Demonstration 1
COSWARA: A Web Application for Data Collection and Real-time Screening of SARS-CoV-2 Infected Subjects
Speaker:
- Debarpan Bhattacharya, Indian Institute of Science
- Neeraj Sharma, Indian Institute of Science
- Srikanth Raj Chetupalli, Indian Institute of Science
- Sriram Ganapathy, Indian Institute of Science
Abstract:
The primary symptom of COVID-19 disease is respiratory discomfort. The pandemic is the most serious international health emergency since the 1918 influenza epidemic. A wide range of information has been gathered to understand how the biggest health crisis of the century affected individuals’ respiratory health in particular. While vaccinations have provided a first line of defence, widespread and ongoing screening remains one of the most important measures to lower the incidence of COVID-19. It is necessary to design a variety of screening procedures that offer a trade-off between cost, testing simplicity, testing speed, and accuracy. In this study, we examine the potential for COVID-19 screening utilising respiratory acoustics, which may be quickly deployed and scaled to a large population. The conception of the data collection strategy, as well as the creation and evaluation of machine learning models that permit COVID-19 screening using audio samples, are presented in this study. The creation of a classifier based on recurrent neural networks for the screening of COVID-19 utilising respiratory sound samples is described in the study. The classifier, which is installed in a cloud-based server, quickly estimates the likelihood of SARS-CoV-2 infection. The screening tool is generally unaffected by the subject’s age, gender, geography, language background, and pre-existing medical issues, according to a bias and fairness analysis.
Demonstration 2
vaAnI - an AI-powered voice-enabled search engine for the visually-challenged community
Speakers:
- Shirsha Ray Chaudhuri, Thomson Reuters
- Prudhvi Nokku, Thomson Reuters
Abstract
EnAble India is a NGO that provides a platform for People with Disabilities to come together and resolve their challenges. It provides many learning facilities, self-help channels, employment training and social navigation aids for People with Disabilities. Enable Vaani is one such platform that this NGO setup, using which the blind population from more than 15 states in our country are able to call in for their asks and get responses. This is a voice-enabled question answer, information retrieval, crowd-sourced recommendation system for the blind. This recommendation system is today executed manually. By this we mean, that mediators work on queries received and match them to a best possible response received in the past. This workflow demands automation with the use of AI. This solution demo shall showcase a solution provided for this platform, where AI-powered services in the cloud are leveraged to provide automation into this workflow, for this platform. Mediators receive auto-suggestions from the system for a given audio query and can select the best-fitting one. Past audio responses and possible suggestions are stored translated, indexed and parsed, so that they may be used for responding against a given query. It is a voice-enabled search and recommendation engine for the visually challenged for their resources.
Demonstration 3
Flexible AutoML: Accelerating AutoML adoption across Amazon
Speaker:
- Abhishek Divekar, Amazon
- Gaurav Manchanda, Amazon
Abstract:
Artificial Intelligence/Machine Learning is facing many hurdles in its journey towards becoming a daily occurrence in our lives. One of the biggest challenges is handling large multi-process, multi-component and multi-algorithm Processing Pipelines in Machine Learning/Ai - and doing so without everything falling apart! It is very easy to build a single Machine Learning or Deep Learning model. The real challenge is to build multiple such models that work together, work in a coordinated manner, and in the end, produce usable, coherent results - that is what will be “True Ai”. The FlAi MLOps Suite is an end-to-end Drag-and-Drop SaaS platform that enables users to create their own Machine Learning pipelines running on Kubernetes. FlAi allows for high-level container-customisation while completely automating the code needed to provision the Infrastructure of the Machine Learning/Ai Pipeline - keeping the user (the Data Scientist) in full control of his/her own Data Science code. The Drag-and-Drop functionality, inherent containerisation, rapid object-level (block and pipeline level) replication, collaborative development all gear up the user towards massive, yet manageable Ai/ML pipelines. Inherent benefits of using Kubernetes under-the-hood include multi fold increase in processing speed, failure-redundancy and container-level replication for components of the Pipeline. It is due to FlAi’s abstraction layer, that even a Data Scientist can easily understand, and exploit the power of Kubernetes without having to necessarily endure the pain of configuring and working it out themselves.
