A Neuromodulable Current-Mode Silicon Neuron for Robust and Adaptive Neuromorphic Systems

TL;DR

Problem: replicate robust, context-dependent computation in neuromorphic hardware via neuromodulation. Approach: develop a fully analog, current-mode mixed-feedback neuron with fast, slow, and ultraslow dynamics, using $I_f$, $I_s$, $I_u$ and sigmoidal blocks, plus a biologically inspired positive-feedback inactivation; implemented in 180 nm CMOS and analyzed with a mathematical model and hardware experiments. Contributions: a tunable steady-state analysis framework linking circuit biases to spiking/bursting, two current-mode blocks for feedback, experimental validation of spiking, bursting, and neuromodulation, plus temperature robustness and quantified energy efficiency. Impact: enables scalable, adaptive neuromorphic systems capable of real-world tasks with intrinsic neuromodulation capabilities.

Abstract

Neuromorphic engineering makes use of mixed-signal analog and digital circuits to directly emulate the computational principles of biological brains. Such electronic systems offer a high degree of adaptability, robustness, and energy efficiency across a wide range of tasks, from edge computing to robotics. Within this context, we investigate a key feature of biological neurons: their ability to carry out robust and reliable computation by adapting their input response and spiking pattern to context through neuromodulation. Achieving analogous levels of robustness and adaptation in neuromorphic circuits through modulatory mechanisms is a largely unexplored path. We present a novel current-mode neuron design that supports robust neuromodulation with minimal model complexity, compatible with standard CMOS technologies. We first introduce a mathematical model of the circuit and provide tools to analyze and tune the neuron behavior; we then demonstrate both theoretically and experimentally the biologically plausible neuromodulation adaptation capabilities of the circuit over a wide range of parameters. All the theoretical predictions were verified in experiments on a low-power 180 nm CMOS implementation of the proposed neuron circuit. Due to the analog underlying feedback structure, the proposed adaptive neuromodulable neuron exhibits a high degree of robustness, flexibility, and scalability across operating ranges of currents and temperatures, making it a perfect candidate for real-world neuromorphic applications.

Figures (13)

• Figure 1: Neuromodulation in thalamocortical relay neurons (top, adapted from sherman_tonic_2001) vs. neuromodulation recorded in our silicon neuron (bottom). For the same input current step (middle trace), neuromodulation enables a switch between tonic spiking (left) and a single transient burst (right), corresponding respectively to a linear rate-based encoding or a nonlinear "wake-up call" response to changes. Subthreshold integration dynamics are also reproduced in our neuron, but are much faster than biological neurons due to the limited capacitor sizes.
• Figure 2: Mixed-feedback neuromodulable neuron structure. Red (resp. blue) highlighted arrows indicate positive (resp. negative) feedback paths. The symbols and circuit implementations of the two types of blocks are detailed in \ref{['fig:DPI_circuit', 'fig:sigmoid_circuit']}. The low-pass filter blocks are denoted using their transfer function, and the sigmoid blocks using the shape of their steady-state response. The dashed arrows represent the positive feedback inactivation illustrated in \ref{['fig:inactivation']}. $I_f$ is the current-mode analog of the membrane potential. $I_s$ and $I_u$ respectively provide slow repolarization and ultraslow spike-frequency adaptation currents. The nonlinear sigmoid blocks provide positive feedback to create the fast spike upstroke and slow regenerative dynamics.
• Figure 3: Spike (A) and burst (B) with (right) and without (left) positive feedback inactivation. The temporal extent of spikes and bursts is reduced thanks to the inactivation, but they remain long enough in time to directly signal to physical systems (in the spirit of deweerth_simple_1991).
• Figure 4: dpi circuit (from bartolozzi_ultra_2006bartolozzi_synaptic_2007livi_current-mode_2009chicca_neuromorphic_2014). Q2-3 are the differential pair receiving the input current $I_\text{in}$. With Q1, Q2 sets the gain $G=\frac{I_\text{th}}{I_\tau}$ of the circuit by diverting part of the input current depending on $I_\text{th}$, while Q3 sets the voltage $V_C$ of the capacitor. Q4 provides the leakage current $I_\tau$ that discharges the capacitor over time. Q5 provides an output current $I_\text{out}$ depending on the capacitor voltage $V_C$.
• Figure 5: Current-mode sigmoid circuit with inactivation mechanism. Q1-2 form a simple current comparator circuit: when $I_\text{in}$ > $I_\text{thr}$, the voltage $V_\text{cmp}$ starts to increase. Q3-5 control the rate of increase of $V_\text{cmp}$: the higher $V_\text{lin}$ is, the more current is drawn by this branch, increasing the response range of the comparator. Q6-7 set the gain of the circuit, defining the maximum output current $I_\text{gain}$ at which the sigmoid saturates. The role of M1-5 is to implement the inactivation mechanism by subtracting $I_\text{mod}$ from the parameter $\hat{I}_\text{gain}$.