Integrated Artificial\nNeural Network with Trainable\nActivation Function Enabled by Topological Insulator-Based Spin–Orbit\nTorque Devices

Nonvolatile memristors offer a salient platform for artificial\nneural network (ANN), yet the integration of different function and\nalgorithm blocks into one hardware system remains challenging. Here\nwe demonstrate the brain-like synaptic (SOT-S) and neuronal (SOT-N)\nfunctions in the Bi<sub>2</sub>Te<sub>3</sub>/CrTe<sub>2</sub> heterostructure-based\nspin–orbit torque (SOT) device. The SOT-S unit exhibits highly\nlinear and symmetrical long-term potentiation/depression process,\nresulting in a fast-training of the MNIST data set with the classification\naccuracy above 90%. Meanwhile, the Sigmoid-shape transition curve\ninherited in the SOT-N cell replaces the software-based activation\nfunction block, hence reducing the system complexity. On this basis,\nwe employ a serial-connected, voltage-mode sensing ANN architecture\nto enhance the vector-matrix multiplication signal strength with low\nreading error of 0.61% while simplifying the peripheral circuitry.\nFurthermore, the trainable activation function of SOT-N enables the\nimplementation of the Batch Normalization algorithm and activation\noperation within one clock cycle, which bring about improved on/off-chip\ntraining performance close to the ideal baseline.