Some time ago I had a chance to interview a great artificial intelligence researcher and Chief AI Scientist in Lindera, Arash Azhand.
We talked about:
- The AI technology behind his work at Lindera
- His career path
- How it is to be a research-centered scientist
- How to become a good leader
- Why it is important to approach AI research from a business perspective
Here is the full video of our conversation:
You can watch the video to see the raw content.
Here, on the other hand, is a summary of our conversation. What follows is not a 1-1 transcript but rather a cleaned-up, structured, and rephrased version of it.
What do you do at Lindera, and what do you build with your team?
What we do at Lindera, from the point of artificial intelligence, is to use these very interesting and novel techniques that are called AI technology in medical technology. Our main focus is to bring computer vision algorithms to detect human skeletons for the purpose of walking analysis. So we take videos of senior people and extract the skeleton out of these 2D video images in three-dimensional space and then try to calculate parameters of walking. For example, step length, step time, walking speed. That’s because we know, from a medical standpoint, that these parameters are specifically important to assess medical disorders such as the prevalence of falling of the elders.
Can you tell us a bit more about how does it work?
Sure. The core technology is actually quite universal. It’s not specifically for any disease or a group of diseases. It’s just taking the video and extracting coordinates of joints. Imagine you have a human skeleton, you have these separate joints like knee, ankle, shoulder, etc., and you want to have the coordinates of these joints over time. For example, when you have the joints of the foot over time, when people are walking you can then use these coordinates to calculate other parameters. So from there, it is just universal.
You can use this technology, this kind of investigation for elder care, you can use it for sports, you can use it for rehabilitation. There is no limitation, it’s quite universal.
Its core has this skeleton tracker, and on top of that, we can just build a tool that takes these coordinates and calculates some other stuff and gives recommendations to people. For example, when we find out that something is wrong in the walking of the person, then we could just calculate some parameters and recommend specific physiotherapy. So the extraction of parameters and recommendations tools are built on top of this core technology.
How big is your team, and how are you organized?
The whole team in Lindera is at the moment around 30 people. It contains a few separate teams.
- We have a sales team together with our CEO and CFO that are trying to push forward our product technology to customers.
- We have a customer success team. The key members of the customer success team go to the customers, present products at the customer’s place, provide help and service, and try to get their feedback in order for us to improve our product.
- Then we have an IT team that develops the frontend and backend system and also the app. And then at the core is data science, which is actually my responsibility area.
- My team at the moment, together with me, is six persons, and we develop these core AI technologies. We develop the algorithms for computer vision algorithms, the recommendations, and also for the gait parameter calculation.
Our main focus in data science is both research and development.
When I compare it to other industry companies and to my time at the university, I would say that data science at Lindera is much more research-centered than in other companies. That means that we usually don’t have the technology or what is behind the technology right from the start. Before that, we have to find out what is out there, we have to read the publication to research the forefront of research and try to extract the knowledge important for us. And then decide which part of this is usable for us, and how we could develop it in our systems.
What was your path to this moment? How did you become a data scientist and eventually, a chief researcher?
My background is in physics. I would say my interest in physics started from my curiosity to understand nature, to understand what are the hidden causes of natural phenomena. And just before I started to study, I was very interested in universe and cosmology. I read a lot of books back at school about cosmology and wanted to go to astrophysics. This was my early, let’s say, love. And eventually, I started to study physics. This was a very good step because we had a very nice astrophysics department and also a very nice professor. We just did the studies from a very philosophical standpoint. And this is also what was very interesting to me.
Eventually during my time at the university I got more interested in biophysics and the self-organization in complex biological systems.
And I was very curious to understand how for example, these very complex systems in biology, like we, humans, or even smaller beings, like insects are able to manage this chaos around.
How they are able to live in this chaotic system, get energy, and transform it into living. This self-organization principle and the biological system was very interesting. So, I actually went into this area and did my PhD in the field of statistical physics. I focused on complex systems and tried to understand pattern formations in these complex systems. You will find complex systems everywhere – some universal patterns, spiral waves. Spiral shaped waves can be found, for example, in the human brain, human heart, but also insect colonies, that try to organize insects through waves. And these waves are mostly spiral-shaped. And during my PhD, I studied the formation of these spiral-shaped waves in chemical systems to try to understand it from there and to understand it specifically for other systems.
And how did you go from this to working at Lindera?
I did my PhD and afterward a post-doc for two years. And at the postdoc, I did some very interesting research on the formation of crystals, which are not organic (so not living systems) but still are resembling those of living systems. They have curvature but are still inorganic. I did some experiments and simulations on these crystals. In the meantime, I just thought about what will I do in the future. Will I go to academia? Or will I step out of academia and go to industry? And my decision was actually to go to the industry because I saw more opportunities to do research. I got very interested in artificial intelligence topics.