Demonstration 4
Automating NLP workflows using AutoML NLP
Speakers:
- SHIPRA JAIN, Microsoft R&D India Pvt Ltd
- Chandra Sekhar Gupta Aravapalli, Microsoft R&D India Pvt Ltd
Abstract:
With recent advancements in technology, everything from the purchase of electronics to household items is happening online at massive scale which results in thousands to millions of data points. Needless to emphasize the importance of reviews and that they can take our scale of growth in an exponential manner. According to the study from Maslowska et al. and the study from Askalidis et al. of online customers read reviews before making a purchase. Its significant for business to address unsatisfactory reviews by providing them a prompt response or by taking feedback to the manufacturer. However, what adds to the complexity is humungous multi-lingual data received from disparate data sources and the complex free flow data semantic in the natural language involved. Thus, there is a need for automated analysis techniques to extract feedback and opinions conveyed in the user-comments. With the products and services increasingly graduating from being data-driven to “data and intelligence” driven and progressing towards more AI-driven world, leveraging AI and automating the solution with AI-driven applications is the key to meeting that need. In this segment we would see how Automated ML solution can come in very handy to quickly train and pick best of the class machine learning model to accept multi-lingual Natural Language data, classify it, and extract meaningful information which is leading to unsatisfactory customer feedback. Thus, enabling manufacturers to harness true potential of their data and make informed data-driven decisions.
Demonstration 5
We built a wall for the better – running online ML experiments at scale
Speakers:
- Shahram Anver, Gojek
- Krithika Sundararajan, Gojek
Abstract:
In the real world, any serious production system must account for performance, scalability, observability, and a slew of other non-functional requirements, to be successful. This is no different for ML systems and the typical Data Scientist either takes on some of the Engineering burden or, in more established companies, relies on a team of savvy Engineers that make the (Jupyter) notebook science ‘production-ready’. But, the relationship between the Engineer and the Data Scientist is not an easy one to navigate – it is fraught with many back-and-forths, misunderstandings, delays and disappointments. That’s hardly surprising given that one desires stability and predictability while the other loves to experiment.
This motivated us to look for a solution and build an MLOps product that would give both camps what they wanted. One such area was model experimentation. We surveyed multiple Data Science projects at Gojek that conducted experiments and realized we had a strong opportunity to standardize. While they each had diverse goals, their requirements and workflows were common enough to extract into a set of abstractions. This would help us provide data scientists the flexibility and control they desired while simultaneously adhering to good engineering standards. And thus, we created Turing, a multi-model orchestration and experimentation platform that is also completely self-service
Demonstration 6
MLOps with enhanced performance control and observability
Speakers:
- Dinesh Mr Ghanta, Oracle
- Prateek Katageri, Oracle
- Girish Nautiyal, oracle
- Indradumna Banerjee, Oracle
- Pradeep Sanchana, Oracle
- Atin Modi, Oracle
Abstract:
The explosion of data and its ever increasing complexity in the last few years, has made MLOps systems more prone to failure, and new tools need to be embedded in such systems to avoid such failure. In this demo, we will introduce crucial tools in the observability module of a MLOps system that target difficult issues like data drift and model version control for optimum model selection. We believe integrating these features in our MLOps pipeline would go a long way in building a robust system immune to early stage ML system failures.
