Higher order neural processing with input-adaptive dynamic weights on MoS2 memtransistor crossbars

The increasing complexity of deep learning systems has pushed conventional computing technologies to their limits. While the memristor is one prevailing for acceleration, it only suited classical layers where two operands, namely weights and inputs, are processed simultaneously. Meanwhile, improve computational efficiency emerging applications, a variety non-traditional requiring concurrent processing many operands becoming popular. For example, hypernetworks predictive robustness by simultaneously inputs against application context. Two-electrode grids cannot directly map layers’ higher-order multiplicative neural interactions. Addressing this unmet need, we present crossbar using dual-gated memtransistors based on two-dimensional semiconductor MoS 2 . Unlike memristor, resistance states can be persistently programmed actively controlled multiple gate electrodes. Thus, discussed memtransistor enables several advanced inference architectures beyond passive crossbar. show that sneak paths effectively suppressed in crossbars, whereas they limit size scalability Similarly, exploiting terminals suppress dynamically reduces biasing power ∼20% crossbars fully connected layer AlexNet. On such as hypernetworks, collocating operations within same cells operating ∼15× considered network cases.