One of the systems which interested me, maybe the most, was the human brain.
So I was very curious to understand how consciousness in the human brain emerges. How are we able to just infer knowledge out of the world with this organ. On the other hand, during the last years, I became interested in artificial intelligence, which in some sense is influenced by neuroscience. I guess my major motivation for work in the industry is to understand the link between original neuroscience and AI.
Did you find the link or only the inspiration?
At first, I found the inspiration, but more and more of the latest research, also in the area of artificial general intelligence, is trying to find a bridge between these two games. So it started with inspiration, but now more and more people see artificial neural networks are inspired by the specialized kind of algorithms to detect patterns and data. But it doesn’t tell us the causal relationships, links between these different phenomena.
So more and more people also try to combine these artificial neural networks with other AI research fields.
One example of another AI research area could be probabilistic models. There is a separate area of the Bayesian brain hypothesis in neuroscience, which builds on that the principle the brain, in some sense, is a kind of great Bayesian machine, which is getting the input through the senses from the environment and tries to model this input in the internal models. So that we just try to match what we see or hear.
What did you work on when you left university?
When I got out of university, I first started as a data scientist for another company. And my daily work was just to do predictive modeling, analytic modeling in the area of electronic sports. eSports teams have championships, and the products of the company called Dojo Madness, where I worked as a data scientist, was to develop models to predict e-sport game outcomes. After some time, I just stepped out of there and started to work here at Lindera. That was by November 2018. I left the university by April 2018. I started then immediately at Dojo Madness and in November at Lindera.
I started at Lindera because the data science work here at Lindera is much more reminiscent of my work at the university.
It’s more of reading scientific publications in this area to extract the knowledge effectively and to understand them. Just to see immediately in which direction we could go with this scientific knowledge that is out there and how to use that for us.
How does your average day look like?
One part is for sure, reading new research. So I developed a kind of system where I gather research papers and set a plan to read them. Then I extract the knowledge effectively out of that and write down kind of documentation of what I found out.
Is this sort of team-wide strategy?
We’re doing the whole development in kind of Kanban style. We have the tickets for the classical development, and we started to do that in some sense also for the research part. Members of my team, who are also research-focused, try to read and test everything that is in some specific paper and make documentation. We write that directly in these Kanban tickets. This way, we can see what the state of research is, what some of us have already found out, in which areas we should not go further.
That’s our idea at the moment, and we’re iteratively developing this method. We have to review it every week, see if it functions and when it does function in some ways how we could go after this technology.
What are other significant parts of your day-to-day activity?
Another major part is the development, of course. When we decide to go further in one specific area, we found out this paper is good and promising, and we should develop that. We check if there is code out there, and someone from the team tests this code tries to develop it, research these findings, and document that. And then this serves as a kind of first proof of kind of tool or algorithm. We can later decide if we put it in our production system. This is the applied development area, and I’m also in that. So I mostly do research but also try to develop specific areas and code.
For chief data science positions, a huge part of the work is communicating with the business, and providing people with infrastructure, access to the data or the legal information. Is it your case as well?
Yes. I worked as a regular data scientist at Lindera for one year, and after that, I had a sort of transition time (2-3 months) to the chief AI. During that time, I gradually stepped back from this research and development part of work. I slowly gave away my tasks in these areas to my team members and took on the overall strategic development and communication. So I’m a kind of link between the data science team, the business, and customers.
I’m a kind of link between the data science team, the business, and customers.
Was there something unexpected for you in transitioning from the role of a contributor data scientist to a managing data scientist?
The interesting thing is that it was not unexpected to me because I’m coming from the physics area where most people do PhD in some research teams. You naturally, at some time, do some management there. You have to advise and manage, for example, master students and other students that also try to develop research ideas. From that standpoint, it was not new for me. And what also has helped me was the year before at Lindera when I worked as a data scientist and received a lot of advice from the former chief and top leadership.
Before I took over this chief position, I felt some kind of anxiety, I was wondering how would it go on and if I would be able to do it. But from the moment I stepped in, there was a kind of switch inside me and my brain, and I knew I could do it.
I just had this feeling that it perfectly fits me and my character to lead this team, advise them, listen to their ideas, and to understand what their strengths are.
To understand how I could take their strengths and use that for the data science and Lindera.
How do you help people improve in some areas? How do you encourage people to be the best of themselves?
There are two areas – soft skills and hard skills. The interesting thing is that I had the impression that the more needed skills are soft skills. Because the hard skills, like the research ability and also the development and programming skills, are kind of craftsmanship. What you learn during the time at the university, and after that, it’s just craftsmanship that you have to do for a while. Then you can do it. For a leading position, soft skills are the major part. For example, the ability to listen closely.