Demonstration 7
EEG Data Visualisation for Blinking Eyes
Speakers
- Darshil Shah, MINRO
Abstract:
Electroencephalography is a technique used to measure the electrical activity in the brain. Electrodes are placed on the scalp of the subject and electrical signals (of varying amplitudes and frequencies) generated from the brain are captured. EEG signals consist of various frequency sub-bands: 1–4 Hz (Delta waves), 4–8 Hz (Theta waves), 8–13 Hz (Alpha waves), 13–30 Hz (Beta waves), and 30–60 Hz (Gamma waves). EEG, being non-invasive, are widely used to assess neurodegenerative conditions, mental load, and sleep patterns. In this work we will demonstrate EEG Data Visualisation for Blinking Eyes. Enobio8 from Neuroelectric is used to capture the EEG data. EEG data is recorded from Pre-Frontal region (FP1 and FP2 locations) which has higher activities during Blinking Eyes state. NIC2 from Neuroelectric is used to visualise EEG Signal, Band Power, Spectrogram and Scalp Map.
Demonstration 8
Sahasraaksha: Video analytics for smart campus surveillance
Speakers:
- Sriram G, MINRO
Abstract:
ISurveillance cameras are ubiquitous nowadays, but the rich visual data is not harnessed enough. Video analytics has extensive applications in traffic, property, and crowd management. We have developed a robust, real-time, and scalable video analytics solution for campus surveillance with advanced deep learning methods. We have features intrusion alerts, dynamic prioritization of camera feeds, number plate recognition (even with general cameras), no parking detector, person image retrieval using text-based description, and detailed analytics and charts depicting the numbers. The product has been implemented at IIIT-Bangalore and NIMHANS. Check out our website for more details.
Demonstration 9
MLOps at FIAI
Speakers:
- Ruju N Alurkar, FlAi MLOps
Abstract:
Artificial Intelligence/Machine Learning is facing many hurdles in its journey towards becoming a daily occurrence in our lives. One of the biggest challenges is handling large multi-process, multi-component and multi-algorithm Processing Pipelines in Machine Learning/Ai - and doing so without everything falling apart! It is very easy to build a single Machine Learning or Deep Learning model. The real challenge is to build multiple such models that work together, work in a coordinated manner, and in the end, produce usable, coherent results - that is what will be “True Ai”. The FlAi MLOps Suite is an end-to-end Drag-and-Drop SaaS platform that enables users to create their own Machine Learning pipelines running on Kubernetes. FlAi allows for high-level container-customisation while completely automating the code needed to provision the Infrastructure of the Machine Learning/Ai Pipeline - keeping the user (the Data Scientist) in full control of his/her own Data Science code. The Drag-and-Drop functionality, inherent containerisation, rapid object-level (block and pipeline level) replication, collaborative development all gear up the user towards massive, yet manageable Ai/ML pipelines. Inherent benefits of using Kubernetes under-the-hood include multi fold increase in processing speed, failure-redundancy and container-level replication for components of the Pipeline. It is due to FlAi’s abstraction layer, that even a Data Scientist can easily understand, and exploit the power of Kubernetes without having to necessarily endure the pain of configuring and working it out themselves.
Demonstration 10
Building Reusable AI Systems
Speakers:
- Vikas Raturi, Intuit
Abstract:
AI capabilities are now the norm in the Industry for building intelligence into products and providing next-gen customer experience. Many such AI solutions however are tailor made to different use-cases and require resources to build and integrate the solution with products. This process becomes highly inefficient as more solutions are built to solve similar problems. Reusable AI Services bridges this gap by providing a well-defined AI capability as a Service allowing more users to adopt AI with much lesser resources. This talk is about our experience in building early reusable AI systems built for faster product integration.
Demonstration 11
Recommendation system for ML model deployment
Speakers:
- John Barboza, Amazon
- Mohan Gandhi, Amazon
Abstract:
Configuring a deployment setup to run Machine Learning models that meets performance at scale is a hard problem to solve. This requires knowing what types of hardware are available, their strengths and weaknesses, trade-offs between performance and cost, tuning parameters of the framework etc. Inference Recommender is a service that aims to take away this burden from the ML customers. It will recommend the best performing instance and tuning parameters to get the lowest cost per inference call.
