Knowledge Exchange and Embed Partnership with Metis Labs
Metis Labs is a London start-up specialising in the application of Artificial Intelligence to optimise manufacturing processes to improve productivity and avoid waste.
Metis Labs applied successfully to CAP-AI Knowledge Exchange and Embed Partnership (KEEP) in July last year, with a 10-month research project to explore different Bayesian approaches for time-series forecasting. Eight months into the project, together with their KEEP researcher, Dr Julio Vargas, Metis Labs have made several breakthroughs into time series data analysis. These have not only improved their product but also has allowed the start-up to develop a strong core Intellectual Property, which is a significant differentiator among other tech players in their industry.
During the project, which is still ongoing, they have investigated existing Machine Learning techniques and their limitations, before coming up with novel approaches to solving their challenges. As Bashir Beikzadeh, Metis Labs’ CTO eloquently put it: “When you apply machine learning to a problem, more often than not, it doesn’t work the first time”. Other interesting learnings include ways of marrying research with start-up life and the importance of having the right team to deliver production-ready software. You can read the entire interview below.
Setting the Scene
The manufacturing industry in the UK is experiencing a radical transformation triggered by the so-called Fourth Industrial Revolution (Industry 4.0 or 4IR). 4IR is the fourth major era since the Industrial Revolution of the 18th century. That period of the steam engine, iron and textile industry was followed by the Second Industrial Revolution of 1870–1914, which brought the growth of the pre-existing industries and the creation of new ones such as oil, electricity and steel. The third era was, of course, the Digital Revolution, which began in the 50s when electronics started replacing mechanical and analogue technology. The fourth industrial revolution is bringing about the integration of physical production with digital technologies such as Artificial Intelligence to boost levels of productivity.
4IR technologies can create value across the entire manufacturing industry. The main areas where they have the biggest potential to deliver new business productivity are
- Smarter supply chains: greater coordination and real-time flow of information across supply chains, for example, better asset tracking, inventory management, fleet routing etc.
- Smarter products: the use of technologies like 3D printing to speed-up time-to-market; collecting data from products to allow for remote and predictive maintenance as well as analysing social sentiment on user feedback data for better marketing targeting
- Smarter production: the use of data science, machine learning and technologies like autonomous robots, multi-purpose production lines, augmented reality to allow for mass customisation, improved yield and speed up production
This is the set-up where Metis Labs, a software company based in London, operates. Metis Labs specialises in the application of Artificial Intelligence to optimise manufacturing processes to improve productivity and avoid waste. They take sensor data from systems like Operational Historians and use Machine Learning algorithms to learn the relationship between the data and the output of these processes, in order to understand how system variables and settings are impacting performance and which to optimise in real-time.
The company was started inside the start-up incubator Entrepreneur First in 2017 as part of EF8 cohort, by Alex Appelbe (CEO) and Bashir Beikzadeh (CTO). They successfully raised €1M Seed round from Speedinvest, AI Seed and Angel investors with strong experience in this space.
Metis Labs has successfully applied for Capital Enterprise CAP-AI’s Knowledge Exchange and Embed Partnership (KEEP) in July last year with a 10-month research project to explore different Bayesian approaches for time-series forecasting.
Dr Julio Vargas, the machine learning researcher, has deep expertise in signal processing and time series analysis. He received his PhD in signal processing from the University of Edinburgh. From the same institution, he received his MBA in 2010. He has published research on novel signal processing algorithms in the IEEE journal. Previously, he was a professor for 16 years at the University of the Andes, in Venezuela. He has also previously co-founded his own deep tech start-up on speech recognition.
Eight months into the project, we are catching up with the Chief Technology Officer, Bashir Beikzadeh at their office inside the Accelerator London. We wanted to understand why they applied for the KEEP in the first place and how has the project gone so far.
Hello, Bashir, thank you so much for taking the time to speak today. Tell me a bit about yourself. What did you do before Metis Labs?
Since childhood, I’ve had a very deep interest in maths and science, and I wanted to use my skills to solve the world’s challenging problems. I enjoyed doing a lot of mathematical challenges and was a national maths olympiad champion. My academic background is in Electrical and Software Engineering and I have previously built another Deep Learning start-up. It was a medical Deep Learning start-up where we analysed medical images to identify different things in them. One of the coolest things we did was to train a Neural Net to detect malaria cells with 98% accuracy.
That sounds cool indeed. What as the beginning of Metis Labs, how has it all started?
Alex Appelbe, my co-founder and I met through Entrepreneur First. I knew that, based on my Machine Learning background, there is a big opportunity for applying this technology to real-world industrial applications. Alex has a decade of experience in bringing innovation to manufacturing, and it seemed that it would be a good match in terms of having excellent domain knowledge of this sector. It was also a meaningful venture in that we were working on a business that would help the environment by reducing waste and energy.
And when did you start?
We started in May 2017.
Tell me a bit about how you found out about Capital Enterprise.
I found out about Capital Enterprise through attending one of the talks by John Spindler at another incubator we were at, called IoT Tribe. That’s when I had a chat with John and he encouraged us to apply for investment from Capital Enterprise.
What has attracted you specifically to the CAP-AI Knowledge Exchange and Embed Partnership (KEEP) and why have you applied for it?
We had just raised a Seed round and we were looking to grow our team. When we found out about the KEEP programme, it was just the perfect match for us, because it was a grant to help support research at Metis Labs.
Was there any other kind of support from Capital Enterprise that was useful for you aside from the KEEP itself?
Yes. John has been very helpful to us. He’s given us good advice on hiring, fundraising and introduced us to potential customers.
