy background sample of length much longer than the utterance of the wake-word. We also discuss how to stop RNN steps early on instances where no signature is found.
Day 3 – 13-June
10:00 AM – 11:15 AM: Applications: Interactive Cane, and other Time-series tasks
In this talk, we will discuss our experience with the interactive cane and other time-series tasks. The interactive cane is an input device for people with visual impairment to do quick tasks on their smart phones. On the hardware side, it consists of a small pod with a microcontroller and IMU sensor on the cane. We trained and deployed a model that predicts gestures (from a pre-defined set) on the cane and communicates the user intent, as detected through gesture, to the smart phone. We will share our experiences in collecting training data, developing the features, and user-studies of the prototype. Separately, we will share our experience working on small audio tasks on Cortex A series chips.
2:00 PM – 5:00 PM: Presentation from the groups on their plans for the workshop (40 minutes each)
Day 4 – 14-June
10:00 AM – 11:15 AM: Anish Arora (opens in new tab), Ohio State University – Machine Learning from Limited Signals at Mote-Scale
This talk describes the growing role of machine learning in battery-powered wireless sensor networks being deployed in rural and urban areas for environmental and wildlife projection. Specifically, we will discuss radar-based HornNet, deployed in a South African rhino reservation for anti-poaching surveillance, and microphone-based SONYC, deployed in New York City for sound complaint mitigation. Characteristic to these applications are classification and counting tasks with signals that are limited in the sense of being low signal to noise ratio (SNR), with significant spatio-temporal variation across different background clutters, rejection of clutter which may or may be accurate, and potential for interference. We will present how our mote-scale implementations have and are dealing with these challenges.
2:00 PM – 3:15 PM: Nagarajan Natarajan, Microsoft Research – Predictive Maintenance in Manufacturing: A Case-Study
Ensuring uptime of machines in manufacturing plants, and predicting faults early on in the production pipelines are crucial problems, that have a direct impact on productivity and in turn on the revenue of the manufacturing industry. Typically, several sensors are deployed in the factories that help monitor the status of the production, as well as enable fine computerized control of the production machines. The goal in predictive maintenance is to enable the plant to allow for convenient ahead-of-time scheduling of maintenance of equipment (and of maintenance engineers), predict imminent faults (production defects or machine downtimes), and thereby ensure throughput of production pipelines. In this talk, I’ll present lessons, techniques and results from the case-study we did with Magna Corp, a leading manufacturer of automobile parts — nature and issues with data arising in a typical industrial predictive maintenance setting, the type of solutions required (edge vs cloud), challenges in formulating the predictive maintenance problem as a machine learning (anomaly detection, in particular) problem, and our solutions for the same.
Week of 18-22 June
Day 7 – 19-June
10:00 AM – 11:15 AM: Bharadwaj Amrutur (opens in new tab) – Chairman, Robert Bosch Centre for Cyber Physical Systems (IISc Bangalore)
A case for an open City Data Exchange and Edge Analytics Stack:
The Indian government is attempting an ambitious aspirational move to “smarten” India’s cities. We do a deep dive analysis of a recent RFP for the same, which has been put out by the city of Agra. From this we extract the contours of a generic, but foundational core – which is a data exchange and edge analytics layer – which we believe should be developed as an open stack. We delve into some details of how such a stack might look – with some examples from an ongoing reference implementation and point out to some interesting technology/research challenges.
Bio: Bharadwaj Amrutur’s recent research is in the area of large scale IoT systems, especially to support AI based autonomous systems. His prior work was in the area of low power VLSI. He is an alumnus of IIT Bombay and Stanford and is currently at IISc Bangalore, as a Professor in ECE dept and Chair of the Robert Bosch Center for Cyber Physical Systems.
Day 8 – 20-June
10:00 AM – 11:15 AM: Tanuja Ganu (opens in new tab), Co-Founder & CTO, DataGlen Technologies
Bits and Joules: Empowering energy consumers through IoT & AI
The Energy vertical is going through a paradigm shift from centralised conventional generation and distributed consumption to non-conventional distributed generation (with renewables and battery technologies) and distributed consumption. This change is posing challenges and driving opportunities and innovations using new age digital technologies such as IoT and machine learning. This talk would describe various open problems and ongoing work using edge analytics and AI in energy vertical, including decentralised demand response, predictive maintenance for various equipments, renewable energy forecasting and resource optimization.
