Data analysis systems in applications ranging from disaster recovery to defense and security, must often operate in environments with no network uplink. These so-called access-denied environments also often impose severe constraints the compute capabilities of systems, as well as constraints on how much data they have access to. The data in question are often real-time multi-modal sensor data measurements (e.g., gas/environmental sensors, audio, low-resolution infrared, embedded LIDAR).
To adapt to their environments in the absence of network connectivity, these systems must not only be able to generate inferences at low computational costs, but must also be able, ideally, to train models in situ, as offloading computation to a remote server is not possible. And, quite often, the available data is limited either by the nature of the modality or by the hardware constraints of the sensing systems.
This special interest group will bring together researchers at the Turing spanning backgrounds ranging from computer science theory, statistics and applied mathematics, to hardware architectures for machine learning at the edge of the network, and low-power sensing systems with limited data.
Explaining the science
Data-constrained machine learning methods
Machine learning is to learn a function from data, using an appropriate representation or hypothesis space combined with an appropriate optimisation algorithm applied to an appropriate cost function. Low-shot and zero-shot learning learn a model of a system with as few observations as possible by using prior knowledge about related systems. They improve performance by using knowledge from related learning tasks to reduce the size of the hypothesis space or to improve the optimisation algorithm.
When learning methods use data from systems constrained by physics, it is possible to exploit these physical constraints to reduce the size of the hypothesis space, which in turn increases how quickly the learning algorithm can learn. Several examples in the research literature show that zero-shot learning benefits from having a general, expressive, and modular knowledge representation.
Resource-constrained machine learning methods
In parallel with challenges to knowledge representation in data-constrained systems, when those systems are also constrained in their computational resources, there are several exciting research challenges in performing both training and inference. These methods range from methods also familiar in the data-constrained context (e.g., exploiting knowledge representations), to methods that are at the surface primarily tradeoffs on numerics (e.g., binarised neural networks).
AI in resource- and data-constrained environments
The interest group brings together a cross-disciplinary group of researchers from the Alan Turing Institute and beyond, to discuss and seed research collaborations on data-constrained machine learning methods and the application of machine learning and artificial intelligence in systems with constrained computational resources. Examples of the kinds of constrained computational environments of interest range from miniature energy-scavenged environmental sensing systems, to autonomous machines with requirements of low size weight and power (low-SWaP systems).
The special interest group will crystalise existing interests being pursued independently by disparate researchers at the Turing, using these mechanisms:
- a reading group / journal club
- an ambition to combine the insights from the reading group into a survey article
- an ambition to organise a workshop
- to facilitate cross-fertilisation of ideas across the disparate activities of the constituent researchers in the interest group and to thereby spark new cross-disciplinary collaborative research
The interest group has an opportunity to influence wider discussions across the Turing. The challenges posed by data privacy concerns means that there will be an ever-increasing push to perform more data analysis on-device and not on a server. The insights and experience from the interest group could therefore contribute to other research at the Turing on trustworthy machine learning.
Machine learning in data-constrained environments
Low-shot / zero-shot learning
Physics-based machine learning model representations, cost functions, and optimization
Machine learning in low-size-weight-and-power systems
Hardware implementations of machine learning in low-power embedded systems (ranging from microcontrollers to FPGAs)
Efficient machine learning hardware based on novel electronic devices (e.g., 2D materials, memristors) with the specific objective of reduced energy usage
Cross fertilisation of ideas (e.g., from compressed sensing)
How to get involved
Soteris Demetriou, Imperial College London
Hamed Haddadi, Imperial College London
Paul Patras, University of Edinburgh
Vasileios Tsoutsouras, University of Cambridge