Can you describe a little bit the project you applied with, what were the objectives for the company?
The project we applied was to investigate different Machine Learning techniques for forecasting process outcomes in factories. We wanted to improve what we had been working on. In Machine Learning there are a variety of techniques you can use and you need to not only keep up to date with a lot of the latest research in this area, but also develop your own novel algorithms to have a competitive edge as a startup.
What were the Machine Learning capabilities you had in your team before the start of the KEEP?
We had good capabilities and had built POC’s on customer datasets.
And was that affecting you? Was that an issue in terms of delivering your product, for example?
Yes, for sure.
Where there any big unknowns and risks at the start of the project, in terms of both tech and business? For example, some estimates say that around 30% of Deep Tech research projects fail to deliver for one reason or another. What is your view on that?
Yeah, that’s absolutely true and I can vouch for that. In my experience, when you apply machine learning to a problem, more often than not, it doesn’t work the first time. There is a wide range of machine learning techniques out there and it takes research to evaluate them for the application we have.
Do you think some of these things can be foreseen at the start of an AI/ML project or do you have to go through the process to discover these limitations?
Yes, to run a deep tech startup you will need to do research to evaluate the best techniques for your application. Machine learning is a fast-moving field with regards to research and so this is a continuous task for the researcher.
What would you say that this means for a young startup, which has to go through this learning process for months and then discover that they might have to change the approach and the algorithms? What could be the impact of something like that?
I think that it can slow them down, and it can even lead them to nowhere.
I think you’ve started to touch a bit on that, but can you summarise how your KEEP, Julio, helped you answer some of those unknown questions?
Julio knows the challenge with time series data analysis, including non-stationarity and seasonality trends. His background in signal processing and time series analysis means he has deep insights into the problem.
And so, how has that helped the company?
Our deep expertise in this area helps us stand out from other competitors and provides our product Kelvin with unique functionality. [If you would like to know more here is a short video explaining the features of Kelvin.]
Your KEEP project is still going. Throughout the project so far, has there been any major changes or breakthroughs that either Julio contributed or you worked together as a team?
The breakthroughs have been in the development of novel machine learning techniques and in implementing them in a robust way in production.
What about in terms of the product and the business side, the part where you deploy these improvements in production and have your clients try them?
The challenge is that factory dynamics are non-stationary — meaning that their statistical properties change over time. To solve this our product Kelvin uses a novel technique using Gaussian Processes that perform Online Learning to continuously learn and forecast process outcomes in industrial systems in real time. We have several paid pilots of our product with large multinationals happening this year.
What about in terms of research collaborations? Has Julio kept in touch with his academic side, have you published any whitepapers for example?
We have written some whitepapers and would like to publish our research at some stage.
Now, let’s talk a little bit about your experience working with a researcher in a start-up environment. Are there any challenges with marrying rigorous academic research with a fast-paced start-up world?
There is definitely a challenge in productionising and testing cutting edge research. I think this is because this area is relatively new. For example, with software development, there are decades of experience on how to build software and methodologies like Agile. But with the ML research and productionising it, there are no widely accepted best practices. Over the past few years, we have learnt a lot about what works and what doesn’t work.
Are there any trends that are appearing in Machine Learning/AI around ways of working, almost like the Agile of AI companies? Have you found anything online to help you answer some of these questions?
I haven’t found much online and this is not common knowledge. In fact, I might actually write a blog post about this to share our experiences.
What are some of these big challenges?
The biggest challenge is communication and collaboration between academic and startup work environments.
So not a Machine Learning thing at all, then?
Well, it is also a technical challenge. A Machine Learning engineer would have to know the research really well to be able to implement it. You cannot simply implement some research if you don’t understand what it’s doing. So you need really talented Machine Learning Engineers. And you also should have people with specific knowledge in your particular area. Machine Learning is a diverse set of techniques. For example, NLP (Natural Language Processing), Image Processing and Time-Series Analysis are all somewhat distinct and rich areas.
What does that mean? Is it a matter of sitting together while solving a particular challenge and testing it or is it more than that?
Testing in Machine Learning is also a new and challenging topic. In software, you have Unit Testing and Integration Testing but with Machine Learning you have a test and validation dataset. Also, Sometimes training can take a long time, so this can slow down testing.
And even more, assuming you figure this out and your algorithms pass the validation and testing phase, then go on production and it turns out the ROI you predicted is not as expected. Is that a risk or a challenge for start-ups in this space?
Yes, there is naturally is a risk with models that need to work on a variety of environments. We tackle this by testing on a diverse set of datasets, some of which have been exclusively given to us by customers.
How has the company grown since the beginning of the KEEP project? Any new hires or funding rounds?
We’ve had a brilliant full-stack developer join us. We have moved to new offices in Shoreditch. In terms of new funding, we have recently won a significant grant funding award from Innovate UK.
Overall, how useful do you think the KEEP programme has been so far? And in terms of shortcomings or improvements that you can think of, to better deliver this programme in the future, what would some of those be?
I think that KEEP has been tremendously helpful for us in supporting cutting edge research. In a way, this program also helps to prevent the monopoly of talent and research in large tech giants, which can definitely help create a more even playing field for startups.
The support from Capital Enterprise has also been great. and this programme has been fantastic for us. I would really like to see more of this, an extension of KEEP or similar type of programmes.
If there was a message to transmit to other start-ups thinking to apply for KEEP, what would that be?
I would say ‘Go for it’!
Thank you, Bashir, and we wish you all the best with your company!
If you’d like to learn more about what Bashir and the Metis Labs team are up to, you can find them at www.metislabs.tech.
This project was part-funded by the European Regional Development Fund (ERDF).