You have to listen to the team members and see how they convey the idea, how they communicate these ideas on a day to day basis, and how they feel when they communicate this. Make them feel appreciated. Give them the room and space to develop ideas and bring them to you. You catch the ideas and then try to integrate these ideas into the overall strategy of your team and also the company. So you have specific key goals of the company. You decide in which area you develop.
Give your team the room and space to develop ideas and bring them to you.
Also, you have to notice that this team member is very research-centered. He or she is very great at research and likes to read papers and extract knowledge. A good leader is just behind you, and also pushes his knowledge to the team.
Are there other skills that you think are crucial to the leadership role in this very researchy data science world?
As a regular data scientist, you do your tasks on a daily or weekly basis, you finish that, and then you document your solution, and everything is fine. But from the leadership position, you have to bring in also some kind of ability of foresight. There are demands from customers and businesses. So you have a timeframe to deliver things to the customer. You have to be able to understand what is the strength of your team, what can they do, if this time window to deliver is doable. This is maybe one of the most difficult things, and I cannot say for sure that I’m on it. So, time will tell if I did it correctly.
We call it here kind of decision intelligence, which is a new discipline in the data science area (some call it machine learning++ or data science++), which is just looking at all the stuff from the other perspective.
Instead of putting data and AI in the first place, it is about starting from questions like “What is the problem and which decisions have to be made from business and customer side?”
What were the things that you think now that have been crucial for you to learn to sort of progress in the field?
I think a major thing is the communication part. When you’re at university, you’re mostly responsible for your work. You can do some research the whole year, and then write about it, and then you have no other responsibilities. However, in most companies you have to, on a daily basis, learn to communicate better solutions and ideas.
When I started as a data scientist, as a perfectionist, I did the research for a week to solve one specific problem, because I had the impression that I have to do that. But in the end, it was not a major problem. I lost a lot of time for micro research and micro improvement while the major stuff was not that. I just had to force myself to prioritize work.
How did you approach that?
Books were helpful – I’ve read some books about the Agile programming area, some about how to prioritize tasks. But I think what helped me the most was talking to other people who had this experience. I had to lose this anxiety of thinking that I have to be perfect and start asking for help. It sounds very trivial, but it’s maybe one of the most difficult things. Remembering I’m not obliged to solve all problems by myself.
And also, from the leadership perspective, it’s very important to give your team the impression that they can always come to you or other team members and ask for help. A recipe for disaster is when there is a kind of competitive environment in the team. So you have to show your team that they can all work together.
Do you have any methodology or system in place that facilitates discussion and exchange of help in your team?
At least on a monthly basis, but even more often, I try to talk to each team member, one on one, and ask them how is their working life, is everything fine, do they have a problem. I take a step back and let them talk. Maybe also tell them, if they have specific problems holding them back, which team member is good at that. So that they work together.
Another thing is the knowledge transfer, both from the data science and also Lindera perspective.
What are three tips that you would give data scientists, junior or regular, on how to progress in the careers and have a better experience doing it?
- First, analyze the company as a whole, but also the data science. Look very carefully into the team structure. What kind of data science leader is there? Do they give you space for improvement? Are there people who are diverse in their expertise? How do they treat you and other team members?
- The next major tip I would give to more research-centered data scientists, like me from mathematics or physics area, to do intensive courses on hard skills, like specific programming courses. Being fluent in the craftsmanship – architecture design and also libraries helps a lot. It’s like speaking a language that allows you to communicate much better with people. When you’re very fluent in the craftsmanship, your brain will have the freedom to create. I learned, for example, that when I didn’t have that craftsmanship, I lost a lot of time and ability for creativity.
- The next major advice is also to look at the company and its learning environment. Do they provide you the space to learn and grow, or is it just what I call it, more of data engineering, just pushing data.
- Lastly, another piece of advice related to the first one is to analyze the team structure. When there are data engineers, maybe also some DevOps in a team, it gives data scientists huge freedom.
Are there any books, resources, blogs, or podcasts that you would recommend?
One thing which I can recommend for sure, probably most of the data scientists who are actually practitioners know that already, is the Medium blog and Towards Data Science blog (which is part of Medium). I decided to pay for a Medium membership since last year, and I’m really into it. I try to read interesting articles about novelties.
Also books –for me personally most valuable are general books, not the specialized ones. General AI books that go more into AI philosophy, neuroscience.
It’s good to take a step back and read something else or do something else, walk around and hear some music, for example, and then come back, and suddenly the problem solves itself.
Is there any message that you want to leave our listeners with?
I think the major message that I can give to the listeners and viewers who wonder if they should go into the AI area is that now is the best time to be here. I have the impression that in the next one to five years, major improvements will come specifically in some kind of artificial general intelligence. The combination of our current AI with other areas of science will lead to major improvements.
