Here at Arm Research, we place a high value on our academic and industry collaborations, and our latest one aiming to drive new developments in Artificial Intelligence (AI) and Machine Learning (ML). The European Lab for Learning and Intelligent Systems (ELLIS) is building a network of some of the brightest minds in AI and ML. They are enabling cutting-edge academic research to be combined with start-up and industry expertise, for maximum impact [1]. In September 2019, a network of smaller ELLIS Units was established with a focus on AI at a local level, while also taking a global view. Arm Research is excited to be part of the Cambridge ELLIS Unit. Taking advantage of the cutting-edge ML and AI infrastructure available at the University of Cambridge (another of our research partners), members of the Unit have expertise in Bayesian statistics and Probabilistic ML – two of the key elements needed to address challenges in data efficiency and flexible adaption.
The success of deep learning depends heavily on the abundance of training data. In a deterministic setting, deep neural networks (DNNs) perform extremely well, but the real world deviates from the typical laboratory environment. DNN models which are trained to high accuracy often perform poorly after deployment due to distribution shifts. This may occur either because the model was not trained with target domain data, or the model is overfitted to the limited training data.
To give a concrete example, in autonomous driving, future states are unknown and the system is being continuously updated by multiple readings taken by many onboard sensors (for example, LiDAR, Radar, Camera). Often, the vehicle’s AI system is designed in a simulated environment and then deployed in a completely unfamiliar environment. Not only is the target environment new, but the intents of the other moving vehicles and pedestrians are also likely to be unknown. In such a scenario, enforcing the model to behave with a fixed set of rules may be too constrained to be realistically useful. A probabilistic framework allows a vehicle’s AI system to learn a distribution over many possible futures, predicting with the right level of uncertainty in ambiguous scenarios.
Another incredibly challenging industrial domain is healthcare, where data is scarce, and every automated decision must be explainable. Often having a yes or no answer is not enough, especially given that models are trained on limited data. Probabilistic models allow the capture of these settings in a principled way. It is honest about model uncertainty which allows domain knowledge to be added without punishing performance. In recent years, research has shown that incorporating probabilistic techniques into deep learning can improve the robustness of many critical applications - see our research on Stochastic-YOLO for robust object detection to learn more. Although the full extent of Probabilistic ML has not yet been generalized to all recent developments of AI, it is nonetheless a success that Bayesian thinking is deeply inherent and crucial to numerous innovations that have recently emerged. The next few years are sure to bring a new wave of practical ML applications, combining the right proportions of Bayesian and deep learning approaches to improve robustness in industrial systems. Our collaboration with the Cambridge ELLIS Unit draws on this, enhancing understanding and solving key problems in this area.
Probabilistic ML plays a vital role in safety-critical AI applications, especially when they are effectively combined with state-of-the-art DNNs. We envisage that going forward, many AI applications, such as autonomous driving, healthcare, and industrial robotics, will see an increased use of a mixed approach that combines the best of both worlds. At the Arm ML Research Lab, we have been exploring how these types of emerging ML workloads could be efficiently implemented on Arm hardware. See our research on Quantised Bayesian Neural Network to learn more. The Bayesian framework for model fitting allows us to describe model uncertainties in a principled manner. However, this is not trivial, as traditional Bayesian models are computationally expensive due to their inherent structure and the way inference works under such assumptions. The Cambridge ELLIS Research Unit is one of the leading groups in Bayesian approaches to modeling and inference. Through our collaboration, we have been investigating efficient ways to extract predictive uncertainties from traditional DNNs and CNNs used on many modern applications. With the Cambridge ELLIS Unit, we hope to share real-life problems with academic researchers and bring their expertise to Arm to improve our investigations into practical aspects of probabilistic deep learning.
"The foundation of the Cambridge Unit was possible thanks to Arm's support, which will enable the recruitment of research and administrative staff, the mobility of Cambridge PhD students to other ELLIS Units in Europe, and the organization of ELLIS events such as workshops and summer schools. Most of the work done at the Cambridge Unit has a strong mathematical foundation, drawing on fundamental frameworks such as probabilistic modeling and Bayesian inference. This is very well aligned with the work done at Arm's ML Research Lab in Cambridge, especially on the efficient implementation of probabilistic ML on resource-constrained devices." Dr. José Miguel Hernández-Lobato, University of Cambridge, UK
"The foundation of the Cambridge Unit was possible thanks to Arm's support, which will enable the recruitment of research and administrative staff, the mobility of Cambridge PhD students to other ELLIS Units in Europe, and the organization of ELLIS events such as workshops and summer schools. Most of the work done at the Cambridge Unit has a strong mathematical foundation, drawing on fundamental frameworks such as probabilistic modeling and Bayesian inference. This is very well aligned with the work done at Arm's ML Research Lab in Cambridge, especially on the efficient implementation of probabilistic ML on resource-constrained devices."
Dr. José Miguel Hernández-Lobato, University of Cambridge, UK
We look forward to seeing what this collaboration brings in the future. Have a question or want to learn more about the Cambridge ELLIS Unit? Click the following links.
Explore Cambridge ELLIS Unit Contact Partha Maji
Find out how we help academic and industry partners achieve research success, on our website.