The 63rd episode of Datacast is my conversation with Azin Asgarian — an applied research scientist on Georgian’s R&D team, where she works with companies to help adopt applied research techniques to overcome business challenges.

Our wide-ranging conversation touches on her foray into studying math and computer science in Iran, her academic research on facial detection analysis at the University of Toronto, the benefits of being a teaching assistant, her interesting projects with Georgian Partners, real-world applications of transfer learning, and much more.

Listen to the show on (1) Spotify, (2) Apple Podcasts, (3) Google Podcasts, (4) TuneIn, (5) RadioPublic, (6) Stitcher, and (7) iHeart Radio.

Key Takeaways

Here are highlights from my conversation with Azin:

On Going To An All-Girl High School in Iran

To be honest, it was one of the best periods of my life. It was very motivating and inspiring to study among the most talented students of my city. We had a healthy and organic competition that helped me a lot to push myself beyond my limit. There’s a huge emphasis on subjects like math, physics, and chemistry in my school, which helps me build a strong foundation. Besides that, there’s also a strong emphasis on extracurricular activities like programming, sports, painting, etc. When I was 14, I participated in a robotics class, which made me passionate about programming in general.

As you know, everything comes with advantages and disadvantages. One disadvantage is that being in a girls-only school, I only interacted with people of the same sex. It’s hard to get to know men and understand the differences between men and women before my Bachelor’s. On the positive side, I was constantly surrounded by independent, successful, and talented women — which made me want to become one of them in the future.

By adopting successful women as my role models, I started to shift my mindset and understood that hard work could lead to great outcomes.

On Undergraduate Study at the University of Tehran

My undergraduate experience is tough, challenging, yet rewarding at the same time. Since childhood, I have always been interested in mathematics and programming. After high school and the university entrance exam, I had to choose a program. The two options that I had were computer science and computer engineering.

There’s a tendency for students to go into engineering programs in Iran because they are generally more prestigious and popular than arts & science programs. However, I decided to study computer science because it is a part of the Math department in Iran and more theoretical. To get some experience in programming, I participated in extra classes in a small team that competes in ACM and ICCP programming contests. Through those activities, I tried to improve my programming skills while studying the theory behind it.

Overall, the combination of the school environment, the professors, and the activities helped me find what I’m interested in.

On Graduate Study at the University of Toronto

My graduate experience is generally very positive. Moving from Iran to Canada was definitely a huge change for me. Everything is the polar opposite: the culture, the education, the economy, etc. When I moved to Canada, it was a bit of a shock at the beginning. Being in a new environment, I didn’t know how to study, prepare for the final exam, or do the homework assignments. I was constantly looking for feedback to understand how I was doing and change strategy if needed. So, the first few months were definitely a bit difficult in adapting to the new environment and understanding what success looks like. It went very smoothly because of the support from my advisors, the community of Iranian students, and other students at the University of Toronto.

After the first few months, it was pure learning, especially from my advisors — whom I am very grateful for. Furthermore, U of T is one of the best schools in the world. It has amazing resources for research, access to facilities like the gym and the library, free classes in English/writing/health and wellness. The researchers you saw around yourself are the best in their fields. Amazing talks and events were going on every day. Being a part of that environment was something very positive in my life.

On Being A Teaching Assistant

Being a TA is one of the best ways to deepen your knowledge in a given area. I can’t highlight that enough. When you try to explain something to someone, you have to think about all the possible questions that you have to know and have answers for. This is very different from reading something by yourself. Throughout my Master’s, I took a few courses in Machine Learning. After that, I really wanted to deepen my knowledge and transfer my knowledge to other undergraduate students. So, I took TA duties for one of those courses, and it definitely helped a lot.

In addition to that, I gained different perspectives while looking at the same problem. Going through students’ assignments and their thought processes, I obtained new and interesting perspectives that might be useful sometimes for my research and sometimes for my own thought process.

