Interview with Megha Srivastava, AI Resident
Tell us about yourself. Where and what did you study? How did you get into that field of study and when you started what did you think you wanted to do with your degree(s)?
Although I recently graduated with a B.S. in Computer Science from Stanford, I have always viewed my minor in Creative Writing and general love for the creative arts—including musical and visual art—as my main motivation for doing research in Artificial Intelligence. Our ability to create, improvise, and empathize is an incredible part of the human experience, yet despite impressive advancements, current AI research is still focused on optimizing a narrow view of “intelligence”. How can we design AI systems that, through interaction and collaboration, are better aligned with the complex behaviors and goals of humans?
My first research experience at Stanford was in a computational neuroscience lab, where I focused on comparing how humans and computer vision systems differ when recognizing novel objects. I observed first-hand how much more powerful the human ability to generalize to unseen images is, which encouraged me to become more interested in the issues seemingly successful AI systems may present, such as the amplification of demographic biases found in data. Moving forward, I will be returning to Stanford after the residency program to pursue a PhD in Computer Science, where I am interested in using ideas from both AI and HCI to more rigorously study how we can design better human-AI interaction.
So now you are here at Microsoft Research Redmond as an AI Resident. What attracted you to this program and how did you get here?
When I was applying to PhD programs last year, I spent a lot of time reflecting on how fast AI technology has been changing each year, both in terms of research advances and the way AI is used in society. I realized I wanted to take a step back before my PhD to think a bit harder about how we should design AI systems in a world where issues such as data privacy, free speech, and automation are influenced by the technological goals and ideas we set to pursue. Microsoft, with initiatives such as the AI and Ethics in Engineering and Research (AETHER) committee and a rich set of impactful products, seemed like a great place to consider the impact AI research can have.
What are the goals you set for yourself while you are an AI Resident?
I view my year at Microsoft as a chance to be a bit more “risky” in exploring research interests that, due to time or availability, I was unable to do during college. For example, apart from my main project, I have been spending some time learning more about causal inference and Microsoft’s use of causal modeling in initiatives such as precision agriculture with FarmBeats. I hope to use both the industry insights and the unique research perspectives Microsoft can provide to spark new ideas and inspiration for my PhD and broader research direction.
I have also been very excited about the extent of Microsoft Research’s involvement in the research community, and have been attending seminars and reading groups on a variety of research topics. Finally, I hope to learn directly from and build close relationships with the other AI residents, each of whom have really unique backgrounds and interests within AI research.
Can you tell us about the project and research you are current working on?
I am working with the Adaptive Systems & Interaction group on the idea of “backwards compatibility” of machine learning systems. In software engineering, backwards compatibility is a term used to design systems and software updates that allow for compatibility with older versions. However, we currently don’t really have a strong notion of this for AI systems—when we update a system with new data or features, even if it achieves high performance or accuracy, it may do so at the cost of introducing new errors. In high-stakes decision making, such as medical diagnostics, this presence of unexpected behavior can be particularly challenging for human users.
I was drawn to this problem because I believe a lot of modern-day AI research doesn’t strongly account for how humans are influenced by AI systems. In general, I want to work towards a world where people have more control over the technologies that use their data and become embedded in their daily lives—and supporting a user’s ability to trust and rely on a system is a big part of this goal.
Outside of your work as an AI Resident, what excites you? What do you spend your free time doing?
I absolutely love photography and creative writing, and am so excited to spend a year surrounded by amazing Pacific Northwest scenery and wildlife for inspiration. I am also enjoying exploring Seattle’s coffee scene—I recently discovered a technology-themed café named after Ada Lovelace!
What advice do you have for any up and coming AI/ML enthusiasts?
There are so many exciting blog posts and resources on the internet, but I think nothing is better than implementing a model yourself when it comes to learning—playing around with basic PyTorch or scikit-learn tutorials is an easy way to understand the different pieces of machine learning pipeline. I also think it is important to reflect on AI use-cases and data that are personal and interesting to you—for example, do you speak a language that you observe AI-powered translation systems struggle with, and are there particular patterns you can observe from examples? Finally, I have found Andrej Karpathy’s blog posts helpful in understanding deep learning concepts, and “the morning paper” blog (opens in new tab) for accessible summaries of interesting papers across computer science.
In this moment, what energizes you the most about the world of AI and ML?
What energizes me most, actually, is the growing engagement in AI/ML by people from areas outside of computer science and statistics—such as law, medicine, and economics. A lot of AI research is motivated by use cases in these fields—such as AI-powered medical diagnostics or credit scoring systems—but I think it is crucial that this research is accompanied with the perspective of domain experts themselves. Understanding how to design better human-AI interaction necessitates an interdisciplinary approach to AI, and I appreciate how Microsoft’s breadth in research and products helps to foster this.