Calda AI

Large time-series models (LTMs) are to time-series data what LLMs are to language. LTMs are on the cusp of delivering state-of-the-art forecasting capability across a wide range of domains, changing the way we analyse time-series data forever. Calda AI researchers Justin Alsing and Benjamin Wandelt (with collaborators from Johns Hopkins, University of Amsterdam and Capital One) made an important step forward in the pursuit of LTMs last week, demonstrating favourable performance scaling with model size, data, and compute. Watch this space for breakthrough foundation models for time-series forecasting!

Calda AI scientists are pushing the limits of what’s possible with Graph Neural Networks (GNNs). Graph Neural Networks are one of the pillars of modern physics x AI modelling, which is steadily revolutionising industrial process optimisation. At their core, GNNs rely on aggregating together lots of little bits of information from across your system or process. But how do you aggregate all of your information in an optimal way, to get the most out of your data? In a recent paper, Calda scientists Justin Alsing and Benjamin Wandelt (together with collaborator Lucas Makinen from Imperial College London) developed “Fishnets” — information-theoretically optimal aggregation for GNNs. By optimizing GNN aggregation we are able to build more accurate GNNs, from less data and smaller architectures, while being significantly more robust when your data in deployment do not follow the same distribution as your training data. Better aggregation -> More accurate predictions, from less data and smaller architectures -> Better process optimisation. Check out the paper here: https://lnkd.in/gPkrkjx2 

Making complex decisions in the face of limited data and uncertainty is the cornerstone of industrial optimisation, from process optimisation through strategic planning and decision-making. Calda AI scientists recently published a key breakthrough in leveraging machine-learning to solve complex decision-making tasks, efficiently and optimally. With our new framework, you can get optimal decision recommendations in the most complex situations, covering many situations that were previously impossible with existing methods. In cases where existing approaches were possible (in theory), our approach is up to 1000x faster than the previous state-of-the-art. Get in touch if you are interested in partnering with us at Calda AI to deploy these new capabilities in your business, to crack your hardest decision-making and planning problems. Quick summary and link to the paper below: -------------------------------------- Making optimal decisions in the face of uncertainty is the end game of reasoning from data, and arguably the most important basic task we undertake as humans. Bayesian decision-theory provides the unique mathematical framework for computing optimal decisions from limited data, assuming some reasonable axioms about what characterises the "rational decision maker". In this framework, you properly account for all of your sources of uncertainty, and find the action that optimises your expected gains while mitigating risks. However, the calculations involved in computing optimal decisions are typically prohibitively expensive - approximations lead to sub-optimal decisions, and poorer business outcomes. We introduce a new, machine-learning based system for calculating optimal Bayesian decisions, which bypasses the expensive (inference, integration, and optimisation) steps in the traditional approach. Our approach relies only on the ability to simulate your problem, making it very generally applicable (for both tractable and intractable likelihoods). It is also incredibly efficient - we need a factor of 100-1000 fewer simulations than you would typically require using standard methods - and it works on problems that were previously impossible due to their complexity (ie., where likelihoods are intractable). This opens up exciting new capabilities in Bayesian decision-making, particularly in the previously challenging regime where likelihoods are intractable and simulation expensive. https://lnkd.in/d5raFatu

Leveraging AI to learn the physics governing a process - directly from the data - is one of the key ways we at Calda AI are building groundbreaking new capabilities in process optimisation and control. Physics-based AI systems allow us to model and optimise processes that were previously too complex, or measurements too indirect or messy. By combining the principles of physical modelling and AI together, we can build process models that are both more accurate and more robust, from less (and lower quality) data than before. In a recent landmark paper, Calda AI scientist Niall Jeffrey and collaborators used Graph Neural Networks and symbolic regression to re-learn the exact formalism for Newtonian celestial mechanics from orbital solar system data. Breakthroughs like these are changing the face of what's possible in industrial process optimisation! Get in touch to find out how Calda AI can help transform your process optimisation capabilities. https://lnkd.in/dCiwYzhF

Delighted to announce early investment and support from SU HOLDING! SU Holding have been great supporters since the beginning of our journey and we're very happy to be part of their community and portfolio. Thanks especially to Mona Wilcke and her team, we look forward to a long a fruitful partnership! https://lnkd.in/dG5XD7-i