5-minute interview Abi Aryan

Once I went back to the industry, I realized that there are not enough people who know how to deploy models in production. A lot of data scientists just come with data science skills, and that's when I sort of transitioned into doing ml engineering and data engineering work.

Meet Abi Aryan, ML engineer and founder of Abide AI. She tells us about her background in research but also why she decided to go back to working in the AI industry and what she finds most fascinating about AI.

Tell us a little bit about yourself

Abi: My name is Abi, I'm currently a machine learning engineer. I didn't actually start there, but I’ve been in the field of data science and machine learning for about eight years now. Currently, I'm the founder of Abide AI. We're working on productionizing machine learning so that AI can be integrated into all the technology and hardware around us.

The key idea behind naming it Abide AI is because machine learning models are stochastic in the way they operate. One of the big problems with those is, and especially how ML is scaling up with deep learning primarily, we don't really have solutions around interpretability. That's something which is still a hard problem. It was like if you're building this technology, the first thing that we need to think about is making sure that it is always constrained, keeping an alignment of people who are working on evaluations.

How did you get into the field of MLOps? 

Abi: So I was doing my Master's in Mathematics and Computer Science, I also studied some Behavioral Finance. But for my dissertation thesis I was thinking about what topic I wanted to work on, and I was studying network theory (basically how information travels along different routes in the universe) and I found it very similar to how information travels within our brain and thinking of it as very similar to neural networks. So I ended up doing my thesis on image recognition using a very different kind of neural network. I was working on Hopfield networks, which is basically an energy based neural network as compared to the deep learning models that we're seeing today. So, Hopfield models were sort of popular in the 1990s. Energy based models are an area where there's not enough research in the field, so I started working with that and eventually I started working in data science doing Business Intelligence stuff, then doing data analytics stuff and then I went into research. I eventually realized that research is fun, but at the end of the day, I'm just writing papers. So, let me go back to the industry. Once I went back to the industry, I realized that there are not enough people who know how to deploy models in production. A lot of data scientists just come with data science skills, and that's when I sort of transitioned into doing ml engineering and data engineering work. So now I've been doing a combination of this for the past three years. I've been involved in ML research but I also have been building data pipelines and doing MLOps stuff as well.

Why is MLOps such an important field?

Abi: So for me, I was always fascinated with Behavioral Sciences and thinking about how we can optimize human performance. That sort of got me into the field of AI because at some level I wanted to understand how our brains work and then sort of replicate the same thing in technology. So we can give away or delegate the things that we're doing which don't have to be done by us, anything boring or repetitive, and allow us to do more creative stuff. There are still big gaps, which is any reason-based work. Models don't really work very well on that yet. They're not able to do inductive reasoning very well. I think they're good at deductive reasoning by now, which is sort of questionable because there's not enough benchmarks on those ones, but we’ll eventually get there.

Any resources you can recommend? 

Abi: So I stay super active on Twitter, even if I don’t post. I spend about two hours on the platform every single day just collecting everything which is going on in the field. Along the lines of learning, there's a really nice podcast by my friend Alex Volkov called “ThursdAI Podcast". He is an AI evangelist with Weights and Biases, so right now I think he hosts one of the best podcasts in the field. It's about two hours long, happening every Thursday. It has its own Substack with the transcriptions as well so even if you miss out on it, you can still catch up later on.