Guillaume Drion
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Guillaume Drion has a total of 64 co-authors. Here are the top co-authors by number of shared publications:
- Rodolphe Sepulchre
University of Cambridge
14 shared publications Alessio Franci
Department of Electrical Engineering and Computer Science University of Liège B-4000 Liège Belgium
10 shared publications- Vincent Seutin
University of Liège
7 shared publications - Alessio Franci
University of Liège
6 shared publications Rodolphe Sepulchre
Department of Engineering, University of Cambridge , Cambridge , United Kingdom
6 shared publications- Kathleen Jacquerie
Volen Center and Biology Department, Brandeis University, Waltham, MA 02454
5 shared publications
Latest Publications
- Open Access
- 23 October 2026
- Published by Elsevier in Neural Networks
- Vol. 206
Parallelizable memory recurrent units
- 23 October 2026
- Published by Elsevier in Neural Networks
- Vol. 206
Open AccessParallelizable memory recurrent units
- Open Access
- 1 October 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (20) , 20LT01
A fully tunable ultra-low power current-mode memory cell in standard CMOS technology
This work introduces a fully tunable, ultra-low power unipolar memory cell inspired by the Schmitt-trigger comparator and designed in complementary metal–oxide–semiconductor (CMOS) using only nine transistors. The proposed circuit operates entirely in the current domain and exploits a novel feedback configuration between two interdependent Heaviside-like thresholding elements to produce tunable bistable switching behavior. Its three key parameters—threshold current, hysteresis width, and output gain—are independently tunable via programmable bias currents, enabling flexibility across diverse analog computing applications. Unlike prior Schmitt-trigger designs, it simultaneously achieves current-mode operation, nanowatt-range power consumption, and full tunability, solely using standard MOSFET elements. Schematic-level and post-layout simulations in a 180 nm CMOS process confirm robust hysteresis together with resilience to device mismatch, supply variation, and process variation, while the residual temperature drift of the switching thresholds can be compensated through a proportional to absolute temperature-based bias reference. The operating current additionally sets a predictable trade-off between bandwidth and power consumption. Building on this circuit, we develop a complete family of spike-based logic gates using three-level current encoding, where the bistable memory retains the polarity of the last spike on each input indefinitely, enabling asynchronous logic operations without temporal windowing or refresh mechanisms. The same circuit also serves as the primitive for Bistable Memory Recurrent Units in analog neural networks, where the quantized hidden states provide inherent noise immunity. Together, these capabilities position the design as a versatile building block for next-generation neuromorphic processors integrating memory, logic, and recurrent computation.
- 1 October 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (20) , 20LT01
Open AccessA fully tunable ultra-low power current-mode memory cell in standard CMOS technology
This work introduces a fully tunable, ultra-low power unipolar memory cell inspired by the Schmitt-trigger comparator and designed in complementary metal–oxide–semiconductor (CMOS) using only nine transistors. The proposed circuit operates entirely in the current domain and exploits a novel feedback configuration between two interdependent Heaviside-like thresholding elements to produce tunable bistable switching behavior. Its three key parameters—threshold current, hysteresis width, and output gain—are independently tunable via programmable bias currents, enabling flexibility across diverse analog computing applications. Unlike prior Schmitt-trigger designs, it simultaneously achieves current-mode operation, nanowatt-range power consumption, and full tunability, solely using standard MOSFET elements. Schematic-level and post-layout simulations in a 180 nm CMOS process confirm robust hysteresis together with resilience to device mismatch, supply variation, and process variation, while the residual temperature drift of the switching thresholds can be compensated through a proportional to absolute temperature-based bias reference. The operating current additionally sets a predictable trade-off between bandwidth and power consumption. Building on this circuit, we develop a complete family of spike-based logic gates using three-level current encoding, where the bistable memory retains the polarity of the last spike on each input indefinitely, enabling asynchronous logic operations without temporal windowing or refresh mechanisms. The same circuit also serves as the primitive for Bistable Memory Recurrent Units in analog neural networks, where the quantized hidden states provide inherent noise immunity. Together, these capabilities position the design as a versatile building block for next-generation neuromorphic processors integrating memory, logic, and recurrent computation.
