What is Secondmind Labs?
Since our company’s inception it has been clear that building the Secondmind Decision Engine would mean charting new territories. This exploration requires state-of-the-art techniques and a team adept at asking the right questions. Having a formidable research team has always been at the core of our company’s strategy, and now Secondmind Labs brings this team to life.
As we have grown, so have our ambitions. We want our research to have an impact on the world around us, not just in academic circles. Secondmind Labs builds a bridge between our cutting-edge research and applied machine learning technology. It blends our founding principle of having a research team, free to explore and experiment, with our collective drive to see those discoveries applied to real-world challenges. The two work in harmony: customer challenges steering our research; research opening up new customer opportunities.
Four areas of expertise are required to derive a decision-making system from first principles: probabilistic modelling, optimisation, optimal control and reinforcement learning. Our multinational team consists of 19 machine learning researchers and engineers that are world leading experts in those critical areas.
Our mission is twofold.
- Provide insights and expertise for building the Decision Engine and supporting our customers. We ensure that our software platform and the people who rely on it to make decisions benefit from the latest advances in the field of machine learning (ML).
- Explore innovative ideas and develop ground-breaking ML libraries to share with the scientific community. We improve the performance of existing methods and solve problems that were previously out of reach.
Building a multipurpose Decision Engine would not be possible without the incredible progress made in the field of ML over the past few decades. This wealth of knowledge inspires us, and in return we do our best to contribute to it. Over the past four years, we’ve published more than 60 articles at leading ML conferences, and we’ve also put tremendous effort into developing code that is now the de facto framework for Gaussian process modelling. This research activity directly benefits our product, partners and customers. Sharing our progress and discoveries with the world also allows our peers to build on top of our work. Secondmind Labs is the on-ramp for promising ML research to be applied and deployed to support real-world decision-making. It’s our vehicle to showcase the company’s scientific expertise, and engage directly with the research community.
Our strategy
The Secondmind Labs mission outlined above requires us to tackle both theoretical and applied problems. Working on customer solutions often requires specialized tools that cannot be found in literature, and they are a great source of inspiration for theoretical work. Furthermore, the datasets used in ML literature to benchmark algorithms often are less noisy and easier to deal with than real-world data. We have a saying that “customer data keeps us honest” and working with this data is the ideal way to ensure that our theoretical contributions address these real world problems that people face.
We want our research to have an impact on the world around us, not just in academic circles.
Our approach consists in deriving decision-making from first principles. The main benefits are to guarantee, in specific contexts, that the decisions we recommend are optimal and can be explained. Furthermore, it provides a principled way for people - the experts tasked with making critical business decisions - to inject their expertise into the decision-making process to achieve the best outcome. We strongly believe that taking a mathematically sound approach is key to closing the gap between people and AI.
Finally, we are developing robust research toolboxes such as GPflow, which are implemented in Python/Tensorflow. These are our go-to software tools to explore new research ideas, enable fast prototyping of proofs of concepts and provide mature libraries to fuel the Secondmind Decision Engine. More toolboxes will be made available in the near future, so stay tuned!
A bright future for decision-making
The more we advance the scalability of probabilistic models and their integration into decision-making algorithms, the more powerful the Secondmind Decision Engine will become in helping people make better business decisions, ranging from demand planning in supply chain to engine tuning in automotive, and many points in-between.
There are various ways to interpret the “scalability” of probabilistic models and to how to advance it. From our point of view, scalability means the ability to quickly train models and make predictions when presented with a new case study. We are thus working along the lines of widening the range of dataset sizes and types supported by our models, developing very flexible models such as deep Gaussian processes and choosing automatically well suited models for the dataset at hand.
On the decision-making side, we are aiming at developing new reinforcement learning algorithms, with a focus on model based and offline policy learning methods that can make the most of our advanced probabilistic models. We are also building new stochastic optimal control approaches that when given a probabilistic model are able to optimise the tradeoff between operational cost and service level for complex and dynamic supply chain networks.
We will keep engaging with the community and sharing our progress of our research on these topics (and plenty of others!). You can expect to hear from us either at machine learning conferences, on Github where we host our open source toolboxes, and on here where we'll add news, blog-posts, and demos of our tech.
Check out our Twitter to stay informed on our latest developments.