Programmable ferroelectric bionic vision hardware with selective attention for high-precision image classification

The human visual system efficiently processes vast amounts of visual information through selective attention, focusing on salient regions while suppressing irrelevant ones. This mechanism, mediated by complex neural networks in the visual cortex, allows for efficient information extraction in crowded visual scenes. Current hardware implementations of selective attention, primarily based on CMOS and conventional transistors, suffer from large footprints and high computational costs. The separation of sensory units from processing systems adds to the complexity. Integrating neuromorphic functions like visual recognition and light adaptation into compact optical sensing devices has shown promise, but these typically lack selective attention. High-efficiency hardware for selective attention is crucial for overcoming insufficient computing power during parallel processing of sensory data. P(VDF-TrFE), a ferroelectric material with tunable remnant polarization, offers potential for increasing memory states in memory cells through charge accumulation and depletion via intermediate polarization states, making it suitable for artificial visual perception systems. This study aims to address this challenge by fabricating a programmable ferroelectric bionic vision hardware with selective attention.

Existing research on hardware implementation of selective attention has largely relied on CMOS and conventional transistor technologies, leading to bulky and computationally expensive systems. While progress has been made in integrating neuromorphic functions like visual recognition and light adaptation into compact optical sensing devices, the crucial element of selective attention has remained largely unaddressed in hardware. Studies have explored various materials and architectures for artificial synapses and neuromorphic computing, but a fully integrated system with efficient selective attention capabilities has been lacking. This paper builds upon the existing body of work on ferroelectric materials and their application in memory devices, extending their functionality towards bio-inspired vision systems. The authors cite various research papers that have focused on artificial synapses, memristors, and organic transistors for neuromorphic computing, highlighting the limitations of these approaches in terms of size, power consumption, and integration of selective attention.

The researchers fabricated a programmable ferroelectric bionic vision hardware (FeBVH) using quantum dots (QDs) and a ferroelectric material, P(VDF-TrFE). The device utilizes a bottom-gate top-contact structure with PDVT-10 as the semiconductor layer and CdSe/ZnS QDs for light sensing. The ferroelectric layer allows for modulation of the energy barrier, controlling the tunneling effect of photogenerated carriers. This enables programmable photonic memory strength. A 5x5 array of these ferroelectric transistors was created to enable selective recording and suppression of UV light information. The polarization direction of the P(VDF-TrFE) layer plays a crucial role in this selective attention mechanism. The device's performance was characterized using various techniques including measuring excited postsynaptic current (EPSC), surface potential using Kelvin probe force microscopy (KPFM), absorption spectra, and transfer curves. The memory factor (ηM), defined as (Ip-Io)/Io x 100%, was used to quantify memory strength. Paired pulse facilitation (PPF) was analyzed to evaluate short-term synaptic plasticity. The signal-to-noise ratio (SNR) was used to assess the efficacy of selective attention in the 5x5 array. Finally, a simple neural network was implemented using the array to perform image classification, evaluating the accuracy of pattern recognition with and without selective attention. The authors also discuss the energy levels involved in the hole generation and tunneling process, supported by literature values.

The study demonstrated that the memory strength of the FeBVH could be modulated from 9.1% to 47.1% by varying the energy barrier using the ferroelectric layer. The device under positive polarization exhibited wavelength-dependent photoresponsivity, selectively responding to shorter wavelengths (UV). The 5x5 array effectively demonstrated selective attention, enabling the clear discrimination of a UV pattern ('E') under positive polarization while suppressing it under negative polarization. Significant differences in signal-to-noise ratio (SNR) were observed between positive and negative polarizations, with positive polarization exhibiting much higher SNR, demonstrating the selective retention and oblivion functions. Importantly, utilizing the selective attention mechanism, the fabricated ferroelectric sensory network achieved a high accuracy of 95.7% in classifying patterns of letters with different wavelengths, significantly outperforming the 69.7% accuracy without selective attention. The device showed long-term plasticity (LTP) under positive polarization and short-term plasticity (STP) under negative polarization, highlighting the tunable synaptic behavior.

The findings demonstrate the successful fabrication and characterization of a programmable ferroelectric bionic vision hardware capable of mimicking selective attention. The use of a ferroelectric material enables the creation of a compact device with multiple nonvolatile memory states, eliminating the need for separate memory units. The wavelength-dependent response under positive polarization, coupled with the high accuracy achieved in image classification, validates the efficacy of the proposed approach. The enhanced SNR under positive polarization directly supports the concept of selective attention, where salient information is retained, and non-salient information is suppressed. The significant improvement in classification accuracy with the implementation of selective attention emphasizes the importance of this biological function in efficient information processing. These findings have significant implications for the development of energy-efficient, high-performance neuromorphic systems.

This research successfully demonstrated a programmable ferroelectric bionic vision hardware capable of selective attention. The device’s tunable photoresponse, low hardware overhead, and high classification accuracy showcase its potential for advanced neuromorphic systems. Future research could focus on expanding the array size, exploring different ferroelectric materials and QD types, and integrating more complex neural network architectures for more sophisticated image processing tasks. Exploring applications beyond image classification, such as object recognition and tracking, would further demonstrate its capabilities.

The current study focused on UV light detection and classification of simple patterns. The generalizability of the findings to other wavelengths and more complex scenes needs further investigation. The size of the array (5x5) is relatively small, limiting its applicability to high-resolution images. Further work is required to scale up the array size while maintaining performance. The chosen neural network is relatively simple; more advanced architectures could potentially further improve classification accuracy.