- Open Access
- 12 June 2026
- Published by Oxford University Press (OUP) in PNAS Nexus
- Vol. 5 (7) , pgag213
Two-factor synaptic plasticity enables memory consolidation during neuronal burst firing
How can brain circuits remain plastic enough to encode new information while still stabilizing synaptic changes that support long-term memory? Many circuits switch between tonic spiking, which encodes external inputs, and burst firing, which is generated collectively; yet how these firing-state changes interact with synaptic plasticity to support consolidation remains unclear. Here, we ask whether burst epochs can provide a minimal, mechanistically interpretable route to stabilizing memories encoded during tonic firing. We introduce a two-factor synaptic plasticity rule in a conductance-based spiking network that switches robustly between tonic and burst regimes. The effective synaptic strength is expressed as the product of two factors: a primary, flexible factor updated by a Hebbian mechanism, and a secondary factor that captures stabilizing processes. The secondary factor is adjusted according to the rate of change of the primary factor. In a pattern recognition case study, the network encodes new inputs during tonic firing and undergoes burst epochs. This two-factor rule stabilizes previously learned patterns, integrates information across samples to support generalization, and improves robustness to noise. Ablation experiments show that these outcomes require a synergy between neural bursting activity and the two-factor plasticity rule: blocking secondary plasticity prevents stable retention, replacing bursts with quiescence leads to fading memories, and replacing bursts with additional tonic firing causes interference and noise sensitivity. Finally, a signal-to-noise ratio analysis across tonic-burst cycles identifies parameter regimes in which bursts either sharpen or weaken synaptic representations, consistent with consolidation or pruning, respectively.
- 12 June 2026
- Published by Oxford University Press (OUP) in PNAS Nexus
- Vol. 5 (7) , pgag213
Open AccessTwo-factor synaptic plasticity enables memory consolidation during neuronal burst firing
How can brain circuits remain plastic enough to encode new information while still stabilizing synaptic changes that support long-term memory? Many circuits switch between tonic spiking, which encodes external inputs, and burst firing, which is generated collectively; yet how these firing-state changes interact with synaptic plasticity to support consolidation remains unclear. Here, we ask whether burst epochs can provide a minimal, mechanistically interpretable route to stabilizing memories encoded during tonic firing. We introduce a two-factor synaptic plasticity rule in a conductance-based spiking network that switches robustly between tonic and burst regimes. The effective synaptic strength is expressed as the product of two factors: a primary, flexible factor updated by a Hebbian mechanism, and a secondary factor that captures stabilizing processes. The secondary factor is adjusted according to the rate of change of the primary factor. In a pattern recognition case study, the network encodes new inputs during tonic firing and undergoes burst epochs. This two-factor rule stabilizes previously learned patterns, integrates information across samples to support generalization, and improves robustness to noise. Ablation experiments show that these outcomes require a synergy between neural bursting activity and the two-factor plasticity rule: blocking secondary plasticity prevents stable retention, replacing bursts with quiescence leads to fading memories, and replacing bursts with additional tonic firing causes interference and noise sensitivity. Finally, a signal-to-noise ratio analysis across tonic-burst cycles identifies parameter regimes in which bursts either sharpen or weaken synaptic representations, consistent with consolidation or pruning, respectively.
