US20260087340
2026-03-26
Physics
G06N3/065
The disclosed methods and apparatus integrate sensory transduction with neuromorphic computation to address challenges in conventional event-based vision systems, such as energy inefficiency, latency, and complexity. This integration occurs through a novel neuro-transducer cell, which combines sensing and computation. Each cell includes a photonic transducer, a membrane capacitor, and a Leaky Integrate-and-Fire (LIF) neuron, which processes raw physical stimuli directly. Adjacent cells are interconnected via programmable RRAM synapses, which store multi-bit weights and are updated using Spike-Timing-Dependent Plasticity (STDP) pulses.
Neuromorphic computing mimics the neural and synaptic structures of the human brain to process information, employing spiking neural networks (SNNs) that use spiking neurons and synapses. Unlike traditional Artificial Neural Networks (ANNs), which process dense numeric matrices, SNNs operate asynchronously and fire only during discrete spikes, encoding temporal and intensity information. This approach provides higher information density and efficiency. Neuromorphic chips, typically based on CMOS technology, support autonomous operation and continuous adaptation, but face challenges such as inefficient sensing and data transfer bottlenecks.
Current neuromorphic systems encounter inefficiencies in sensing and data processing. Conventional sensors, like CMOS cameras, generate large amounts of redundant data, leading to significant power consumption and latency. The need to convert dense data into sparse spike trains for neuromorphic processing further exacerbates energy and time inefficiencies. These limitations hinder the practical application of neuromorphic computing in real-time, low-power AI systems, as they require more than 30 milliseconds to sense and recognize real-world events.
The proposed monolithic device fuses sensory transduction and neuromorphic computation into a single unit, termed a neuro-transducer. This device directly converts physical stimuli into computationally rich spike patterns, serving as the initial analytical layer of the network. Unlike traditional systems that require multiple processing steps, this approach allows for ultra-low power consumption and instantaneous perception. It is particularly beneficial for applications requiring long battery life, rapid response times, and edge-native features, such as wearables, robotics, and privacy-sensitive environments.
The neuromorphic computing performer circuitry includes several components: optics/mechanical coupling, a neuro-transducer tile array, calibration DACs, a local synaptic array, AER encoders/spike bus, a global arbiter/PCIe-AER bridge, and a downstream ML accelerator/MCU. This circuitry can be instantiated by various programmable devices, such as CPUs, FPGAs, and ASICs, to perform the required operations. The integration of these components enables efficient fusion of sensing and computation, promoting advancements in neuromorphic applications.