Synaptic plasticity in self-powered artificial striate cortex for binocular orientation selectivity

Classical computer vision still struggles with robust motion detection, object recognition, navigation, and activity recognition. In contrast, the biological visual system is a hierarchical pathway (retina → optic nerve → lateral geniculate nucleus (LGN) → striate cortex (V1)) where each stage processes visual information with distinct receptive fields. Retinal ganglion cells and LGN neurons exhibit antagonistic center-surround receptive fields, while striate cortical neurons possess narrow, elongated, orientation-selective receptive fields that are critical for edge and corner detection and underlie motion processing. These cortical receptive fields emerge via convergence from many lower-level inputs, implying that synaptic modulation along the pathway is key to developing orientation selectivity. Rate-based plasticity, particularly the Bienenstock-Cooper-Munro (BCM) learning rule with a sliding modification threshold, provides a more biorealistic account than purely Hebbian learning for experience-dependent cortical selectivity. Triplet-STDP can reproduce frequency effects and, in an All-to-All framework, generalize to BCM-like rate rules. Moreover, neurons in V1 are binocular and require interocular matching of orientation tuning for coherent depth perception; deprivation paradigms reveal experience-dependent modifications of orientation selectivity. While many retina-inspired in-sensor processing systems exist, hardware emulation of experience-dependent synaptic modifications to form binocular, orientation-selective receptive fields in striate cortex remains underexplored. This work reports an artificial striate cortex using a crossbar array of self-powered, monolithic all-perovskite memristors (each node integrates a CsFAPbI₃ solar cell with a CsPbBr₂I memristor), enabling optical-to-electrical transduction and plasticity governed by triplet-STDP/BCM for pattern learning and binocular orientation selectivity.

The paper situates its contribution within several bodies of work: (1) Visual neurophysiology establishing hierarchical receptive fields from retina through LGN to V1, and the orientation-selective, binocular nature of striate cortical neurons (Hubel & Wiesel; subsequent work on development and binocular matching). (2) Synaptic plasticity theories where rate-based BCM with a sliding threshold better matches experience-dependent cortical modifications than pure Hebbian rules; pair- and triplet-STDP frameworks have been shown to reproduce BCM-like behavior under certain conditions (Izhikevich & Desai; Pfister & Gerstner; Gjorgjieva et al.). (3) Neuromorphic devices and memristors emulating synaptic functions (PPF, LTP/LTD, STDP), including perovskite-based memristors and self-powered artificial retina concepts; however, most prior demonstrations focus on monocular, electrically driven synapses and lack binocular orientation selectivity and light-driven, experience-dependent BCM implementation. The authors aim to bridge these gaps by integrating a perovskite solar cell with a perovskite memristor to implement triplet-STDP-derived BCM under optical stimulation and emulate binocular orientation selectivity.