- Open Access
- 1 June 2026
- Published by IOP Publishing in Neuromorphic Computing and Engineering
- Vol. 6 (2) , 024022
Energy-efficient implementation of spiking recurrent cells on FPGA
Spiking neural networks (SNNs) can significantly reduce energy consumption compared to conventional artificial neural networks when spiking activity is sparse and the neuron model is hardware-friendly. However, biologically faithful models are often too costly for hardware implementations like field-programmable gate arrays (FPGAs), whereas highly simplified models, such as the leaky integrate-and-fire (LIF) neuron, sacrifice essential neuronal dynamics. In this work, we present an FPGA accelerator for an SNN utilizing the spiking recurrent cell (SRC) model, which offers an intermediate level of biological plausibility and hardware efficiency between simple LIF and complex conductance-based models. The SRC model features continuous spikes and critical dynamical properties like the refractory period, yet remains mathematically simple enough for efficient FPGA deployment. To optimize SRC computation, we propose a set of mathematical simplifications such as piecewise-defined approximations that eliminate costly non-linear functions ( tanh , exp ) and using scaling to avoid floating-point arithmetic. We then integrate these units into a cohesive hardware architecture using a dedicated binding layer to permit modularity and scalability of the network VHDL architecture. To demonstrate this modularity, two complete networks, with respectively 1 and 4 SRC layers, were implemented and validated using Spiking Traces (SpT) derived from, respectively, the MNIST and Fashion-MNIST datasets. Weight matrices computed offline are stored directly in LUT-registers without retraining or hardware-specific adaptation to strictly evaluate the robustness of SRC. The reference implementation achieves 96.31 % accuracy on MNIST with a 220 -image SpT and a processing time of 1.7424 ms per digit. We further investigate accuracy-energy trade-offs by reducing the SpT length and quantizing synaptic weights down to 4 bits , achieving 93.32 % accuracy at 0.492 mJ per digit ( 55 images , 5 -bit weights) and 92.89 % at 0.394 mJ ( 44 images , 4 -bit weights). These results demonstrate that SRC-based SNNs deliver competitive performance and low energy consumption while preserving richer neuronal dynamics than standard LIF models.
- 1 June 2026
- Published by IOP Publishing in Neuromorphic Computing and Engineering
- Vol. 6 (2) , 024022
Open AccessEnergy-efficient implementation of spiking recurrent cells on FPGA
Spiking neural networks (SNNs) can significantly reduce energy consumption compared to conventional artificial neural networks when spiking activity is sparse and the neuron model is hardware-friendly. However, biologically faithful models are often too costly for hardware implementations like field-programmable gate arrays (FPGAs), whereas highly simplified models, such as the leaky integrate-and-fire (LIF) neuron, sacrifice essential neuronal dynamics. In this work, we present an FPGA accelerator for an SNN utilizing the spiking recurrent cell (SRC) model, which offers an intermediate level of biological plausibility and hardware efficiency between simple LIF and complex conductance-based models. The SRC model features continuous spikes and critical dynamical properties like the refractory period, yet remains mathematically simple enough for efficient FPGA deployment. To optimize SRC computation, we propose a set of mathematical simplifications such as piecewise-defined approximations that eliminate costly non-linear functions ( tanh , exp ) and using scaling to avoid floating-point arithmetic. We then integrate these units into a cohesive hardware architecture using a dedicated binding layer to permit modularity and scalability of the network VHDL architecture. To demonstrate this modularity, two complete networks, with respectively 1 and 4 SRC layers, were implemented and validated using Spiking Traces (SpT) derived from, respectively, the MNIST and Fashion-MNIST datasets. Weight matrices computed offline are stored directly in LUT-registers without retraining or hardware-specific adaptation to strictly evaluate the robustness of SRC. The reference implementation achieves 96.31 % accuracy on MNIST with a 220 -image SpT and a processing time of 1.7424 ms per digit. We further investigate accuracy-energy trade-offs by reducing the SpT length and quantizing synaptic weights down to 4 bits , achieving 93.32 % accuracy at 0.492 mJ per digit ( 55 images , 5 -bit weights) and 92.89 % at 0.394 mJ ( 44 images , 4 -bit weights). These results demonstrate that SRC-based SNNs deliver competitive performance and low energy consumption while preserving richer neuronal dynamics than standard LIF models.
