The AI Fellowship: Mastering Machine Learning on the Job
How our AI Fellowship helps Machine Learning engineers realize their full potential and stay on top of the game
Every single day, a huge amount of AI research is being published. As a Machine Learning (ML) Engineer, it’s getting harder and harder to keep up. The Stanford University AI index report highlights that the number of publications in the field of AI has increased exponentially in recent years. You could argue we need AI to track AI. 😊
And keeping track of AI research is important.
Knowing what is currently state-of-the-art is crucial for delivering the best quality AI products and solving real user needs. Additionally, Machine Learning engineers get excited by trying new techniques and want to stay relevant in their field.
At Slimmer AI, an applied AI B2B venture studio, we are also continuously scouting for new opportunities where data and machine learning can create a competitive edge. Knowing what is — and what isn’t — possible with AI enables us to focus on opportunities in exciting new areas and stay away from others.
There are many platforms where AI research is collected and summarized, such as YouTube channels, AI newsletters, blogs, or following the right people on Twitter or LinkedIn. However, all of these supply a high-level overview. They only guide you to focus on certain research areas while disregarding others. For those of us that do applied machine learning, they provide neither deep theoretical knowledge nor hands-on experience. And they don’t provide answers to practical questions such as: Do the results on the academic dataset translate to acceptable results in practice? Can we implement it? What’s the runtime like? To answer these questions, we need to spend more effort.
Within Slimmer AI we’ve attempted to tackle this need for more AI R&D many times and learned what did and didn’t work — sometimes the hard way. After several iterations, we now have a truly successful formula: the AI Fellowship. And our ML engineers love it!
The AI Fellowship is our take on an R&D program for ML engineers. It was launched in January 2021 and is still going strong 18 months later. The program is continuously evolving and improving by incorporating feedback from our engineers. It is one of our biggest assets and unique selling points for talent development, and tech talents are often blown away when we mention its existence. Our ML engineers consistently rate the program as very or extremely valuable and indicate that the learnings have made an impact in their product development and advanced their technical skills. This approach has helped us consistently increase our eNPS (employee satisfaction) scores, develop new AI capabilities, and attract the right talent.
In this blog post, I will introduce you to the fundamentals of the AI Fellowship and explain why we believe it is crucial for our innovation and an essential part of our tech talent strategy.
In a separate article I outline our previous failed attempts towards open R&D and tips & tricks for implementing an AI Fellowship in your own company.
The AI Fellowship
The AI Fellowship centers around four main objectives:
1Accelerate Innovation We work on topics that are aligned with Slimmer AI’s near-future needs, such that we stay ahead of the game.
2 Attract, Develop, and Retain Talent We work on topics that are aligned with both gaps and interests in our employees’ expertise.
3 Thought Leadership We strengthen our thought leadership position by sharing and publishing our findings.
4 Strengthen Venture Ties As a venture studio, we care not only about the Slimmer AI team but the ML capabilities of our portfolio companies as well. We include venture employees in our Fellowship and help them set up their own programs.
So, how do we organize ourselves to achieve these objectives? Based on the identified patterns of our past failures, we centered the Fellowship around the following key principles:
- Scheduled: We have four R&D cycles per year with preselected topics
- Expected: Participation is mandatory and is part of your job description
- Prioritized: There is dedicated time per week to work on the R&D program
- Structured: There are appointed people that drive the innovation work in the AI Fellowship
R&D Cycles
Each R&D round or cycle takes three months to complete and has a clear start and end date. The cycles coincide with the four quarters of the year.
Before a cycle starts, a few topics are chosen to focus on. This typically comes down to two or three topics for about 15 ML engineers, but the number of topics depends on the team size. Topics are chosen to accommodate our focus areas (e.g., expand our expertise in NLP), but also to explore new ones (e.g., the latest research in graph neural networks). ML engineers are encouraged to write a mini proposal if they have an idea of their own.
Each cycle consists of (1) a research phase and (2) a development phase. The research phase takes around 7 weeks to complete, whereas the development phase takes 2 weeks. The time that is left is meant as a buffer (e.g., when someone was out sick or went on a holiday) or it is used to write a blog post when something interesting can be shared. The buffer is also useful as a breather, as continuous learning can be mentally taxing. At the end of each cycle there’s a retrospective where we identify possible improvements. To optimize continuous knowledge sharing during the cycles, we have a weekly AI stand-up where new learnings are discussed with the entire team of ML engineers.
