Neuromorphic technology is evolving rapidly, but it still faces two critical problems. Firstly, few compelling applications exist that demonstrate the superiority of neuromorphic technology over classical computing, limiting its widespread adoption and commercialisation. Some insightful applications include keyword spotting [BCHE19],, real-time modelling of microcortical circuits [RPR+20] the implementation of nearest-neighbor searches [FOF+20] and LASSO optimisation via the spiking locally competityive algorithm [DSL+18]. Secondly, the use of neuromorphic technology by neuroscientists is scarce, with physicists, mathematicians, engineers and computer scientists as the principal user communities. Discrepancies remain between the variables of interest in the laboratory to experimental neuroscientists and the parameterisations realisable on neuromorphic hardware, making the models of the latter too abstract or simplified. For example, while experimental data is acquired in the form of ion-channel conductances from patchlamp experiments, local field potentials, effects of parmacological blockers and neurotransmitter on neurons, and intra-cellular and extra-cellular ion concentrations, the neuromorphic hardware is configured in terms of synapse level connectivity (point-to-point adjacency matrices), membrane and postsynaptic potential time constants, inter spike intervals and firing probabilities. We contribute to addressing both issues by implementing stochastic processes arising in neuronal dynamics, developing applications for neuromorphic hardware of both biological and technological interest. On the application level, we harness recent theoretical developments and results from conventional hardware on the computational power of stochastic neuronal dynamics for problem-solving. We do so by replicating and improving on the solution of constraint satisfaction problems (CSPs) with stochastic networks of spiking neurons. For this, we have used both the SpiNNaker and the Loihi neuromorphic chips, harnessing the advantages of each one. Our results demonstrate the usability of neuromorphic technology to solve hard problems with industrial application for which conventional machine learning faces challenges. The performance of our CSP solver is comparable to that of the state of the art solutions, and is a basic module for implementing solution strategies of increasing sophistication as well as for gaining insights into how living beings solve CSP problems in the real world. To bridge the gap with experimental neuroscience, we demonstrate the implementation on SpiNNaker of models of the intrinsic currents generated by voltage-gated ion channels, as well as of realistic postsynaptic potentials. Both of these arise in the neuronal membrane from complex ion-channel dynamics which are stochastic by their very nature. Our work paves the way to integrate neuromorphic technology with the worlds of neurophysiology and neurogenetics, allowing a direct relation with processes of interest in neuropharmacology, such as protein-drug interaction, as well as in whole-cell recordings of phenomena such as homeostasis and intrinsic plasticity. Hence these results at the cellular level open the way for the use of neuromorphics in medical applications and scientific enterprise in neuroscience.
Date of Award | 1 Aug 2020 |
---|
Original language | English |
---|
Awarding Institution | - The University of Manchester
|
---|
Supervisor | David Lester (Supervisor) & Steve Furber (Supervisor) |
---|
- Voltage gated ion channel currents
- Postsynaptic Potentials
- Constraint Satisfaction
- SpiNNaker
- Loihi
- Neuromorphic Hardware
Stochastic Processes For Neuromorphic Hardware
Fonseca Guerra, G. (Author). 1 Aug 2020
Student thesis: Phd