Women in AI Interview, Anusha Balakrishnan: Conversational AI at Facebook

Conversational systems like Siri and Google Assistant, have been around for quite a few years now and have recently started to play increasingly ubiquitous roles and many people's daily lives, through smart home devices, phones or social media. Despite this, the conversational experience that these systems provide has evolved only incrementally. We spoke with Anusha at the Women in AI Dinner in San Francisco, on the Women in AI Podcast, to hear about how she's working as a research engineer at Facebook on the conversational AI team. Here is the transcript of our chat.

Yazmin: So hey, Anusha, thank you so much for joining me today. Before we get started, can you introduce yourself? Tell me about your current role?

Anusha: Yeah, sure. So I'm a Research Engineer at Facebook. I work on natural language generation for conversational AI and I've been at Facebook around one and a half years now.

Yazmin: Okay, amazing. So can you give me a bit of an overview of your background? How did you start your work in conversational AI and just AI more generally?

Anusha: Yeah. So during my undergrad I went to Columbia, I studied computer science and linguistics. I was always really interested in language, and just how people converse and discourse, things like that. I was Googling one day and there was like this thing called natural language processing, I was like, oh, that sounds really cool. I like computer science. Maybe I should learn more. I reached out to this professor at my university who had some research positions open and just kind of started learning more about NLP things like that, and then after my undergrad, I got to work at Apple for six months at Siri, which was really great. That was like my first real introduction to the conversational AI space and assistants and things like that. So after that, when I went on to my Masters at Stanford, I was like really keen on just continuing that line of research and so I reached out to some professors. There I worked with Percy Wang, on dialogue systems and adding like knowledge to, and reasoning, to dialogue systems, which was really exciting. Then it just kind of naturally led to like my role at Facebook, my team works on conversational AI and so it was just like a really good fit.

Yazmin: Yeah, that's really exciting. So what was your master's in?

Anusha: In computer science, with a specialization in artificial intelligence.

Yazmin: I know you mentioned that you started Googling and you saw about natural language processing, was that when you first became aware of AI? And natural language processing more generally?

Anusha: Yeah, that was pretty much the first time, just online, and then I just looked online at my course listing and I found that there was a couple of classes that were being offered, this was almost like pre deep learning, I would say, so it was just like very, very basic stuff. But it was really exciting to me. So I just took a few classes. I don't think, because I was still an undergrad. I don't think I learned a lot lot from them, honestly. But it was still really exciting to just kind of get all that knowledge and I didn't get a lot of depth, but it was a very good introduction to the field.

Yazmin: And just to be aware of everything. That's good. Yeah. What does your role as research engineer at Facebook entail at the moment?

Anusha: Yeah, so I can't really talk about very specific projects. But the basic goal of my work and our team is to enrich the way that people converse with compositional systems, and specifically thinking about how human language interacts with either Facebook hardware or Facebook surfaces. And for me, personally, a lot of my work is kind of bridging the gap between research and product. So thinking about all of the different exciting advancements that are happening in the field, contributing to those but then also advancing, using those advancements to actually add to product and thinking what how can we make those solutions practically scalable? How can we put them in the hands of Facebook users.

Yazmin: Yeah, definitely. There are so many challenges still present in conversational AI, even though people have made huge advancements in recent years. But what are the main challenges that you're still facing? And why do they exist?

Anusha: Yeah. Oh, yeah, there are so many. So I think one of the biggest ones honestly, it's just language is so complex. And it's almost like an onion, every time you peel off one sub problem that you think you solve, you get to this entire, set of new things that you've never thought about. So just the way that people talk is so complex and in designing a conversational AI system, you have to account for all of that. Something else that has been recently happening, which is really exciting, is people are actually trying to think about when we use language, there are so many implicit sources of knowledge that we draw on and a lot of it is just common sense. knowing that apples are red or knowing there are 7 billion people in the world. Then how do you incorporate all of that information into a model, that's actually supposed to converse with people at a human level? That's something that's really, really hard to do. But I think there are interesting advancements being made on that front now and another really big challenge, honestly, is just having the right datasets to even evaluate success on, because nearly every single time that people come out with a good data set, and it's really exciting, but then there are all these findings that oh, maybe this data set is still too narrow. It's just because conversation is such a broad space and so it's really hard to know how to focus but I think there are like good studies coming out now on trying to establish good benchmarks and good, actually good, concrete, well rounded data sets. I think we're making good progress on that front.

Yazmin: And I think, like you're saying, language is so varied. It's so unique and personalized, as well as thinking about how can a model account for every single use case and every single different person so complicated.

Anusha: Exactly.

Yazmin: We obviously hear a lot in the news and in the AI world as well at the moment about using AI for like a positive social impact. I can think of loads of different ways that conversational AI can help here. But what are some of the other industries that you can see benefiting massively from the kind of work you're doing and the work that's being done in this space at the moment?

Anusha: Yeah, so for me personally, I think the things I'm most excited about are these two. So the first one is just the impact of conversational AI specifically on wellness and just like improving lifestyle. So I think there are a couple of different there are quite a few different companies and like research efforts on this so, you know, improving people's mental wellness, for example, providing a trusted source of companionship, providing a confidant almost, even like coaching people on lifestyles, diets, things like that. The other thing that I'm really excited about is just the impact of conversational AI on, educational and informational applications. So, for example, coaching students in almost like a virtual classroom kind of setting, which I think that goes back to what you were saying about it, mass personalization, if we can figure out how to make language very personalized, then we can figure out how to give students unique and personalized learning experiences, which is what teachers give in a classroom experience. Then also, just purely informational things, improving knowledge about voting rights, and political platforms, things like that. So I think there's definitely a lot of space for conversational AI that's engaging, to really improve how people access information and interact with it.

