Recent advances in the field of artificial intelligence (AI) have brought significant changes across various fields, establishing a new technological standard. However, this progress demands considerable computational and energy resources, thereby acc...
Recent advances in the field of artificial intelligence (AI) have brought significant changes across various fields, establishing a new technological standard. However, this progress demands considerable computational and energy resources, thereby accelerating research into neuromorphic systems as a promising approach for resource-efficient AI. In this regard, hardware neural networks (HNNs) that perform vector-matrix multiplication (VMM) within synapse arrays have attracted substantial attention. In particular, on-chip trainable HNNs are regarded as an effective method of compensating for performance degradation caused by synaptic device non-idealities, yet they remain relatively underexplored.
This dissertation presents a comprehensive framework for implementing on-chip trainable HNNs based on a NOR-type dual-gate field-effect transistor (DG FET) array. The proposed architecture achieves low-power operation through selective weight updates enabled by Fowler–Nordheim (FN) tunneling. Furthermore, its bidirectional VMM capability facilitates on-chip training without requiring a separate array for backpropagation, thereby simplifying the overall system architecture.
To further address the challenges of weight initialization and efficient learning in hardware, a novel weight transfer scheme based on the outer product form of Singular Value Decomposition (SVD) is proposed. This scheme is inherently compatible with the gradient descent-based fine-tuning algorithm, significantly reducing the latency of both the initial weight transfer and the subsequent training. Simulation results on Fashion-MNIST classification tasks validate the effectiveness of the proposed approach, demonstrating classification accuracy comparable to software-based networks while compensating for device-level non-idealities. Furthermore, the proposed system achieves more than 8.1 and 64 times improvement in energy efficiency and latency over systems based on conventional NOR-type flash arrays. These results demonstrate the potential of the NOR-type DG FET array as a promising solution for next-generation low-power, on-chip trainable HNNs.