The dynamic processing of optoelectronic signals carrying temporal and sequential information is critical to various machine learning applications including language processing and computer vision. Despite extensive efforts to emulate the visual cortex of human brain, large energy/time overhead and extra hardware costs are incurred by the physically separated sensing, memory, and processing units. The challenge is further intensified by the tedious training of conventional recurrent neural networks for edge deployment. Here, we report in-sensor reservoir computing for language learning. High dimensionality, nonlinearity, and fading memory for the in-sensor reservoir were achieved via two-dimensional memristors based on tin sulfide (SnS), uniquely having dual-type defect states associated with Sn and S vacancies. Our in-sensor reservoir computing demonstrates an accuracy of 91% to classify short sentences of language, thus shedding light on a low training cost and the real-time solution for processing temporal and sequential signals for machine learning applications at the edge.

The large-scale crossbar array is a promising architecture for hardware-amenable energy efficient three-dimensional memory and neuromorphic computing systems. While accessing a memory cell with negligible sneak currents remains a fundamental issue in the crossbar array architecture, up-to-date memory cells for large-scale crossbar arrays suffer from process and device integration (one selector one resistor) or destructive read operation (complementary resistive switching). Here, we introduce a self-selective memory cell based on hexagonal boron nitride and graphene in a vertical heterostructure. Combining non-volatile and volatile memory operations in the two hexagonal boron nitride layers, we demonstrate a self-selectivity of 10 10 with an on/off resistance ratio larger than 10 3 . The graphene layer efficiently blocks the diffusion of volatile silver filaments to integrate the volatile and non-volatile kinetics in a novel way. Our self-selective memory minimizes sneak currents on large-scale memory operation, thereby achieving a practical readout margin for terabit-scale and energy-efficient memory integration.

Synaptic computation, which is vital for information processing and decision making in neural networks, has remained technically challenging to be demonstrated without using numerous transistors and capacitors, though significant efforts have been made to emulate the biological synaptic transmission such as short-term and long-term plasticity and memory. Here, we report synaptic computation based on Joule heating and versatile doping induced metal-insulator transition in a scalable monolayer-molybdenum disulfide (MoS) device with a biologically comparable energy consumption (∼10 fJ). A circuit with our tunable excitatory and inhibitory synaptic devices demonstrates a key function for realizing the most precise temporal computation in the human brain, sound localization: detecting an interaural time difference by suppressing sound intensity- or frequency-dependent synaptic connectivity. This Letter opens a way to implement synaptic computing in neuromorphic applications, overcoming the limitation of scalability and power consumption in conventional CMOS-based neuromorphic devices.