Olfactory system as a blueprint for novel neural architectures for spacecraft autonomy

Biology has been a major source of inspiration for artificial intelligence and hence for advanced spacecraft autonomy. A prominent example is the convolutional neural network architecture - nowadays the standard model for image recognition tasks - which was inspired by the visual neural system of mammals and initiated the ongoing deep learning revolution in 2012. However, despite this success, other senses have been mostly overlooked in the pursuit of novel neural network architectures and learning paradigms.

One such biosensor modality is olfaction. In general, olfactory systems are highly plastic, show high robustness towards noise and generalise well to unseen stimuli. Similar to vision systems, olfactory systems are composed of stereotypical motifs that can be utilised as a blueprint for exploring novel neural network architectures for machine learning applications. For instance, olfaction combines concepts from biology such as Hebbian learning, random projections, neuromodulation of plasticity and lateral competition between neurons in a clear and approachable setting.

In recent work, the olfactory system of insects has been investigating for solving machine learning problems like few-shot digit classification [1,2] as well as odour localization tasks on the planetary surface [3]. In this project, we extend these studies by identifying interesting architectural and functional concepts observed in olfactory systems and applying those concepts to a variety of benchmark tasks that are relevant for the space sector, such as image classification and optimal control.

Project overview

This project has three main goals:

1. Identifying and implementing the essential aspects of olfactory systems in a pyTorch model.
2. Investigating possible extensions of this model by combining it with other neural network architectures or by applying modern optimisation techniques, e.g., evolutionary optimisation, end-to-end learning using gradient descent or meta-learning strategies.
3. Benchmarking of the model on few-shot learning tasks relevant for onboard AI applications.

The result of this study will drive forward the search for computational architectures that are energy efficient, performant and potentially realisable on emerging hardware platforms, such as neuromorphic processors. Moreover, this work will contribute to the search for novel computing paradigms that might re-envision how we realise artificial intelligence onboard spacecraft.

References:

[1] C. B. Delahunt & J. Nathan Kutz, Putting a bug in ML: The moth olfactory network learns to read MNIST (2018).

[2] R. Huerta & T. Nowotny, Fast and robust learning by reinforcement signals: explorations in the insect brain (2009).

[3] de Croon, G. C., O'connor, L. M., Nicol, C., & Izzo, D. Evolutionary robotics approach to odor source localization. Neurocomputing (2013).