Device architecture and materials: - Memristor-only device: Au/P3HT/CsPbBr₂I/ITO. CsPbBr₂I (~100 nm) is the switching layer; ultrathin P3HT acts as a passivation/reservoir layer to reduce defects and retain halide ions; Au is TE, ITO is BE. - Self-powered memristor (monolithic all-perovskite stack): ITO/CsPbBr₂I/P3HT/Au/ITO/Au/Spiro-OMeTAD/CsFAPbI₃/SnO₂/ITO. A CsFAPbI₃ perovskite solar cell is directly stacked atop the CsPbBr₂I memristor, enabling optical-to-electrical conversion (V_OC ~1 V at 100 mW/cm²) to drive synaptic plasticity without external power. Materials preparation (key steps): - CsPbBr₂I precursor: PbI₂ (0.461 g), PbBr₂ (0.367 g), CsBr (0.425 g) in 4 mL DMSO:DMF (1:1), stirred 30 min at 25 °C, filtered. - P3HT precursor: P3HT in CHCl₃ (3 mg/mL), ultrasonicated 1 h, filtered. - CsFAPbI₃ precursor (1.5 M): FAI (245.05 mg), CsI (19.47 mg), PbI₂ (681.13 mg), PbBr₂ (8.53 mg), MACl (35 mg) in DMF:DMSO (4:1). - Spiro-OMeTAD solution: Spiro (72.3 mg), tBP (28.8 µL), Li-TFSI stock (17.5 µL of 520 mg/mL in acetonitrile) in 1 mL chlorobenzene. Device fabrication: - Memristor: Clean ITO; spin-coat CsPbBr₂I (500 rpm 5 s + 2000 rpm 30 s), anneal 100 °C 10 min; spin-coat P3HT (500 rpm 5 s + 3000 rpm 10 s), anneal 100 °C 10 min; evaporate Au TE through shadow mask. - Solar cell: Spin-coat SnO₂ (3.75 wt%) on ITO at 3000 rpm 30 s; anneal 150 °C 30 min; spin-coat CsFAPbI₃, anneal 150 °C 30 min; spin-coat filtered Spiro-OMeTAD (3000 rpm 30 s); evaporate Au cathode. - Integration: Sputter ~600 nm ITO onto the solar cell Au cathode (acts as robust interlayer), evaporate Au, spin-coat ultrathin P3HT reservoir, spin-coat CsPbBr₂I switching layer, deposit ITO by vacuum evaporation to complete the self-powered memristor stack. Layer continuity verified by cross-sectional SEM; optical/electrical characterizations confirm functionality. Electrical/optical characterization: - Memristor I–V with B1500A; EPSC/PPF/LTP/LTD and STDP via voltage pulses; optical generation via solar cell (AM1.5G, 100 mW/cm²) with optical chopper to modulate frequency; J–V of solar cell via Keithley 2400; PL, TRPL, UV–Vis, EIS, TPC/TPV, XRD, XPS, AFM, SEM per standard instrumentation. Synaptic function emulation: - Second-order dynamics assessed via EPSC decay and hysteresis; PPF quantified with two 0.5 V, 5 ms pulses varying inter-spike interval; fit PPF = c₁e^{−t/τ₁} + c₂e^{−t/τ₂}. - LTP/LTD: 100 potentiating pulses (0.5 V/5 ms) followed by 100 depressing pulses (−0.5 V/5 ms); study dependence on pulse interval and width; evaluate conductance window and nonlinearity. - History-dependent plasticity: sequential pulse trains (0.5 V/5 ms) at 100 Hz → 10 Hz → 1 Hz → 10 Hz to probe metaplastic effects. - Pair-STDP: apply pre/post pulses (0.5 V, 5 ms) with variable Δt = t_post − t_pre; evaluate dependence on initial conductance G₀ (100, 200, 300 µs) and Δt sign/magnitude. Triplet-STDP and BCM implementation (self-powered mode): - Circuit: Solar cell provides pre-synaptic optical spikes (V_OC ~1 V) to memristor TE; BE receives post-synaptic pulses; triplet protocols include post-pre-post and pre-post-pre with varying Δt₁, Δt₂ to capture paired and triplet terms. - Map ΔG across quadrants (combinations of Δt₁, Δt₂), showing asymmetry and first-spike-dominant suppression; fit parameters support quantitative BCM derivation. BCM learning and pattern recognition on array: - Fabricate flexible 5×5 crossbar array; use a 3×3 subset (nine solar cells/memristors) for demonstration. All electrodes routed to an STM32 controller on PCB for real-time control. - Learning loop: Optical Poisson input at rate p_x drives solar cells; STM32 measures postsynaptic current I = ∫_{t−200ms}^{t} w_o G dt (w_o=0.5 V), computes feedback post firing p_y = g I (g = 50 Hz/mA), effectively p_y ≈ G p_x for linear neuron; update synapses per generalized BCM derived from triplet-STDP: dG/dt = φ(p_y, θ) p_x with sliding threshold θ dependent on postsynaptic activity/history (tuned via G₀). - Pattern encoding: High-rate (30 Hz) for pattern pixels vs low-rate (14 Hz) for background; monitor weight evolution until pattern emerges. Binocular orientation selectivity simulation: - 2-layer SNN: two 9×9 input layers (left/right eyes) and one output neuron; inputs are oriented bars (0°, 45°, 90°, 135°) with high (30 Hz) vs low (14 Hz) rates per pixel; initialize low random weights; in each epoch, randomly present orientations to both eyes with equal probability; compute p_y = G_l p_{xl} + G_r p_{xr}; update G and θ via BCM; winner orientation is the one driving p_y above θ and remains stable through competitive dynamics. - Conditions: (1) normal binocular contour vision, (2) monocular deprivation (one eye noise 4–6 Hz), (3) binocular deprivation (both eyes noise). Track synaptic maps, p_y, and θ over time.

