Day 3 Keynotes & Panels

Building AI & Data products requires strategizing across data collection & scaling techniques, accounting for uncertainty in AI outcomes, and building feedback loops for the AI flywheel effect. In addition to ML techniques strategies, MLOps and Data architectures are also critical. This talk will go through frameworks, challenges and best practises on these topics.

Being data-driven organisation is table stakes in today’s hyper competitive environment. In this talk, we will start by sharing our Data Platform strategy at Udaan, which has enabled us to unlock value from our data by embedding data driven decision making in all workflows. Then we will touch upon the worrying trend that nearly 85% of AI/ML models don’t make it to production and will share personal experience on how to mitigate the same. We will end by sharing the data science problem landscape for B2B e-commerce and will go through one use case demonstrating the various ML Ops capabilities that has enabled successful deployment of sophisticated, state-of-the-art ML model.

AI is undoubtedly impacting the way of life, though not without its challenges. ‘Life 2.0’ is what is believed to emerge as this technological revolution advances. At present, Governments globally are looking to invest in it as well as blend it into policies that will not only help economic growth but individual corporate growth. When considering other stakeholders, questions now arise on what impact it currently has in our day-to-day, what can corporations do (particularly where AI and diversity is concerned) and how can we ensure that this technology is used responsibly.

Machine Learning (ML) has quickly become ubiquitous in banking for both predictive analytics and process automation applications. However, banks in the US remain cautious in adopting ML for high-risk, regulated areas like credit underwriting due to key concerns including ML explainability and robustness. To address the inadequacy of post-hoc explainability tools in XAI for critical applications, we have developed inherently interpretable machine learning models, including Deep ReLU networks and Functional ANOVA-based ML models. Model robustness is a key requirement as models will be subjected to a continually changing environment in production. A conceptually sound model must be able to function properly – without continuous retraining – while remaining relatively invariant in a changing environment. Recently, we released the PiML (Python Interpretable Machine Learning) package as a tool to design inherently interpretable models, and to test machine learning robustness, reliability and resilience.

AI-driven facial recognition software is probably the most divisive technology deployed today. This is not only because of the social implications it can have on individuals or certain sections of society but also because the technology itself has proven to be unreliable. Several accidents or instances of misuse have led law enforcement agencies to shut down its use in the US and other countries. While that is the case, Indian governments are on a spree to implement automated FRTs. Also, India doesn’t have the appropriate legal frameworks to avoid its implications, such as mass surveillance and misuse of personal data, or accountability frameworks when it works erroneously. All socio-technical factors considered, should FRTs be banned altogether in India?

While we understand the cognitive (and logical) algorithms “on the surface of the mind” how much do we understand the unconscious part of our mind and personality? While exploring the unconscious in human beings itself is a difficult task, would AI be able to incorporate both conscious and unconscious part of our awareness?

The session looks at Human unconscious and how it influence our behavior and explore what could be a future possibility where machines could also have an unconscious. Santhosh also would be demonstrating hypnosis in this session on some volunteers