14 October 2019 in IEEE Journal of Solid-State Circuits
Recent advances in neural network (NN) and machine learning algorithms have sparked a wide array of research in specialized hardware, ranging from high-performance NN accelerators for use inside the server systems to energy-efficient edge computing systems. While most of these studies have focused on designing inference engines, implementing the training process of an NN for energy-constrained mobile devices has remained to be a challenge due to the requirement of higher numerical precision. In this article, we aim to build an on-chip learning system that would show highly energy-efficient training for NNs without degradation in the performance for machine learning tasks. To achieve this goal, we adapt and optimize a neuromorphic learning algorithm and propose hardware design techniques to fully exploit the properties of the modifications. We verify that our system achieves energy-efficient training with only 7.5% more energy consumption compared with its highly efficient inference of 236 nJ/image on the handwritten digit [Modified National Institute of Standards and Technology database (MNIST)] images. Moreover, our system achieves 97.83% classification accuracy on the MNIST test data set, which outperforms prior neuromorphic on-chip learning systems and is close to the performance of the conventional method for training deep neural networks (NNs), the backpropagation.
1 March 2023 in IEEE Robotics and Automation Letters
Curriculum learning has proven highly effective to speed up training convergence with improved performance in a variety of tasks. Researchers have been studying how a curriculum can be constituted to train reinforcement learning (RL) agents in various application domains. However, discovering curriculum sequencing requires the ranking of sub-tasks or samples in order of difficulty, which is not yet sufficiently studied for robot navigation problems. It is still an open question what navigation strategies can be learned and transferred during multi-stage transfer learning from easy to hard. Furthermore, despite of some attempts of learning real robot manipulation tasks using curriculum, most of existing works are limited to toy or simulated settings rather than realistic scenarios. To address those issues, we first investigated how the model convergence in diverse environments relates to the navigation strategies and difficulty metrics. We found that only some of the environments can be trained from scratch, such as in a relatively open tunnel-like environment that only required wall following. We then carried out a two-stage transfer learning for more difficult environments. We found such approach effective for goal navigation, but failed for more complex tasks where movable obstacles may be on the navigation path. To facilitate more complex policies in the navigation among movable obstacles (NAMO) task, another curriculum with distance and pace functions appropriate to the difficulty of the environment was developed. The proposed scheme was proved effective and the strategies learned were discussed via comprehensive evaluations conducted in simulated and real environments.
1 November 2018 in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
This paper presents a power- and area-efficient spike sorting processor (SSP) for real-time neural recordings. The proposed SSP includes novel detection, feature extraction, and improved K-means algorithms for better clustering accuracy, online clustering performance, and lower power and smaller area per channel. Time-multiplexed registers are utilized in the detector for dynamic power reduction. Finally, an ultralow-voltage 8T static random access memory (SRAM) is developed to reduce area and leakage consumption when compared to D flip-flop-based memory. The proposed SSP, fabricated in 65-nm CMOS process technology, consumes only 0.175 μW/channel when processing 128 input channels at 3.2 MHz and 0.54 V, which is the lowest among the compared state-of-the-art SSPs. The proposed SSP also occupies 0.003 mm²/channel, which allows 333 channels/mm².
1 February 2019 in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
Recently, analog and mixed-signal neural network processors have been extensively studied due to their better energy efficiency and small footprint. However, analog computing is more vulnerable to circuit nonidealities such as process variation than their digital counterparts. On-chip calibration circuits can be adopted to measure and compensate for those effects, but it leads to unavoidable area and power overheads. In this brief, we propose a variation and noise-tolerant learning algorithm and postsilicon process variation compensation technique which does not require any additional monitoring circuitry. The proposed techniques reduce the accuracy degradation in the corrupted fully connected network down to 1% under large amount of variations including 10% unit capacitor mismatch, 8-mVrms comparator noise and 20-mVrms comparator offset.
20 February 2017 in IEEE Journal of Solid-State Circuits
This paper presents an energy-efficient face detection and recognition processor aimed at mobile applications. The algorithmic optimizations including hybrid search scheme for face detection significantly reduce computational complexity and architecture modification such as feature memory segmentation and further reduce energy consumption. We utilize characteristics of the implemented algorithm and propose a 5T SRAM design heavily optimized for mostly-read operations. Systematic reset and write schemes allow for reliable data write operation. The 5T SRAM reduces the cell area by 7.2% compared to a conventional 6T bit cell in logic rule while significantly improving read margin and voltage scalability due to a decoupled read path. The fabricated processor consumes only 23 mW while processing both face detection and recognition in real time at 5.5 frames/s throughput.
13 December 2021 in IEEE Transactions on Biomedical Circuits and Systems
This paper presents a spike sorting processor based on an accurate spike clustering algorithm. The proposed spike sorting algorithm employs an L2-normalized convolutional autoencoder to extract features from the input, where the autoencoder is trained using the proposed spike sorting-aware loss. In addition, we propose a similarity-based K-means clustering algorithm that conditionally updates the means by observing the cosine similarity. The modified K-means algorithm exhibits better convergence and enables online clustering with higher classification accuracy. We implement a spike sorting processor based on the proposed algorithm using an efficient time-multiplexed hardware architecture in a 40-nm CMOS process. Experimental results show that the processor consumes 224.75μW/mm 2 when processing 16 input channels at 7.68 MHz and 0.55 V. Our design achieves 95.54% clustering accuracy, outperforming prior spike sorting processor designs.