A key end result of our AI Fellowship is the creation of technical content, typically in the form of blog posts. The blogs are usually technical and primarily intended for the benefit of fellow ML engineers. They’re a means of giving back to the community, not only showing thought leadership as a company, but also to get yourself (as an engineer) known to the world. If the research is really promising, there’s the possibility of continuing the work with the aim of publishing a paper or presenting at a conference. If you’re interested in some of our previous work, you can find our posts on Medium.
Another technical output of a cycle could be an open-source software package. Generating open-source software is extremely rewarding for engineers as having their code benefiting others is a big reason they create it in the first place. Slimmer AI’s open-source repositories can be found on our GitHub.
Mandatory Participation
Because the world of AI is moving so rapidly, the sad reality is that we ML engineers become outdated really quickly. This is something my team cannot afford. Therefore, it only makes sense to add ‘learning’ to the job description and make it part of yearly assessments. This means that if ML engineers haven’t actively participated in the program, this will be reflected in their reviews.
By making it part of the job description and reviews, everyone is reminded that management is not only on board but also expects that time is spent on projects that are not business critical at that time.
Dedicated Time
All product owners and team leads are aware of the AI Fellowship’s timeline and plan accordingly. This way, the ML engineers are confident they can spend their time on R&D.
During the research phase of an R&D cycle we set aside 4 hours per week. Engineers are free to choose how they plan this time. Some like to spend half a day per week, some rather spend one full day every two weeks. Some work alone, others work together.
During the development phase we block two weeks where people can spend 3 full days each week on development. Again, they are free to plan it slightly differently. Some rather spend 2 full days for 3 weeks and some rather switch between the research and development phases multiple times. As long as the end result is there, we encourage people to do what makes sense for their agenda.
Driving Innovation
In the ideal world, all engineers will always feel empowered and energetic to self-organize and behave proactively. In reality, given the failed attempts at R&D programs that we’ve experienced, this simply isn’t the case. So, the final piece of the puzzle involves having someone drive the AI Fellowship.
For this reason, we added a new role to our company and made the success of the Fellowship its primary target. After a few cycles, many of these responsibilities are now being shared between a few senior ML engineers that form the steering committee of the Fellowship. Together, we make sure that there are interesting topics to work on given the priorities set out by product and venture managers. We facilitate interesting discussions in our weekly stand-ups, and ensure that the goals we set are being met.
Mixing it up
People say “variety is the spice of life”, so to keep things interesting we mix it up now and then.
Internal AI Competition
After the first three successful R&D cycles, we decided to step away from our traditional approach and organize our very first internal AI competition. The competition was centered around the topics we tackled earlier in the year and can be compared to a Kaggle-like competition, with the ultimate goal of obtaining eternal glory and a place in our company’s Wall of Celebration.
People were divided into small teams and competed on the same task where they not only had to optimize for performance, but also for efficiency. Many team members didn’t work together before, which contributed to the team coming closer together. In the end it was a lot of fun and by mainly focusing on development this time, people gained a lot of valuable hands-on experience. We’re planning a similar event for this year.
If you want to learn more about the competition we organized, have a look at this blog post by Ayla Kangur, which contains all the competition’s rules, the lessons learned, and the 10 steps for running a successful competition yourself.
Venture Collaboration
As a venture studio we’re adding more and more ventures to our portfolio. To stay connected with the ML teams within those ventures, we plan to organize joint R&D cycles. The first edition will start this year.
In this collaborative cycle, we invite the AI/ML department from one of our portfolio ventures to our weekly AI stand-ups and jointly tackle a topic they’re interested in. This will not only help them with their business problem, but there will also be an opportunity for mutual learning. They usually have a different focus area, which means they can learn from us and we can learn from them. In what other company do you have the chance to collaborate and learn from so many different people?
We are very much looking forward to this collaboration, to learn from it, and to apply this format more frequently going forward.
In conclusion
AI is a fast moving field and as an ML engineer you want and need to stay up to date with the latest developments. Not only to stay ahead of the game, but also to stay relevant.
In this blog post I introduced the AI Fellowship: our R&D program for ML engineers. It was created based on the learnings of several failed attempts and can be implemented in any company that develops and applies AI. Our engineers love it and it has become one of our unique selling points for attracting and retaining talent.
If you want to learn more about our previous attempts, our learnings from them (i.e., how to not set it up), and our tips and tricks for implementing a successful AI Fellowship in your company, please have a look here.
And if you want to know more about our work at Slimmer AI, feel free to reach out or head to our website to learn more.
