What problem does this paper attempt to address?

The paper attempts to address the problem of achieving efficient, low-power neural network computation in edge computing devices. Specifically, the main challenges faced by the researchers include: 1. **Bottlenecks of traditional computing architecture**: In the traditional von Neumann computing architecture, data storage and computation are performed in separate units, resulting in slow computation speed, high energy consumption, and large stacking scale. These factors limit the processing capability of edge learning devices in complex scenarios. 2. **Demand for neuromorphic computing**: Inspired by the neural networks of the human brain, researchers aim to develop new computing devices that can simulate the functions of biological synapses to overcome existing computational bottlenecks. The human brain achieves powerful memory and high-speed computation through distributed and parallel computing, making artificial synapses a promising approach to achieve this goal. 3. **Integration and optimization of memristors**: Memristors, as devices that simulate the functions of biological synapses, have nonlinear resistance characteristics and can combine data storage and computation by adjusting conductance values. However, practical applications of memristors still face obstacles such as instability of ion migration, difficulty in high-precision programming, and IR drop issues in large-scale integration. To address these challenges, the researchers developed a fully integrated chip based on neuromorphic memristors that can achieve efficient object recognition while maintaining low power consumption in noisy environments. Specific technical innovations include: - **Stable memristor design**: Using a "TiN/HfOx/TaOy/TiN" structured memristor to improve conductance stability and programming success rate. - **Simulation of neuromorphic systems**: By simulating dendritic structures, the accuracy of object recognition in noisy environments is improved, and dynamic energy consumption is significantly reduced. - **Crossbar array memristors**: Utilizing a 2T2R configured crossbar array to solve the IR drop problem, achieving efficient analog computation and in-memory computation. - **STELLAR architecture**: Proposing a STELLAR architecture based on a periodic parallel conductance tuning scheme to achieve on-chip learning, reducing energy consumption and improving durability. - **Hybrid training framework**: Combining on-chip improved training and off-chip training to further enhance system accuracy. In summary, this paper aims to solve the problem of efficient, low-power neural network computation in edge computing devices in complex scenarios by developing a fully integrated memristor chip.

A Fully-Integrated Memristor Chip for Edge Learning

Edge learning using a fully integrated neuro-inspired memristor chip

Code for Edge Learning Using a Fully Integrated Neuro-Inspired Memristor Chip

Transforming edge hardware with in situ learning features

CMOS-integrated memristive non-volatile computing-in-memory for AI edge processors

Fully Memristive SNNs with Temporal Coding for Fast and Low-power Edge Computing

Fusion of memristor and digital compute-in-memory processing for energy-efficient edge computing

Fully memristive spiking-neuron learning framework and its applications on pattern recognition and edge detection

End-to-end Edge Detection on Self-rectifying Resistive Memory Array

A Brain-Inspired In-Memory Computing System for Neuronal Communication Via Memristive Circuits.

A fully integrated reprogrammable memristor–CMOS system for efficient multiply–accumulate operations

Fully Hardware Memristive Neuromorphic Computing Enabled by the Integration of Trainable Dendritic Neurons and High‐Density RRAM Chip

Hardware Implementation of Edge Neural Network Computing for Sensor With Memristors Based on Single-Crystalline LiNbO 3 Thin Film

In-Memory Computing With Memristor Arrays

Edge AI without Compromise: Efficient, Versatile and Accurate Neurocomputing in Resistive Random-Access Memory

Memristor-Based Signal Processing for Edge Computing

Memristive brain-like computing

In-memory and in-sensor reservoir computing with memristive devices

Powering AI at the edge: A robust, memristor-based binarized neural network with near-memory computing and miniaturized solar cell

Memristor‐Based Neuromorphic Chips

A compute-in-memory chip based on resistive random-access memory