Single Neuromorphic Memristor closely Emulates Multiple Synaptic Mechanisms for Energy Efficient Neural Networks

TL;DR

This work shows that a two-terminal SrTiO$_3$ memristor can intrinsically realize six synaptic functions—long- and short-term memory, plasticity, meta-plasticity, and in-memory multiplication—within a single device, enabling bio-inspired ST-Hebb synapses. By integrating these memristive synapses into a deep short-term plasticity network (m-STPN) and adapting the model to hardware constraints, the authors train an Atari Pong agent with energy-efficient in-memory computation. Energy analyses reveal large gains over GPU implementations (up to at least $96\times$), driven by reduced memory traffic and sparse short-term updates, with further gains anticipated from hardware scaling and improved retention. The results highlight a promising path toward scalable, energy-efficient neuromorphic hardware that leverages intrinsic synaptic dynamics for real-time learning in dynamic environments.

Abstract

Biological neural networks do not only include long-term memory and weight multiplication capabilities, as commonly assumed in artificial neural networks, but also more complex functions such as short-term memory, short-term plasticity, and meta-plasticity - all collocated within each synapse. Here, we demonstrate memristive nano-devices based on SrTiO3 that inherently emulate all these synaptic functions. These memristors operate in a non-filamentary, low conductance regime, which enables stable and energy efficient operation. They can act as multi-functional hardware synapses in a class of bio-inspired deep neural networks (DNN) that make use of both long- and short-term synaptic dynamics and are capable of meta-learning or "learning-to-learn". The resulting bio-inspired DNN is then trained to play the video game Atari Pong, a complex reinforcement learning task in a dynamic environment. Our analysis shows that the energy consumption of the DNN with multi-functional memristive synapses decreases by about two orders of magnitude as compared to a pure GPU implementation. Based on this finding, we infer that memristive devices with a better emulation of the synaptic functionalities do not only broaden the applicability of neuromorphic computing, but could also improve the performance and energy costs of certain artificial intelligence applications.

Figures (4)

