Brain-Inspired In-Memory Data Pruning and Computing with TaO_x Mem-Selectors

The selective attention mechanisms inherent in the human visual system provide a promising framework for developing edge systems that can simultaneously prune and process critical information from visual input. However, conventional complementary metal-oxide-semiconductor-based edge vision systems rely on complex digital logic for data pruning, alongside the physical separation of pruning, memory, and processing. This increases both power consumption and latency. Herein, a Mem-Selector (M-S) device that features reconfigurable non-volatile resistive memory and volatile threshold switching in a Ta/TaO_x/Ta₂O₅ stack is presented. For the first time, using transmission electron microscopy, the formation and rupture of conductive oxygen vacancy filaments are observed when the device operates as a resistive memory, as well as the growth of Ta-rich nanocrystalline clusters when it switches to threshold mode. This suggests the coexistence of ionic and electronic switching mechanisms. By leveraging a multifunctional M-S device, an in-memory pruning-computing (IMPC) system that simultaneously prunes and processes information is constructed. The IMPC system, inspired by human visual-selective attention, the IMPC system adaptively extracts essential information while pruning trivial inputs based on task complexity. This approach optimizes the balance between hardware cost and classification performance. Compared to conventional in-memory computing systems, the integrated IMPC system reduces input energy consumption by 29%, 54%, and 90% with less than 1% accuracy loss. Additionally, it shows robustness improvements of 7.6%, 29.8%, and 80.7% on the CIFAR-10, FashionMNIST, and MNIST datasets, respectively. This demonstrates the potential of hardware-software co-design for energy-efficient, high-performance edge hardware.