- Open Access
A fast and tiny current as common generator of slow regular pacemaking in brain and heart
Slow pacemaking is a key physiological process in specific excitable cells, yet its underlying mechanisms remain debated. Here we identify a conserved, voltage-dependent pacemaker current that is essential for slow, regular firing in both midbrain dopaminergic neurons and sinoatrial node cells. Conductance-based models incorporating this current reproduce stable pacemaking, requiring a fast-activating, small-amplitude current. This is further confirmed by dynamic-clamp experiments in dopaminergic neurons. Replacing the pacemaker current in a model with a voltage-independent conductance such as the non-selective sodium leak channels fails to sustain slow rhythmicity, highlighting the necessity for an adequate voltage dependence. Our results suggest a novel and shared biophysical mechanism for slow pacemaking in neuronal and cardiac systems
Open AccessA fast and tiny current as common generator of slow regular pacemaking in brain and heart
Slow pacemaking is a key physiological process in specific excitable cells, yet its underlying mechanisms remain debated. Here we identify a conserved, voltage-dependent pacemaker current that is essential for slow, regular firing in both midbrain dopaminergic neurons and sinoatrial node cells. Conductance-based models incorporating this current reproduce stable pacemaking, requiring a fast-activating, small-amplitude current. This is further confirmed by dynamic-clamp experiments in dopaminergic neurons. Replacing the pacemaker current in a model with a voltage-independent conductance such as the non-selective sodium leak channels fails to sustain slow rhythmicity, highlighting the necessity for an adequate voltage dependence. Our results suggest a novel and shared biophysical mechanism for slow pacemaking in neuronal and cardiac systems
- Open Access
Two-stage synaptic plasticity enables memory consolidation during neuronal burst firing regimes
Neural circuits routinely alternate between input-driven tonic activity and collective burst firing. In the presence of Hebbian plasticity, bursts generate a robust attractor in weight space, creating a built-in drift that can be repurposed into a stabilizing trace of prior learning. We show that this phenomenon can be harnessed for memory consolidation through the introduction of a two-stage synaptic rule. The effective synaptic weight is defined as the product of a primary weight—updated by a Hebbian rule during both tonic and burst periods—and a secondary weight that updates in proportion with a coupling gain to the negative time-derivative of the primary weight. In a MNIST-like task, alternating tonic and burst epochs preserves earlier patterns, improves generalization to unseen inputs, and resists interference and noise, whereas replacing burst by quiescence or additional tonic epochs does not. Parameter sweeps reveal that coupling gain and the initial synaptic weights control whether bursts consolidate (“up-selection”) or prune (“down-selection”) synapses. Pairing the rule with alternative primary plasticity models yields distinct treatments of overlapping inputs, enabling either integration or separation. Studying switches in firing activity with a two-stage synaptic plasticity provides a plausible route to consolidation in biological and neuromorphic networks. Significance Statement: Neural circuits alternate between tonic spiking and burst firing, yet most models of synaptic plasticity are limited to a single firing regime. We introduce a two—stage synaptic rule in which a primary weight encodes activity during both states, while a secondary weight—engaged only during bursts— stabilizes learning from tonic periods. In conductance-based networks and a pattern recognition task, this rule preserves memories, improves generalization, and resists interference, whereas quiescence or extended tonic activity do not. The model further shows that bursts can consolidate or prune synapses depending on coupling gain and initial conditions. These findings identify a plausible, biologically motivated mechanism for how activity state transitions shape memory consolidation. Graphical
Open AccessTwo-stage synaptic plasticity enables memory consolidation during neuronal burst firing regimes
Neural circuits routinely alternate between input-driven tonic activity and collective burst firing. In the presence of Hebbian plasticity, bursts generate a robust attractor in weight space, creating a built-in drift that can be repurposed into a stabilizing trace of prior learning. We show that this phenomenon can be harnessed for memory consolidation through the introduction of a two-stage synaptic rule. The effective synaptic weight is defined as the product of a primary weight—updated by a Hebbian rule during both tonic and burst periods—and a secondary weight that updates in proportion with a coupling gain to the negative time-derivative of the primary weight. In a MNIST-like task, alternating tonic and burst epochs preserves earlier patterns, improves generalization to unseen inputs, and resists interference