Conventional means of implementing artificial intelligence (AI) introduce serious impediments related to data centralisation, model bias, and exclusion of certain users. These hazards become particularly concerning as our reliance on AI grows and we start employing it in a range of life-critical tasks, from medical diagnostics to autonomous driving. With the goal of making AI more inclusive and more efficient, in this project we strive to set the ground for such AI to become feasible in years to come. XS is based on the insight that the key limiting factors for efficient AI the computational burden of model training can be selectively reduced to enable the incorporation of broader populations and attainment of more detailed learning goals.
To realise XS, we will:
- Advance beyond backpropagation-based deep learning training methods and enable lightweight training particularly well suited for edge computing devices;
- Design and implement dynamically-tunable approximation of on-device neural network training, thus enabling heterogeneous devices to synchronously work towards joint AI models;
- Explore a range of novel paradigms for distributed learning that combine federated and split learning, and thus enable flexible distribution of computation (neural network training) over a set of heterogeneous computing devices.
XS unites previously independent perspectives on distributed machine learning, approximate computing, and edge systems computing, and breaks new g