Reliable Quantum Neuromorphic Graph Neural Networks

Mimicking the operation of biological brains, neuromorphic engineering (NE) aims at
improving the energy efficiency of computational tasks that can be carried out by
encoding information with time. Recently, time-encoded spiking models have found
widespread applications for the implementation of attention mechanisms in transformers,
as the detection of timing coincidences can efficiently replace attention weights.
Neuromorphic architectures also offer a natural platform to study probabilistic computing
applications (PC), e.g. Monte Carlo simulations and Bayesian Neural Networks, whose
inherent stochasticity, in turn, makes them ideal choices to interface with nascent
quantum computing (QC) devices.
Quantum models can effectively propagate timing information by leveraging noisy gates
and state relaxations, making spiking neurons promising candidate computational
models for noisy intermediate scale quantum (NISQ) computers.
This project will explore to what extent the principles of neuromorphic modelling can be
embedded or applied to quantum devices, in two-fold directions.
- One focus will be uncertainty, expressivity or trainability quantification of NISQ
applications: by leveraging the inherent randomness of the computational
substrate, a central working principle in neuromorphic engineering, we will explore
further the implications of parallelisms across NE and digital/analog
parameterised Quantum Circuits. A grand goal in this direction would be
producing reliable decision-making machines that can quantify their own
uncertainty.
- In parallel, novel quantum neuromorphic models could be found, ranging from
attention-based transformers, to more general graph neural networks, or hybrid
intermediate representations for NISQ circuits, with the target of
increasing the usability and applicability of (particularly analog) NISQ devices.