Bio: Tanuja Ganu is the Co-Founder and CTO of DataGlen Technologies Private Limited, an early-stage startup that focuses on achieving decarbonization and rapid adaption of distributed and renewable energy resources using IoT & Machine Learning technologies. DataGlen was one of the top twelve global energy start-ups selected for FreeElectrons accelerator program and was also selected to participate in the Cisco Launchpad program.
In the past, she has worked as Research Engineer at IBM Research, India. At IBM, her work was focused on developing low cost innovative solutions to address energy shortage and peak demand problems that are applicable for developing as well as developed countries. With research interests in machine learning, embedded analytics and data driven optimisation, she has published more than 20 scientific publications and has 4 granted US patents.
Tanuja has been recognized as MIT Technology Review’s Innovator Under 35 (MIT TR 35) in 2014. She has also served on the judges committee for MIT TR35 2015, 2016 and 2017. She has won IBM Eminence and Excellence award and IBM first invention plateau award. Her work was covered by top technical media (IEEE Spectrum, MIT Technology Review, IBM Research blog and Innovation 26X26: 26 innovations by 26 IBM women). Recently she was also invited as guest speaker for Cisco Women Rock IT TV series, ACM N2Women Event and Cisco SecCon 2017 security conference.
Prior to joining IBM Research, she led SharePoint Center-of-Excellence team at Tata Consultancy Services Ltd. She pursued her Masters in Computer Science at Indian Institute of Science (IISc), Bangalore and Bachelor in Computer Science at Walchand College of Engineering, Sangli, India.
Day 9 – 21-June
10:00 AM – 11:15 AM: Sriram Rajamani (opens in new tab), Managing Director, Microsoft Research India
Overview of Microsoft Research with specific focus on the research and other initiatives of the lab in India.
Week of 02-06 July
Day 18 – 04th July
3:00 PM – 04:15 PM: Vyas Sekar (opens in new tab) , Carnegie Mellon University
Rethinking IoT Analytics with Universal Monitoring
Many IoT analytics tasks require accurate estimates of metrics for many applications such as heavy hitters, anomaly detection (e.g., entropy of source addresses), and security (e.g., DoS detection). Obtaining accurate estimates given CPU, memory, energy, and bandwidth constraints on IoT devices is a challenging problem. Existing approaches fall in one of two undesirable extremes: (1) low fidelity general purpose approaches such as sampling, or (2) high fidelity but complex sketching algorithms customized to specific application level metrics. Ideally, a solution should be both general (i.e., supports many applications) and provide accuracy comparable to custom algorithms. In this talk, I will present our recent work on leveraging recent theoretical advances in the area of “universal sketching” to demonstrate that it is possible to achieve both generality and high accuracy. Our solution called UnivMon uses an application-agnostic data plane monitoring primitive; different (and possibly unforeseen) estimation algorithms run in the control plane, and use the statistics from the data plane to compute application-level metrics. I will describe our experiences in using this for network-flow monitoring and highlight interesting directions for future research in the IoT analytics domain.
I will also provide a brief overview of: (1) a new project effort called CONIX (conix.io (opens in new tab)) that aims to provide a new middle tier of distributed computing that tightly couples the cloud and edge by pushing increased levels of autonomy and intelligence into the network and (2) interesting applications of machine learning to IoT security and privacy.
Bio: Vyas Sekar is the Angel Jordan Early Career Chair Associate Professor in the ECE Department at Carnegie Mellon University, with a courtesy appointment in the Computer Science Department. His research is in the area of networking, security, and systems and spans network appliances or middleboxes, network management, network security, Internet video, and datacenter networks. Vyas received a B.Tech from the Indian Institute of Technology, Madras where he was awarded the President of India Gold Medal, and a Ph.D from Carnegie Mellon University. He is the recipient of the NSF CAREER award and the ACM SIGCOMM Rising Star Award. His work has received best paper awards at ACM Sigcomm, ACM CoNext, and ACM Multimedia, the NSA Science of Security prize, the CSAW Applied Security Research Prize.
Day 20 – 06th July
11:00 AM to 01:30 PM : Final presentations from the groups