As an Applied Linguist, Suzana Ilic was introduced to machine learning through her specialization in data and text analysis. Through working on projects in sentiment analysis and emotion recognition, she began writing code and now works on deep learning projects for natural language processing. Currently based in Tokyo, Japan, her work has seen her collaborate with companies like Google and research organizations such as RIKEN.
Ilic is also the founder of Machine Learning Tokyo, a non-profit organization that has brought together a passionate community of engineers and researchers to work on projects, study new technology, and gather for regular events with industry experts.
In this interview, we ask Ilic about ML trends in Japan, challenges facing engineers, how MLT grew into a community of +4800 members, and what’s in store for the future.
How do you see the current state of AI and machine learning in Tokyo and Japan?
In the field of NLP we’ve seen major breakthroughs with pre-trained Language Models such as BERT or GPT-2 that can be quickly fine-tuned to adapt to a variety of language tasks. At Machine Learning Tokyo (MLT) we’ve also seen more and more interest in Edge AI and ML for IoT applications, a field where Japan excels.
We also see the push for ML use cases in the real world – from business applications to tackling social issues. Production-level ML comes with technical and non-technical challenges, from infrastructure and production pipelines at scale to realistically assessing and measuring impact. I recently read a report by BCG that stated that many AI initiatives fail and 7 out of 10 companies they have surveyed reported minimal or no impact from AI. And yet AI is still seen as a key competitive advantage in an increasingly technical market.
Have you noticed any interesting trends in machine learning, whether in Tokyo or the world?
In Japan I’ve seen a lot of research labs that are at the intersection of science and industry. We also see university-incubated AI Startups, many of which arise from the University of Tokyo. Some are even supported by university venture capital funds such as UTEC (University of Tokyo Edge Capital). I’m currently a visiting scientist at the RIKEN CSB-Toyota Collaboration Center, where Prof. Rei Akaishi works on Decision Science and Data Science with industry partners. That’s one trend I’ve noticed.
Another trend is, as earlier mentioned, the push for computing at the edge. In Tokyo there are companies like Preferred Networks or LeapMind with impressive projects and results in that field, but also many new players like EdgeCortix, working on distributed edge infrastructure and hardware design optimization.
How did Machine Learning Tokyo start?
It started very small in July, 2017. We were just two people doing what we needed for ourselves. My co-founder Yoovraj Shinde and I come from different backgrounds, but both of us realized that machine learning was something that could advance our fields.
We started meeting twice a week to write machine learning code at a coworking space. Each week by word of mouth more people joined, and at some point we weren’t able to fit all the people in that one space. So we decided to start doing meetups, and ever since it’s been growing pretty fast.
Over the past two and a half years we’ve grown to around 4,800 members in Tokyo. Our activities are mostly hands on, focused on writing code and building things. We organize a lot of workshops and study sessions where we go through algorithms, train models, and support engineers and researchers in the field.
So the key role is to support engineers and researchers?
In May of 2019 we established MLT as a non-profit organization because the community was growing fast and we started working on collaborations and partnerships with companies and universities. We now have a core team of 15 members and all our main efforts go into open education, open source, and open research. We host workshops with instructors who teach how to think about and implement new algorithms. We also work on open source projects that are reproducible and reusable, as well as on independent research, with two workshop contributions at NeurIPS in Vancouver last year.
How often do you organize talks and events?
MLT hosts regular workshops and study sessions 3-5 times per month, for instance monthly sessions for recommendation systems and or workshops at our Edge AI lab. For that we provide different hardware; edge devices and microcontrollers, from the Jetson Nano to the Tinker Edge T board and the Raspberry Pi 4. In these sessions we have 5-6 hours to build a working POC. Our members came up with some interesting demos so far and it’s a lot of fun to experiment.
We also hold remote sessions for people all over the world, hosted by people from different countries in different timezones. These sessions are dedicated to machine learning math, with more than 1,000 sign ups.
What are your future goals for MLT?
Last year showed it was a good decision to focus on building open source projects and research. I think we’ll continue putting more effort into projects, and increase the amount of focused sessions we do.
Another goal is to expand beyond Japan. We’ve gotten such a great response to our work from outside of Japan, so we’re thinking of ways to support an international community.
With all your experience interacting with engineers and researchers, what challenges do you see for people wanting to get into machine learning?
There are general challenges in tech because things move very fast. Research papers, tools, and libraries are being published and released every day. Sometimes it’s hard to keep up with the state of the art and everything that’s happening around ML. There’s also a lot of pressure in ML research, publishing at conferences has become extremely competitive and has been criticized for several reasons.
If you’re new to ML, I would advise you find out what you want to do as early as possible. It has never been easier to get into the field, with excellent resources and course material that is provided by top Universities like Stanford, MIT, or Berkeley for free. From theory and math to top down approaches like fast.ai, we have the tutorials, tools, and libraries that allow us to prototype, fail, and learn fast. But it can also be overwhelming. By knowing what you’re specializing in, curation will be a lot easier.
Many mid-career professionals ask me how to move to ML. My answer is usually to leverage your domain expertise and professional experience, and dive into ML top-down and hands-on. This will be your competitive advantage.
Finally, what’s the best way to learn more about MLT and take part in events?
At MLT, all events, resources, and project opportunities are provided for free to the community. If members want to support us, we’re on Patreon where supporters get early access to events or invitations to special opportunities. There are also many ways to contribute to MLT – you can find out more on GitHub.