and noise, whereas replacing burst by quiescence or additional tonic epochs does not. Parameter sweeps reveal that coupling gain and the initial synaptic weights control whether bursts consolidate (“up-selection”) or prune (“down-selection”) synapses. Pairing the rule with alternative primary plasticity models yields distinct treatments of overlapping inputs, enabling either integration or separation. Studying switches in firing activity with a two-stage synaptic plasticity provides a plausible route to consolidation in biological and neuromorphic networks. Significance Statement: Neural circuits alternate between tonic spiking and burst firing, yet most models of synaptic plasticity are limited to a single firing regime. We introduce a two—stage synaptic rule in which a primary weight encodes activity during both states, while a secondary weight—engaged only during bursts— stabilizes learning from tonic periods. In conductance-based networks and a pattern recognition task, this rule preserves memories, improves generalization, and resists interference, whereas quiescence or extended tonic activity do not. The model further shows that bursts can consolidate or prune synapses depending on coupling gain and initial conditions. These findings identify a plausible, biologically motivated mechanism for how activity state transitions shape memory consolidation. Graphical
- Open Access
- 19 September 2024
- Published by Oxford University Press (OUP) in PNAS Nexus
- Vol. 3 (10) , pgae415
Dimensionality reduction of neuronal degeneracy reveals two interfering physiological mechanisms
Neuronal systems maintain stable functions despite large variability in their physiological components. Ion channel expression, in particular, is highly variable in neurons exhibiting similar electrophysiological phenotypes, which raises questions regarding how specific ion channel subsets reliably shape intrinsic properties of neurons. Here, we use detailed conductance-based modeling to explore how stable neuronal function is achieved despite variability in channel composition among neurons. Using dimensionality reduction, we uncover two principal dimensions in the channel conductance space that capture most of the variance of the observed variability. These two dimensions correspond to two sources of variability that originate from distinct physiologically relevant mechanisms underlying the regulation of neuronal activity, providing quantitative insights into how channel composition is linked to the electrophysiological activity of neurons. These insights allow us to understand and design a model-independent, reliable neuromodulation rule for variable neuronal populations.
- 19 September 2024
- Published by Oxford University Press (OUP) in PNAS Nexus
- Vol. 3 (10) , pgae415
Open AccessDimensionality reduction of neuronal degeneracy reveals two interfering physiological mechanisms
Neuronal systems maintain stable functions despite large variability in their physiological components. Ion channel expression, in particular, is highly variable in neurons exhibiting similar electrophysiological phenotypes, which raises questions regarding how specific ion channel subsets reliably shape intrinsic properties of neurons. Here, we use detailed conductance-based modeling to explore how stable neuronal function is achieved despite variability in channel composition among neurons. Using dimensionality reduction, we uncover two principal dimensions in the channel conductance space that capture most of the variance of the observed variability. These two dimensions correspond to two sources of variability that originate from distinct physiologically relevant mechanisms underlying the regulation of neuronal activity, providing quantitative insights into how channel composition is linked to the electrophysiological activity of neurons. These insights allow us to understand and design a model-independent, reliable neuromodulation rule for variable neuronal populations.
- Open Access
Dimensionality reduction of neuronal degeneracy reveals two interfering physiological mechanisms
Neuronal systems maintain stable functions despite large variability in their physiological components. Ion channel expression, in particular, is highly variable in neurons exhibiting similar electrophysiological phenotypes, which poses questions regarding how specific ion channel subsets reliably shape neuron intrinsic properties. Here, we use detailed conductance-based modeling to explore the origin of stable neuronal function from variable channel composition. Using dimensionality reduction, we uncover two principal dimensions in the channel conductance space that capture most of the variance of the observed variability. Those two dimensions correspond to two physiologically relevant sources of variability that can be explained by feedback mechanisms underlying regulation of neuronal activity, providing quantitative insights into how channel composition links to neuronal electrophysiological activity. These insights allowed us to understand and design a model-independent, reliable neuromodulation rule for variable neuronal populations.