On Her MS Thesis

When starting my Master’s, I participated in a research project about developing a pain detection system for older adults with dementia to detect their pains based on their facial expressions. People with severe dementia usually can’t communicate their pains verbally, so their pains are missed most of the time when they live in nursing homes or long-term healthcare facilities. That leads to a high mortality rate because when the pains are missed, it’s a likely indicator that bigger problems are missed.

Initially, we had some data collected from older adults. We were thinking about developing our AI model to detect their facial expressions. But halfway through the project, we realized that our data was minimal, so we started to look for public datasets. However, there was not much data with older adults, especially those with dementia.

Most of the publicly available datasets contain normal faces of healthy individuals. Our targeted population was different from what we could find on the Internet. Healthy individuals have clear eye-browns, no wrinkles, and usually dark hair. But people whom we worked with have thin eye-browns, lots of wrinkles, double chins, and white hair. When our model was trained on publicly available datasets, they performed poorly on the images of our patients.

We realized that, intuitively, there is a huge difference between the appearances of healthy individuals and older adults with dementia (whom we wanted to deploy our models on). The question then became how to remove this gap and train our algorithm on a combination of these datasets, such that our algorithm learns the general features from the public dataset and the specific features of older adults from the tiny, limited data that we had.

The idea behind our paper is that:

If we have a source dataset and a target dataset, we can still benefit from training our model on both datasets if the algorithm can smartly pick the features from them.

We accomplished this via a combination of dimensionality reduction methods and active appearance models (which are generative models for capturing facial structure).

On Georgian Partners

Georgian is an innovator and venture capital space focused on investing in B2B SaaS companies with large datasets and interesting problems. We had different teams, including investment, R&D, operations, customer success, and more. As part of the post-investment process, we help our portfolio companies scale and growth as much as possible on different aspects (operationally, financially, and technologically).

What makes Georgian unique compared to other VC firms is our R&D team, which I am a part of. My team helps our portfolio companies solve their technical challenges. We do it in different ways:

1. We play an advisor role with them for building the foundation of their data science capabilities.

2. We run hackathons with them when there is a particular problem to solve.

3. We focus on a very research-y problem that can’t be solved in a week but maybe in 6 months or a year for deeper engagement.

4. We help them find the technology trends in the industry and stay ahead of those (via a monthly paper club).

On Transfer Learning

In this blog post, I laid down the foundation of transfer learning and explained it mathematically in a not-so-heavy way such that people can absorb it. I also talked about different types of problems that can be solved with transfer learning and different solutions that can be applied. In the upcoming posts, I plan to highlight real applications of transfer learning and a few use cases in more depth to show people the advantages and disadvantages of using these methods. Like any other algorithms, transfer learning has its own limitations, so it’s good to be aware of them (especially in the industry where you have limited time and resources).

Most of the classic work in transfer learning lies in homogeneous transfer learning, where we share knowledge between two domains with the same feature space (image-to-image, text-to-text, etc.). However, in many real-world cases, we have different modalities of data and need heterogeneous transfer learning to transfer information between two domains with different feature spaces. Recent research that develops robust algorithms focused on multi-modal environments can be valuable in this scenario. Most of this research uses deep learning and evaluates on unstructured data. However, we often have a lot of structured data in the real world. I think it would be good to work on a combination of structured and unstructured data.

The most interesting and wide application of transfer learning in industry is the “cold-start” problem.

A lot of growing and scaling companies already face it. When they try to onboard new customers and do not have enough data to train new models on those customers, that creates delays and increases on-boarding time significantly. This problem can be solved by using different transfer learning methods. In the upcoming posts, I plan to cover two use cases at Georgian. Through collaboration with our portfolio companies, we helped decrease their onboarding times and enhance the time-value-add to them significantly while using different transfer learning (and other ML) techniques.

On Data-Efficient Methods

I am passionate about methods that can make AI models data-efficient (transfer learning, active learning, representation learning, and related research areas). Thanks to the past few years working in the industry, I have realized the value that these methods can bring to our daily lives by reducing the need for labeled data and increasing the scalability of AI companies. I am eager to work and make more progress in those research areas. I believe there’s a huge untapped potential that can be converted into a real value if we focus on those methods.