Interesting articles and resources shared by Arash Azhand
Here some stuff about Zero-Shot Learning (ZSL):
- Intro article on ZSL and its applications: https://towardsdatascience.com/applications-of-zero-shot-learning-f65bb232963f “Zero-shot learning refers to a specific use case of machine learning (and therefore deep learning) where you want the model to classify data based on very few or even no labeled example, which means classifying on the fly.”
- Another article on one latest research contribution: https://medium.com/@alitech_2017/from-zero-to-hero-shaking-up-the-field-of-zero-shot-learning-c43208f71332 “zero-shot learning refers to the process by which a machine learns how to recognize objects in an image without any labeled training data to help in the classification.”
- Max Planck Research group working on ZSL with some of the greatest work on the field: https://www.mpi-inf.mpg.de/departments/computer-vision-and-machine-learning/research/zero-shot-learning/zero-shot-learning-the-good-the-bad-and-the-ugly/
- … and one of their best papers:
Introduction to the Field of Decision Intelligence (Data Science plus plus):
- Intro article by Cassie Kozyrkov (Chief Decision Intelligence Scientist at Google): https://towardsdatascience.com/introduction-to-decision-intelligence-5d147ddab767 “Curious to know what the psychology of avoiding lions on the savannah has in common with responsible AI leadership and the challenges of designing data warehouses? Welcome to decision intelligence!”
- Talk by Cassie Kozyrkov on the topic: https://www.youtube.com/watch?v=bCjMhZZYlP4
- Blog article by Cassie Kozyrkov, “What on Earth is Data Science?” https://hackernoon.com/what-on-earth-is-data-science-eb1237d8cb37
A new Theory to understand Deep Learning in neuronal Networks better, based upon the classical Information Theory of Shannon, and called the Information Bottleneck” Theory of Deep Learning:
- Article in Quanta Magazine on it: https://www.quantamagazine.org/new-theory-cracks-open-the-black-box-of-deep-learning-20170921/ “A new idea called the “information bottleneck” is helping to explain the puzzling success of today’s artificial-intelligence algorithms — and might also explain how human brains learn.”
- My own blog articleon the “Information Bottleneck” Theory, written recently as the first article of our own Lindera AI Blog: “A View into the Black-Box – An Information-Theoretic Approach to Understand Deep Learning” https://www.lindera.de/2020/03/05/a-view-into-the-black-box-an-information-theoretic-approach-to-understand-deep-learning/
A series of very interesting medium articles by Manuel Brenner from neuroscience perspective:
- Blog article “How We Can Learn from The Brain to Learn How the Brain Learns”: https://towardsdatascience.com/how-we-can-learn-from-the-brain-to-learn-how-the-brain-learns-2b81286a1a7b
- Blog article “Why Intelligence might be simpler than we think”: https://towardsdatascience.com/why-intelligence-might-be-simpler-than-we-think-1d3d7feb5d34
- Blog article “The Thermodynamics of Free Will”: https://medium.com/@manuel_brenner/the-thermodynamics-of-free-will-940cacd02401
- Blog article “The Bayesian Brain Hypothesis”: https://towardsdatascience.com/the-bayesian-brain-hypothesis-35b98847d331
- Blog article “The Geometry of Thought”: https://towardsdatascience.com/the-geometry-of-thought-700047775956
Last but not least, work by Prof. Karl Friston
Maybe one of the most renowned neuroscientists these times, who developed the Free Energy Principle as driving force for human intelligence and thought. Some of his work also was introduced in the above articles by Manuel Brenner. But here also one article by himself on Aeon: “The mathematics of mind-time – Consciousness is not a thing but a process of inference” https://aeon.co/essays/consciousness-is-not-a-thing-but-a-process-of-inference.
ML Experiment Tracking: What It Is, Why It Matters, and How to Implement It
Jakub Czakon | Posted November 26, 2020
Let me share a story that I’ve heard too many times.
”… We were developing an ML model with my team, we ran a lot of experiments and got promising results…
…unfortunately, we couldn’t tell exactly what performed best because we forgot to save some model parameters and dataset versions…
…after a few weeks, we weren’t even sure what we have actually tried and we needed to re-run pretty much everything”
– unfortunate ML researcher.
And the truth is, when you develop ML models you will run a lot of experiments.
Those experiments may:
- use different models and model hyperparameters
- use different training or evaluation data,
- run different code (including this small change that you wanted to test quickly)
- run the same code in a different environment (not knowing which PyTorch or Tensorflow version was installed)
And as a result, they can produce completely different evaluation metrics.
Keeping track of all that information can very quickly become really hard. Especially if you want to organize and compare those experiments and feel confident that you know which setup produced the best result.
This is where ML experiment tracking comes in.Continue reading ->