Open AccessDimensionality reduction of neuronal degeneracy reveals two interfering physiological mechanisms
Neuronal systems maintain stable functions despite large variability in their physiological components. Ion channel expression, in particular, is highly variable in neurons exhibiting similar electrophysiological phenotypes, which poses questions regarding how specific ion channel subsets reliably shape neuron intrinsic properties. Here, we use detailed conductance-based modeling to explore the origin of stable neuronal function from variable channel composition. Using dimensionality reduction, we uncover two principal dimensions in the channel conductance space that capture most of the variance of the observed variability. Those two dimensions correspond to two physiologically relevant sources of variability that can be explained by feedback mechanisms underlying regulation of neuronal activity, providing quantitative insights into how channel composition links to neuronal electrophysiological activity. These insights allowed us to understand and design a model-independent, reliable neuromodulation rule for variable neuronal populations.
- 7 August 2023
- Published by Elsevier in Neural Networks
- Vol. 166, 645-669
Warming up recurrent neural networks to maximise reachable multistability greatly improves learning
- 7 August 2023
- Published by Elsevier in Neural Networks
- Vol. 166, 645-669
Warming up recurrent neural networks to maximise reachable multistability greatly improves learning
- Open Access
Warming up recurrent neural networks to maximise reachable multistability greatly improves learning
Training recurrent neural networks is known to be difficult when time dependencies become long. In this work, we show that most standard cells only have one stable equilibrium at initialisation, and that learning on tasks with long time dependencies generally occurs once the number of network stable equilibria increases; a property known as multistability. Multistability is often not easily attained by initially monostable networks, making learning of long time dependencies between inputs and outputs difficult. This insight leads to the design of a novel way to initialise any recurrent cell connectivity through a procedure called "warmup" to improve its capability to learn arbitrarily long time dependencies. This initialisation procedure is designed to maximise network reachable multistability, i.e., the number of equilibria within the network that can be reached through relevant input trajectories, in few gradient steps. We show on several information restitution, sequence classification, and reinforcement learning benchmarks that warming up greatly improves learning speed and performance, for multiple recurrent cells, but sometimes impedes precision. We therefore introduce a double-layer architecture initialised with a partial warmup that is shown to greatly improve learning of long time dependencies while maintaining high levels of precision. This approach provides a general framework for improving learning abilities of any recurrent cell when long time dependencies are present. We also show empirically that other initialisation and pretraining procedures from the literature implicitly foster reachable multistability of recurrent cells.
Open AccessWarming up recurrent neural networks to maximise reachable multistability greatly improves learning
Training recurrent neural networks is known to be difficult when time dependencies become long. In this work, we show that most standard cells only have one stable equilibrium at initialisation, and that learning on tasks with long time dependencies generally occurs once the number of network stable equilibria increases; a property known as multistability. Multistability is often not easily attained by initially monostable networks, making learning of long time dependencies between inputs and outputs difficult. This insight leads to the design of a novel way to initialise any recurrent cell connectivity through a procedure called "warmup" to improve its capability to learn arbitrarily long time dependencies. This initialisation procedure is designed to maximise network reachable multistability, i.e., the number of equilibria within the network that can be reached through relevant input trajectories, in few gradient steps. We show on several information restitution, sequence classification, and reinforcement learning benchmarks that warming up greatly improves learning speed and performance, for multiple recurrent cells, but sometimes impedes precision. We therefore introduce a double-layer architecture initialised with a partial warmup that is shown to greatly improve learning of long time dependencies while maintaining high levels of precision. This approach provides a general framework for improving learning abilities of any recurrent cell when long time dependencies are present. We also show empirically that other initialisation and pretraining procedures from the literature implicitly foster reachable multistability of recurrent cells.