We believe that an energy-aware perspective is imperative for developing true artificial intelligence and that an AI-first approach is imperative for solving the challenge of transitioning to a sustainable future based on renewable energy. Our research is organized under two broad themes:
- Energy in Artificial Intelligence (the use of energy in developing AI). This area encapsulates research on the tradeoffs and limitations of energy usage in solving basic AI tasks such as classification, ranking or planning & search. The notion of what constitutes AI in a machine is most famously captured by the notion of the Turing test. Currently, this is an unfair test because AI systems that offer impressive performance on tasks in restricted domains (e.g., AlphaGo) are not judged fairly: while the human brain uses around 20W-25W of energy to perform such feasts, current AI methods use three to four orders more energy. Energy-efficient AI methods will also be required in the years to come to reduce their energy footprint in edge and cloud computing scenarios.
Artificial Intelligence in Energy (the use of AI technology in generating, storing and managing energy). This area encapsulates research into the development and application of state-of-the-art AI methods to particular problems arising from the larger reliance on renewable energy, starting with more refined modeling of batteries to enable the extension of their working lifetime, all the way to the development of AI methods for control of domestic energy consumption, necessitating research into computational privacy and explainable AI.
Energy in Artificial Intelligence
Up until today, the primary focus of research on AI methods has been on predictive accuracy and its dependence on the amount of training data rather than the amount of energy that is necessary — this is understandable as digital data of the physical world was scarce in the early decades of AI. Going forward, data is (typically) abundant for many tasks but energy is becoming a limiting and economical factor. This is not sustainable in the long term, in particular as the amount and dependence on AI methods in people’s life is only going to increase while the amount of energy that is available to society remains constant. The focus of this new strand of research will be three-fold:
Methods. Develop algorithms for low-energy learning and prediction as well as simulation. This includes learning algorithms which operate on data and parameters that are represented through spiking networks or low-precision numbers rather than high-precision numbers as well as hardware-accelerated sampling and simulation algorithms for multi-modal likelihoods such as those found in recommender models. Moreover, we will investigate AI algorithms which work on noisy computation and storage systems.
Measurement. Establishing a widely-used benchmark environment for measuring both the accuracy and energy consumption of learning and prediction algorithms on real-world image & audio analysis, text understanding, time-series prediction and game-playing tasks (whether they are using purely statistical- and/or simulation-based AI algorithms).
Theory. Developing a theory of learning which is characterized by both the predictive accuracy of the learned functions as well as the energy necessary to achieve this. We begin to formulate a version of the probably-approximately correct (PAC) framework that incorporates a characterization of the learning and prediction algorithm in terms of irreversible information processing steps and evaluate the tightness of such bounds for basic AI algorithms such as the perceptron learning algorithm or drop-out in deep learning.
Artificial Intelligence in Energy
- Batteries. One of the key technical solutions that enables the use of renewable energy are rechargeable batteries; they not only enable carbon-neutral transportation but they are increasingly used in people’s homes. Today, there are already over 10 million electric vehicles (EV) on roads worldwide. Battery aging is fundamentally characterized by the material degradation during an electro-chemical process (between anode, cathode, separator and electrolyte) but we can neither measure the ground truth of degradation at the molecular level (of ions) nor is the aggregation mechanism from ions all the way to measurable quantities such as currents, voltages and temperatures fully understood. We will start with developing large-scale and efficient, approximate inference methods for probabilistic models of battery material degradation based on observable (aggregated) data such as current, voltage and temperature time series only.
- Smart Homes. In a second line of research, we base our work on the belief that the home is the most private space for any individual and both sensing and automation conducted in this space by algorithms requires full trust and transparency from and with its users. Our research is focused on privacy-aware learning algorithms to predict the occupancy of individual spaces (i.e., rooms) in people’s home as well as explainable predictions for the predicted visitation schedule of spaces, as well as learning rules in (probabilistic) knowledge bases in order to combine common knowledge of physical spaces and everyday life with the observable patterns of usage behavior.