- Monolithic self-powered memristor: First demonstration of a solution-processed, monolithic all-perovskite self-powered memristor integrating a CsFAPbI₃ solar cell directly stacked on a CsPbBr₂I memristor, enabling optical stimulation without external bias and strong biomimic functional congruence (retina to cortex pathway). - Second-order synaptic dynamics: Devices exhibit pinched hysteresis with gradual conductance modulation; EPSC decays to baseline in ~200 ms; PPF fits yield τ₁ ≈ 0.28 ms and τ₂ ≈ 10.86 ms, comparable to biological synapses. - Long-term plasticity: Robust LTP/LTD with 100 pulses (±0.5 V, 5 ms); conductance window (Gmax−Gmin) and nonlinearity (NL) of LTP/LTD reported as 3.00/3.98; stronger LTP for shorter inter-pulse intervals and longer pulse widths. - Metaplasticity/history dependence: Identical stimulus (10 Hz) can induce opposite conductance trends depending on prior activity (preceded by 100 Hz vs 1 Hz), evidencing history-dependent plasticity. - Pair-STDP with state dependence: Sign and magnitude of ΔG depend on Δt and initial conductance G₀ (100, 200, 300 µs), including decay-dominated regimes where LTD occurs even for Δt > 0. - Triplet-STDP under optical driving: Full mapping of ΔG across Δt₁, Δt₂ shows asymmetric first-spike-dominant behavior; quadrant-specific transitions between LTP/LTD; additional triplet configurations reduce to pure LTP (quadrant I) or LTD (quadrant III). Fitted parameters enable quantitative BCM realization. - Generalized BCM implementation: Using triplet-STDP-derived terms within an All-to-All framework, achieved BCM with sliding threshold; threshold θ tunable via history state G₀, sliding from ~40 Hz (G₀=100 µs) to ~60 Hz (200 µs) and ~80 Hz (300 µs); depression at low p_y and potentiation at high p_y confirmed. - Pattern learning on array: With a flexible 5×5 array (3×3 subset used), optical-encoded ‘X’ pattern (30 Hz vs 14 Hz) learned via BCM; synapses corresponding to pattern pixels potentiated (1,3,5,7,9), background depressed, yielding a clear ‘X’ mapping. - Binocular orientation selectivity (simulation): In normal binocular input, the network randomly selects a winner orientation (vertical in the reported run) with p^0 rising above θ (~64 Hz) while others remain below; under monocular deprivation (right eye noise 4–6 Hz), the intact eye selects horizontal with lower winner rate (~54 Hz) and fluctuating θ due to noisy binocular input; under binocular deprivation, orientation selectivity is lost. These outcomes align with experimental observations in kitten V1. - System-level implications: Crossbar-compatible, low-power operation via self-powered optical input and two-terminal memristive synapses supports scalable machine vision with embedded edge/corner detection capabilities.

The study addresses the challenge of emulating experience-dependent cortical receptive field formation by implementing a hardware platform that maps optical input (retinal function) to synaptic plasticity (cortical function). By demonstrating triplet-STDP under optical stimulation and deriving a generalized BCM rule with a sliding threshold controllable via synaptic history, the work reproduces key mechanisms believed to shape orientation selectivity and binocular matching in V1. Device-level second-order ionic dynamics (halide vacancies) provide the biophysical substrate for rate- and history-dependent plasticity, enabling both short-term (PPF) and long-term (LTP/LTD) behaviors. The array-level demonstration validates that rate-based learning can perform optical pattern mapping, while simulations extrapolate to binocular orientation selectivity under different rearing conditions (normal, monocular deprivation, binocular deprivation), consistent with neurophysiology. The alignment between device physics, synaptic learning rules, and visual processing supports the relevance of self-powered perovskite memristor arrays as a platform for neuromorphic vision, offering a pathway to integrated, low-power, and scalable edge/corner detection and motion-processing front-ends.

This work introduces a monolithic all-perovskite, self-powered memristor array that emulates a striate cortex with binocular and orientation-selective receptive fields. Key contributions include: (1) integration of a CsFAPbI₃ solar cell with a CsPbBr₂I memristor to form a self-powered synapse; (2) demonstration of optical triplet-STDP and derivation of a generalized BCM learning rule with a tunable sliding threshold; (3) array-level optical pattern learning; and (4) simulation of binocular orientation selectivity under normal and deprivation conditions, reproducing known neurophysiological phenomena. The crossbar-compatible, two-terminal architecture and material homogeneity support high-density, low-power machine vision systems. Future work could focus on scaling array size and density, on-chip implementation of binocular SNNs with hardware inhibition/competition, real-time dynamic scene processing, robustness across spectra and illumination levels, and closing the loop from hardware to behavior via end-to-end sensor-compute systems.

- Scale of hardware demonstrations: Orientation selectivity and binocular interactions were validated via simulations grounded in device data; the large-scale 9×9 binocular networks were not realized in hardware in this study. - Array utilization: Although a flexible 5×5 array was fabricated, pattern learning used a 3×3 subset; broader utilization and larger arrays will be needed to confirm scalability and mapping fidelity. - Sliding-threshold assumption: The BCM implementation assumes active pre- and postsynaptic firing for pathway activation; performance under sparse or asynchronous conditions may vary. - Device variability and stability: While perovskite devices enable rich plasticity, long-term stability, variability across devices, and environmental sensitivity (e.g., moisture, temperature) can impact consistency and durability. - Optical input constraints: The reported array size and electrode geometries are limited by sub-solar-cell dimensions; optical intensity and spectrum dependence could affect performance in diverse environments.