Designing Neuromorphic Processor with On-Chip Learning
2
조회
0
다운로드
4
인용
0
참고문헌
키워드
초록
In this paper, we present a neuromorphic processor that could learn to classify images through spiking information between neurons. Data is only computed locally, and there is no need for energy-hungry data transmissions that is required for any other machine learning algorithms. Careful algorithmic adaptations, along with novel hardware implementation of softmax functions are introduced to deliver maximum performance on image classification tasks while minimizing energy consumption. The design was fabricated in TSMC 65nm technology and consumes mere 23.6mW at a scaled voltage of 0.8V with 20MHz clock frequency, while showing a throughput of 73.9M pixels/second for supervised training on handwritten images
Abstract
In this paper, we present a neuromorphic processor that could learn to classify images through spiking information between neurons. Data is only computed locally, and there is no need for energy-hungry data transmissions that is required for any other machine learning algorithms. Careful algorithmic adaptations, along with novel hardware implementation of softmax functions are introduced to deliver maximum performance on image classification tasks while minimizing energy consumption. The design was fabricated in TSMC 65nm technology and consumes mere 23.6mW at a scaled voltage of 0.8V with 20MHz clock frequency, while showing a throughput of 73.9M pixels/second for supervised training on handwritten images
참고문헌 (16)
- Han, Song, "Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding", ~ (2015)
- P. A. Merolla, "A million spiking-neuron integrated circuit with a scalable communication network and interface", Science, 345(6197), pp.668~673 (2014)
- P. Knag, "A sparse coding neural network ASIC with on-chip learning for feature extraction and encoding", IEEE Journal of Solid-State Circuits, 50(4), pp.1070~1079 (2015)
- Guerguiev, Jordan, "Towards deep learning with segregated dendrites", eLife, 6 (2017)
- D.-H. Lee, "Difference target propagation", Joint European Conference on Machine Learning and Knowledge Discovery in Databases (2015), Springer
- J. K. Kim, "Efficient hardware architecture for sparse coding", IEEE Transactions on Signal Processing, 62(16), pp.4173~4186 (2014)
- S. Gonugondla, "A 42pJ", Decision 3.12TOPS, IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig. Tech. Papers, pp.2018[DOI]
- A. Amravati, "A 55nm time-domain mixed-signal neuromorphic accelerator with stochastic synapses and embedded reinforcement learning for autonomous micro-robots", IEEE Int. Solid-State Circuits Conf. (ISSCC) Dig. Tech. Papers, pp.124~125 (2018)
- C. Tsai, "A 41.3", 26,7pJ per neuron weight RBM processor supporting on-chip learning, IEEE Journal on Solid-State Circuits (JSSC), pp.2017, 52(10)[DOI]
- P. Whatmough, "A 28nm SoC with a 1.2GHz 568nJ", prediction sparse deep-neural-network engine with >0.1 timing error rate tolerance for IoT applications, 2017, 242~243
- F. Buhler, "A 3.43TOPS", W 48.9pJ, Classification 512 analog neuron sparse coding neural network with on-chip learning and classification in 40nm CMOS, pp.IEEE Int. Symp. VLSI Circuits Dig. Tech. Papers, 2017
- J. K. Kim, "A 640M pixel", s 3.65mW sparse event-driven neuromorphic object recognition processor with on-chip Learning, 2015, 61~62
- S. Esser, "Back propagation for energy-efficient neuromorphic computing", Advances in Neural Information Processing Systems (NIPS), ~ (2015)
- A. Vaswani, "Attention is all you need", ~ (2017)
- B. Yuan, "Efficient hardware architecture of softmax layer in deep neural network", Proc. IEEE International System-on-Chip Conference (SOCC) (2016)
- J. Park, "A 65-nm Neuromorphic Image Classification Processor with Energy-Efficient Training Through Direct Spike-Only Feedback", IEEE Journal of Solid-State Circuits, 55(1), pp.108~119 (2019)