Show Notes

• (02:06) Azin described her childhood growing up in Iran and going to a girls-only high school in Tehran designed specifically for extraordinary talents.

• (05:08) Azin went over her undergraduate experience studying Computer Science at the University of Tehran.

• (10:41) Azin shared her academic experience getting a Computer Science MS degree at the University of Toronto, supervised by Babak Taati and David Fleet.

• (14:07) Azin talked about her teaching assistant experience for a variety of CS courses at Toronto.

• (15:54) Azin briefly discussed her 2017 report titled “Barriers to Adoption of Information Technology in Healthcare,” which takes a system thinking perspective to identify barriers to the application of IT in healthcare and outline the solutions.

• (19:35) Azin unpacked her MS thesis called “Subspace Selection to Suppress Confounding Source Domain Information in AAM Transfer Learning,” which explores transfer learning in the context of facial analysis.

• (28:48) Azin discussed her work as a research assistant at the Toronto Rehabilitation Institute, working on a research project that addressed algorithmic biases in facial detection technology for older adults with dementia.

• (33:02) Azin has been an Applied Research Scientist at Georgian since 2018, a venture capital firm in Canada that focuses on investing in companies operating in the IT sectors.

• (38:20) Azin shared the details of her initial Georgian project to develop a robust and accurate injury prediction model using a hybrid instance-based transfer learning method.

• (42:12) Azin unpacked her Medium blog post discussing transfer learning in-depth (problems, approaches, and applications).

• (48:18) Azin explained how transfer learning could address the widespread “cold-start” problem in the industry.

• (49:50) Azin shared the challenges of working on a fintech platform with a team of engineers at Georgian on various areas such as supervised learning, explainability, and representation learning.

• (51:46) Azin went over her project with Tractable AI, a UK-based company that develops AI applications for accident and disaster recovery.

• (55:26) Azin shared her excitement for ML applications using data-efficient methods to enhance life quality.

• (57:46) Closing segment.

Azin’s Contact Info

• Website

• Twitter

• LinkedIn

• Google Scholar

• GitHub

Mentioned Content

Publications

• “Barriers to Adoption of Information Technology in Healthcare” (2017)

• “Subspace Selection to Suppress Confounding Source Domain Information in AAM TransferLearning” (2017)

• “A Hybrid Instance-based Transfer Learning Method” (2018)

• “Prediction of Workplace Injuries” (2019)

• “Algorithmic Bias in Clinical Populations — Evaluating and Improving Facial Analysis Technology in Older Adults with Dementia” (2019)

• “Limitations and Biases in Facial Landmark Detection” (2019)

Blog Post

• “An Introduction to Transfer Learning” (Dec 2018)

• “Overcoming The Cold-Start Problem: How We Make Intractable Tasks Tractable” (April 2021)

People

• Yoshua Bengio (Professor of Computer Science and Operations Research at University of Montreal)

• Geoffrey Hinton (Professor of Computer Science at University of Toronto)

• Louis-Philippe Morency (Associate Professor of Computer Science at Carnegie Mellon University)

Book

• “Machine Learning: A Probabilistic Approach” (by Kevin Murphy)

Note: Azin and her collaborator are going to give a talk at ODSC Europe 2021 in June about a Georgian’s project with a portfolio company, Tractable. They have written a short blog post about it too which you can find HERE.

About the show

Datacast features long-form conversations with practitioners and researchers in the data community to walk through their professional journey and unpack the lessons learned along the way. I invite guests coming from a wide range of career paths - from scientists and analysts to founders and investors — to analyze the case for using data in the real world and extract their mental models (“the WHY”) behind their pursuits. Hopefully, these conversations can serve as valuable tools for early-stage data professionals as they navigate their own careers in the exciting data universe.

Datacast is produced and edited by James Le. Get in touch with feedback or guest suggestions by emailing khanhle.1013@gmail.com.

Subscribe by searching for Datacast wherever you get podcasts or click one of the links below:

• Listen on Spotify

• Listen on Apple Podcasts

• Listen on Google Podcasts

If you're new, see the podcast homepage for the most recent episodes to listen to, or browse the full guest list.