Classic computer vision systems still struggle with robust solutions for tasks like motion detection, object recognition, and navigation. The biological visual system, a hierarchical structure encompassing the retina, optic nerve, lateral geniculate nucleus (LGN), and striate cortex, offers insights. Understanding each level's role in processing visual information can help address the challenges faced by computer vision. Receptive fields, areas in the retina influencing the firing rate of corresponding units, vary across the visual pathway. Ganglion cells have concentric receptive fields with antagonistic "on" and "off" regions. The LGN possesses similar fields. However, striate cortex cells have narrow, elongated, vertically oriented receptive fields crucial for edge and corner detection, which underlies motion detection. The convergence process along the visual pathway, from photosensory cells to ganglion cells and finally to striate cortical cells, shapes the receptive field's development. Modulating synaptic connections between these levels is critical for developing the narrow, slit-shaped, orientation-selective receptive fields in the striate cortex. Compared to purely Hebbian learning, the rate-based Bienenstock-Cooper-Munro (BCM) learning rule offers a more biorealistic model for synaptic modification and neuronal response selectivity in the striate cortex. BCM describes how synaptic weight modification is determined by whether postsynaptic response exceeds a threshold. A sliding threshold, dependent on average postsynaptic neuron activity, ensures a history-dependent characteristic. Triplet-STDP, introducing a third pre- or post-synaptic spike to pair-STDP, reproduces the frequency effects of the pair protocol and can realize the rate-based BCM learning rule. Striate cortical neurons also exhibit binocularity, receiving signals from both eyes. In biological systems, interocular neurons initially have different orientation preferences, requiring a matching process to form normal binocular perception. This research aims to emulate the experience-dependent modifications of synaptic strength to create a binocular, orientation-selective receptive field similar to the striate cortex.

The existing literature extensively covers the hierarchical structure of the biological visual system and the characteristics of receptive fields at different levels, from the retina to the striate cortex. Studies highlight the antagonistic nature of "on" and "off" regions in ganglion cell receptive fields and the elongated, orientation-selective receptive fields of striate cortex neurons. The BCM learning rule has been established as a more biologically plausible model for synaptic plasticity compared to purely Hebbian learning, accurately capturing the sliding threshold and history-dependent nature of synaptic weight modification. Triplet-STDP has been demonstrated as a mechanism for achieving frequency effects in synaptic plasticity, aligning with the observed behavior in the striate cortex. The importance of binocularity and the developmental process of orientation selectivity in the striate cortex have also been widely documented, emphasizing the need for a model that incorporates both features. However, the hardware implementation of a system that emulates the experience-dependent development of binocular, orientation-selective receptive fields lags behind.

This study employed a self-powered memristor crossbar array as the foundation for emulating the striate cortex. Each cross-point in the array consists of a CsFAPbI3 perovskite solar cell acting as a photosensory neuron and a CsPbBr2I perovskite memristor functioning as a striate cortical synapse. The memristor's second-order dynamics, stemming from mobile halogen vacancies, mimic rate-based plasticity. The solar cell converts optical signals into electrical signals driving the memristor. The bottom electrode of the memristor represents the post-cortical neuron, and conductance changes simulate synaptic weight modulation. The fabrication process involved solution processing to create a monolithic all-perovskite system, ensuring facile integration of memristor and solar cell. The CsPbBr2I memristor's second-order effects were verified by demonstrating its ability to exhibit paired-pulse facilitation (PPF), long-term potentiation (LTP), and long-term depression (LTD) behaviors. The history-dependent plasticity of the memristor was also confirmed through experiments demonstrating conductance modulation based on previous activity. To realize the BCM learning rule, triplet-STDP was implemented using light modulation of the self-powered memristor. A 3x3 crossbar array was used to demonstrate pattern learning based on BCM, showing successful synapse depression/potentiation at low/high postsynaptic firing rates and a history-dependent sliding threshold. For the artificial striate cortex emulation, a simulation based on two 9x9 self-powered memristor networks was conducted using the generalized BCM learning rule. Three conditions were simulated: normal binocular contour vision, monocular deprivation, and binocular deprivation. The results were compared with experimental findings of kitten striate cortex development.

The research successfully demonstrated a self-powered artificial striate cortex using a monolithic all-perovskite system. This system utilizes a crossbar array where each crosspoint comprises a CsFAPbI3 perovskite solar cell (photosensory neuron) and a CsPbBr2I perovskite memristor (cortical synapse). The CsPbBr2I memristor exhibited second-order dynamics, accurately mimicking rate-based synaptic plasticity. Experiments confirmed the memristor's capacity for paired-pulse facilitation (PPF), long-term potentiation (LTP), and long-term depression (LTD). The implementation of triplet-STDP under optical stimuli showed an asymmetry beneficial for realizing the BCM learning rule. Using a 3x3 crossbar array, the key characteristics of BCM learning—synapse depression/potentiation at low/high postsynaptic firing rates and a sliding threshold—were successfully demonstrated. A pattern recognition task ('X' pattern) was successfully achieved with the self-powered memristor array. Simulations using two 9x9 memristor networks, emulating a striate cortex, reproduced the experience-dependent modifications observed in kitten striate cortex under three conditions: normal binocular vision, monocular deprivation, and binocular deprivation. The results closely matched experimental observations, validating the model's accuracy. The two-terminal structure of the self-powered memristor, based on the monolithic all-perovskite system, makes the bio-inspired striate cortex scalable for high-density, low-power consumption machine vision applications.

The successful emulation of the striate cortex's functionality using a self-powered memristor array represents a significant advancement in neuromorphic computing. The close alignment of the simulation results with experimental findings validates the model's biological plausibility and its potential for accurately capturing the complex dynamics of synaptic plasticity and receptive field development in the visual cortex. The use of a monolithic all-perovskite system offers advantages in terms of scalability and energy efficiency, opening possibilities for high-density, low-power machine vision applications. The findings demonstrate the potential of this bio-inspired approach to surpass the limitations of traditional computer vision systems by leveraging the inherent efficiency and adaptability of biological neural networks.

This study successfully demonstrated the first hardware implementation of an artificial striate cortex with binocularity and orientation selectivity, using a self-powered memristor array. The system accurately emulated key aspects of biological visual processing, including triplet-STDP, BCM learning, and experience-dependent plasticity. The use of a monolithic all-perovskite system makes it highly scalable and energy-efficient, promising advancements in high-density machine vision. Future work could focus on expanding the array size, investigating different learning rules, and exploring applications in real-world vision tasks.

The current study used simulations for the large-scale (9x9) artificial striate cortex due to the constraints of fabricating such a large array. While the 3x3 array demonstrated successful BCM learning, the 9x9 results are based on simulations parameterized by the smaller array experiments. Further research is needed to validate the performance of larger arrays. Another limitation is the simplified model of the striate cortex used in the simulations; future work should explore more complex models incorporating additional layers and neural interactions. While this study focused on visual processing, the underlying principles of plasticity and BCM learning may be applicable to other brain regions. Exploring the versatility of the developed self-powered memristor array in emulating other neural circuits is a promising area for future research.