• Figure 1: Biologically inspired synaptic functions and their memristor implementation. a) Organisation of the mammalian brain with several biological neurons connected through synapses. When a postsynaptic spike (light blue) coincides with a presynaptic spike (light green) the corresponding synaptic coupling is strengthened (Hebbian plasticity) for a limited amount of time (short-term plasticity). This bio-physical process is illustrated in the circular insets: (I) an influx of ions (e.g. $Ca^{2+}$) through the postsynaptic voltage gated ion channels leads to (II) an increased number of synaptic receptors, which increases the synaptic weight. (III) The weight subsequently decays back to its original value due to the receptors gradually detaching from the membrane. b) Table comparing the synaptic functions of artificial synapses in standard ANNs ("Artificial" column) and biological synapses ("Biological" column). The plot on the right shows the weight of a biological synapse as a function of time. The short-term weight (F) is updated ($\Delta F$) when the pre- and post-synaptic spikes coincide. Additionally, the decay time of F can be controlled, which corresponds to meta-plasticity. c) Bio-inspired Short-Term Plasticity Neuron (STPN) model combining a conventional neuron model with short-term Hebbian (ST-Hebb) synapses. d) Hardware implementation of a neuromorphic ST-Hebb synapse with a Cr/Pt-SrTiO3-Ti memristor. The device measurement on the right mirrors the biological functions of ST-Hebb combining memory and computation as well as long- (W) and short-term (F) dynamics.
• Figure 2: DC and dynamical behavior of multi-functional memristive synapses. a) Conductance vs. voltage characteristic of the fabricated Cr/Pt-STO-Ti memristors. The black arrows indicate the counter-clockwise switching direction. b) Sketch of the device stack and of the underlying switching mechanism. The two insets zoom into the Pt-STO interface at different applied voltages, showing the dynamics of interfacial oxygen ions (O) and oxygen vacancies ($V_O$): ($V_{app} > 0$) At positive voltages $V_O$ formation and migration occurs. The negatively charged oxygen migrates towards the interface and into the porous Pt electrode and the positively charged $V_O$ move along the electric field away from the Pt electrode and towards the grounded Ti electrode. ($V_{app} \leq 0$) At zero applied voltage, $V_O$'s move back towards the Pt, driven by the built-in electrochemical gradient, where they recombine with O. A negative voltage accelerates this process. c) Conductance change from low to high under the application of 100 SET pulses with an amplitude of 4V and a duration of 500 $\mu$s. d) Time-dependent conductance measurement (read out at 0.6V) when a series of voltage pulses with an amplitude of 2V, 2.5V, and 3V and a duration of 100$\mu$s are applied. The long-term conductance (red area) remains constant. The pulses induce short-term increases of the conductance with subsequent decay. e) Short-term conductance changes due to the last three voltage pulses of the measurement in d), indicated by the dotted rectangle. Only the conductance values during the read voltage are shown here. f) Aggregate plot showing short-term plasticity for different values of the long-term weight W. The measurement data was obtained by first applying the protocol in d) to characterize the short-term plasticity for the minimum long-term weight (W1). The long-term weight was then changed by 100 SET pulses (c) and, after a waiting period of 240s, the short-term plasticity was measured again.
• Figure 3: Control over the magnitude and dynamics of short-term conductance updates. a) Mean (solid line) and standard deviation (shaded area) of pulse-induced short-term conductance updates ($\Delta F$) from five conductance measurements and using four different pulse voltages (2, 2.5, 3, and 3.5V). The read voltage is set to 0.6V and the pulse width to 100 $\mu$s. To better compare the measurements, the conductance values were adjusted by subtracting the initial conductance at t=0 from the data. b) Heatmap of the achieved $\Delta F$ for the different pulse voltages and widths. c) Heatmap of the required pulse energy for the same voltage and width combinations as in b). d) Applied voltage protocol on a linear x-axis (top) and corresponding conductance values using 0.6V read pulses shown on a logarithmic x-axis (bottom). In between the read pulses a constant bias voltage ($V_{bias}$) of variable amplitude is applied. The main pulse voltage and width are set to 3.5V and 500 $\mu s$, respectively, in all measurements. The mean and standard deviation of the adjusted conductance values are shown for 5 measurements on a semi-log plot. e) Extracted decay time constant $\Lambda$ from the measurements in d) as a function of $V_{bias}$. The experimental data points were fitted with a sigmoid function.
• Figure 4: Simulation and energy consumption of an STPN network with multi-functional synapses. a) Sketch of the full STPN network. A frame of the Atari game is fed into two convolutional layers: Conv(kernel, stride) plus a ReLU activation function. The features are then flattened into a 1D array and fed into the m-STPN layer consisting of m-STPN units (blue ellipses) and the corresponding ST-Hebb synapses (blue lines). The layer's output is split into "actions" and a "value" by two fully connected linear layers. b) Average reward as a function of the agent steps during training for five different ranges of $\Lambda$. Each curve represents the average reward of 16 agents with different random parameter initialization. In the inset, the cases $\Lambda = 0$ and $\Lambda = [0.08, 0.92]$ (i.e., the achievable device range) are shown with their standard deviation (shaded area). c) Total synaptic weight (long- and short-term component) of a single synapse of the trained network ($S_{max\{\Delta F\}}$) during an entire game. d) Zoom-in of the dotted area in c). The long-term weight W is marked in red. The black bars represent the $\Delta F$ at each time step. e) Experimental values of the energy consumed by the voltage pulses required to change the short-term weight as a function of $\Delta F$. These values follow a power-law relationship. f) Calculated power consumption of a single synapse as a function of the total synaptic weight at different $V_{bias}$ used to vary the short-term decay constant. g) Time evolution of the estimated energy consumption of the synapse in c) during an entire Pong game (roughly 50 seconds) comparing a memristive synapse with a pure GPU implementation. The different components as well as the total energy consumption are shown. h) Histograms of all synapses in the network, indicating how many synapses consume a specific amount of energy during the whole Pong game. The two different components of the energy consumption ($\Delta F$ and Decay) are shown analogous to g). A worst case scenario of $V_{bias} = 0.6$ for all synapses is assumed. i) Total energy ($\Delta F$ plus Decay) consumed per synapse for the entire game.