With the accumulation of vast amounts of data, deep learning has advanced rapidly. Deep learning involves two key processes: the inference process, which employs matrix-matrix multiply-accumulate (MAC) operations, and the training process, which utili...
With the accumulation of vast amounts of data, deep learning has advanced rapidly. Deep learning involves two key processes: the inference process, which employs matrix-matrix multiply-accumulate (MAC) operations, and the training process, which utilizes outer product operations. Performing these processes on von Neumann architecture-based digital systems requires computation and memory access proportional to the matrix size. However, in the field of analog in-memory computing (AiMC), memristor-based crossbar arrays have been proposed as a novel solution. These arrays leverage Kirchhoff's laws to perform MAC operations in a single step and enable outer product operations by stochastically generating pulses and updating the conductance of memristors at the crosspoints in one operation. While the advantages of AiMC have driven extensive research, the non-ideal programming characteristics of analog devices have led to a focus on accelerating inference by mapping pre-trained parameters from the cloud onto analog device arrays. However, such cloud-based training approaches pose challenges related to latency and privacy, making the development of on-chip training methodologies and memory devices essential.
In the first part of this research, a 6T1C device based on six InGaZnO (IGZO) thin-film transistors (TFTs) and a capacitor was designed to meet the required specifications for on-chip training, and fabricated. After fabrication, measurements were conducted to demonstrate that the 6T1C device met the required specifications for on-chip training. A 16×16 6T1C crossbar array combined with an FPGA successfully demonstrated on-chip training with over 97% accuracy on the MNIST dataset. To validate on-chip training in the 6T1C array, detailed electrical characterizations were conducted, particularly focusing on temporal variations in ADC outputs during the training process. These measurements confirmed the feasibility of dynamic weight updates in the fabricated devices.
The second part of this study addresses the scalability limitations inherent to the 6T1C architecture, particularly when attempting to increase integration density by reducing the size of capacitor. Such reduction introduces severe retention degradation and heightened susceptibility to disturbance effects. To alleviate these issues, novel operation schemes were devised to suppress the impact of disturbances. To harness the intrinsic non-idealities, retention decay and disturbance, as a form of hardware-embedded regularization, on-chip training architecture separating the inference path and training path was adopted. In addition, the excessive regularization effect induced by disturbance in convolutional neural network (CNN) was analyzed. To mitigate this, pulse scheduling method and selective update algorithm were developed, leading to robust training. The proposed system successfully trained a ResNet18 model on the CIFAR-10 dataset to a software-equivalent accuracy level and achieved 100% classification accuracy for checkerboard pattern recognition in on-chip training demonstration.
In the final part of this work, an energy-efficient synaptic circuit tailored for on-chip training was proposed and fabricated. The design adopts a charge-sharing-based voltage sensing scheme, which inherently enables intrinsic regularization and facilitates on-chip training. Furthermore, a 1.5-bit ADC was introduced to realize a deterministic weight transfer mechanism, replacing the high-resolution ADC typically required in architectures that separate inference and training paths. This proposed approach significantly reduced the number of update pulses by avoiding unnecessary weight update in NVM array, thereby enhancing energy efficiency and prolonging the endurance of NVM devices. The architecture demonstrated stable and effective training behavior, offering a viable alternative to conventional designs with improved power and reliability characteristics.