Yazmin: Yeah, definitely and I guess another big thing in the news at the moment has been about the ethical implications of AI and how we can stop bias creeping into these systems. Do you think we're close to being able to eliminate bias and conversational AI? Or do you think that is something that is possible to be done or is it just too complex?

Anusha: I don't know if we're close. I think it's possible. But I think the biggest challenge still that I see, and there's a lot of good progress in this direction, but is that we don't actually still know how to quantify this kind of bias in meaningful ways. I think that's, before you can eliminate bias, you have to know how to identify and quantify it and so some of the studies that are coming out now, about, first, just even starting with a data set and looking at bias in a data set, and then seeing whether models trained on that data set actually exhibit that same kind of bias and coming up with good frameworks and mental models for even quantifying that bias, I think is really important. There's been good work on that front and there's been some really exciting work on even trying to eliminate that bias. So I don't think it's impossible, but it definitely does require a lot of like, very focused thinking on that space, which I do see now. Another challenge is that, people inherently have so many biases that and that's why I think it's so important to have a lot of different people thinking about the same problem. Conversational AI, AI in general, I think just need so much diversity. Because the more diverse set, the more diverse the set of people working on AI is, the easier it is to call out biases and things like that.

Yazmin: Definitely. Obviously, you've mentioned diverse, kind of types of people with different backgrounds working on AI. But do you think there any other ways that we can ensure that AI doesn't kind of inherit these intrinsic faults of humans?

Anusha: Yeah, I think coming up, honestly, so going back to just coming up with like framework to identify bias, I think can really help us in building data sets that don't exhibit bias. This is kind of hard, because I think as humans, we are always looking at things through a bias lens. So what does it really mean for someone to look at a data set and say this data that has no bias? I'm not sure what that looks like.

Yazmin: Because you've got a biased opinion, whoever you are, looking at it

Anusha: Exactly, but if we can come up with, numbers don't lie. So if we can come up with metrics that actually quantify that, and then we can say, hey, this data set has like a score of zero on bias based on this metric that is empirically proven, then I think that's a step in the right direction. Again, it's a lot of these things are, they're like chicken and egg almost and it's really hard to find, but I do think it's possible and I think attacking it from data is definitely one way to think about it. Another way is, there are probably like modeling techniques that we can actually use and some of this has been started to be tapped into, like there's some recent work from Google on devising word embedding, for example, and word embedding such a foundational thing to almost any NLP task and if you can quantify what bias is in a word embedding and then de-bias them, that I think that's a really foundational step. And I think more like fundamental advancements like that can really help.

Yazmin: And obviously in order to have such a diverse range of opinions coming into the like programming and everything like that, we need lots of different backgrounds in AI. So how do you think we can encourage, firstly, more young girls and women into AI, but also just people from lots of different types of backgrounds? And what do you think needs to be done there to help encourage these people into the space?

Anusha: Yeah, I think something that can really help and again, this is also a really hard problem. But I think where solutions are more imminent, I think one thing that can really help is just creating more awareness, just of what it means to work on AI because I think that's something that honestly even I didn't really know what that meant until I got like a research opportunity, which if you think about how many people have access to that kind of opportunities, is so little. So creating just awareness, going to high schools going to, you know, middle schools, going to college campuses, just talking about what it means to work on AI and honestly, it's such a fun field and there's so much scope for impact and it's just one of the best fields to work in right now, if you're interested in computer science. So I think just creating that kind of awareness is really important. Another thing that I found, in my own personal experience, and from talking to other people is, if you think about, any kind of AI class, they can seem really intimidating because they have, so many prerequisites, you know, and you look at those, you look at like the class reading list, and you look at the set of assignments, and you're like, wow, I could never get through one of those classes. And I think, in general, in college campuses, there's been a lot of effort to kind of demystify introductory programming classes. And so you know, for example, making programming classes that are geared towards liberal arts majors, making programming classes that are geared towards people who've never coded like a day in their life. I think AI needs something similar, if you can build more AI classes that are focused, maybe more on the applications and have less of an assumption on certain theoretical knowledge. I think that could be really impactful and that could also help encourage more people to take that up.

Yazmin: Definitely. Yeah. So what's next for you and your work? What are you going to be doing? Maybe say 12 months? Where do you hope to be?

Anusha: Yeah, I think, just in my work, we've been making a lot of really interesting advancements on language generation and I think we're making some cool discoveries around, you know, how do you incorporate knowledge into models for language generation, how do you make them sound more natural? How do you make their responses more diverse? So pushing on that, I think, is really exciting. And for me personally, like I would love to be able to publish some papers on that and just honestly, just keep learning more from everyone else's in the industry, because it's such a large space right now that I think there's just so much scope to learn, there's so much cross pollination for me.

Yazmin: Yeah. And it's still growing so quickly, isn't it? So there's always more to learn, but there's always knowledge that you can share with other people.

Anusha: Exactly. Yeah.

Yazmin: Thank you so much for joining me. And it's been great to chat. But where can we kind of keep up to date with you? You mentioned that you might be publishing some papers but do you have like a Twitter or website or anything where people could check out your work?

Anusha: Yeah, my Facebook AI website is probably the best place, but I'll try to keep updating. As I like publish more and stuff. I also have a Google Scholar page to kind of follow my publications.

Yazmin: So amazing. Thank you so much. Yeah. Thank you for joining me.

Anusha: Thank you so much.

Listen to more episodes on the weekly Women in AI Podcast here. We meet with leading female minds in AI, Deep Learning and Machine Learning. We will speak to CEOs, CTOs, Data Scientists, Engineers, Researchers and Industry Professionals to learn about their cutting edge work and advances, as well as their impact on AI and their place in the industry.