2015-now, Ph.D. candidate, EE, Princeton University, USA.
- Efficient inference for deep neural networks
- Latency driven deep neural network compression
- Machine learning ensembles
- Thesis advisor: Prof. Niraj K. Jha
2011-15, B.Eng. (highest honor), EEE, Nanyang Technological University, Singapore.
- GPA: 3.94/4.00, Minor in Business: GPA 4.00/4.00
- Dean’s list for all 4 academic years
- NTU President Scholar, with Distinction (3-year)
- Honorable University Best Industrial-Orientation Prize
- Gold Medal: Defence Science & Technology Agency Gold Medal (Best Final Year Project / Thesis)
- Gold Medal: Thomson Asia Pacific Holdings Gold Medal (Best RF Circuits)
05-09/2018 Research Intern, Machine Learning Group, Alibaba Group, Sunnyvale, CA, USA.
Hardware-aware low latency inference for deep neural networks
Significantly reduce the inference latency
Compress the model sizes by one order of magnitude
Improve the accuracy
CNN compression and architecture learning
- Train both weights and network architecture based on the gradient information
- Enable effective gradient descent in the architecture space
- Achieve 15.7x /4.6x reduction for AlexNet params./FLOPs
- Achieve 30.2x /8.6x reduction for VGG-16 params./FLOPs
CNN platform-aware model adaptation
- Predictive-model based algorithm that cuts down search cost to CPU minutes
- Beat hand-crafted and NAS models through exploiting the potential of existing building blocks
- Achieve absolute accuracy gain over MobileNetV2
- Achieve absolute accuracy gain over MnasNet
Execution-efficient LSTM synthesis
- Propose hidden-layer LSTM cells with enhanced control gates
- Grow and prune the recurrent model for extra compactness against pruning-only methods
- 38.7x/45.5x reduction for NeuralTalk params./FLOPs, 4.5x faster, +2.6 CIDEr score
- 19.4x/23.5x reduction for DeepSpeech2 params./FLOPs, 1.2x faster, +4.2% accuracy
Deployable machine learning ensembles
- Exploit ensemble methods for disease diagnosis using wearable sensors: 77.4~99.3% diagnostic accuracies for 6 diseases in 4 disease categories
- Propose efficient ensemble hierarchy for IoT: 3.2-60x transmission load reduction
05-09/2015 Project Manager, Pattern Discovery Technologies Inc., Singapore.
- Machine-learning-based trend forecasting and vision for Port of Singapore Authority (PSA).
05-09/2014 Research Intern, Temasek Laboratories, Singapore.
FPGA-based real-time signal blind reconstruction
- Multiband signal blind reconstruction based on multi-coset theory
- 36-50% latency reduction
- Paper publised in IEEE ISCAS-15
• Hongxu Yin, Guoyang Chen, Yingmin Li, Shuai Che, Weifeng Zhang, and Niraj K. Jha, “Hardware-guided symbiotic training for compact, accurate, yet execution-efficient LSTMs,” to be submitted, Sept. 2018.
• Xiaoliang Dai*, Hongxu Yin* (eql. contrib.), and Niraj K. Jha, “Grow and prune compact, fast, and accurate LSTMs,” arXiv 1805.11797, Jun. 2018.
• Xiaoliang Dai, Hongxu Yin, and Niraj K. Jha, “NeST: A neural network synthesis tool based on a grow-and-prune paradigm,” arXiv 1711.02017, Nov. 2017.
• Hongxu Yin, Zeyu Wang, and Niraj K. Jha, “A hierarchical inference model for Internet-of-Things,” IEEE Trans. Multi-scale Computing Systems, Mar. 2018.
• Hongxu Yin and Niraj K. Jha, “A hierarchical health decision support system for disease diagnosis based on wearable medical sensors and machine learning ensembles,” IEEE Trans. Multi-scale Computing Systems, Oct. 2017. (most popular paper, featured by Tech Xplore, Home Health, etc.)
• Hongxu Yin, Ayten O. Akmandor, Arsalan Mosenia, and Niraj K. Jha, “Smart healthcare,” Foundations and Trends® in Electronic Design Automation, 2017.
• Ayten O. Akmandor*, Hongxu Yin* (eql. contrib.), and Niraj K. Jha, “Smart, secure, yet energy-efficient, Internet-of-Things sensors,” IEEE Trans. Multi-scale Computing Systems, Sept. 2017.
• Ayten O. Akmandor, Hongxu Yin, and Niraj K. Jha, “Simultaneously ensuring smartness, security, and energy efficiency in Internet-of-Things sensors,” in Proc. IEEE Custom Integrated Circuits Conference (CICC), Apr. 2018.
• Hongxu Yin, Bah H. Gwee, Zhiping Lin, Kumar Anil, G. R. Sirajudeen, and C. M. S. See, “Novel real-time system design for floating-point sub-Nyquist multi-coset signal blind reconstruction,” in Proc. IEEE Int. Symp. on Circuits and Systems, pp. 954– 957, Lisbon, Portugal, 24 -27 May 2015.
• Hongxu Yin, Bah H. Gwee, and Zhiping Lin, “Implementation of bin-wise multi-coset signal blind reconstruction algorithm on FPGA using sub-Nyquist signals,” in Proc. URECA@NTU 2013-2014, pp. 652 - 657, Nov. 2014.
• Hongxu Yin and Yuanjin Zheng, “Non-invasive magnetic resonance based stimulation system design on pain management,” in Proc. URECA@NTU 2012-2013, pp. 724-727, Nov. 2013.
Invited Talk and Poster
- H. Yin, 'Applied Machine Learning: From Theory to Practice', Invited Talk by IEEE Circuits and Systems Singapore Chapter & Centre for Infocomm Technology (INFINITUS), Singapore, Feb. 01, 2018.
- H. Yin, 'A Health Decision Support System for Disease Diagnosis based on Wearable Medical Sensors and Machine Learning Ensembles', Poster at New Jersey Tech Concil's 'What's Next in Medical Devices' Forum, Princeton, Jun. 13, 2016.
- (17-18 F') ELE364 Machine Learning for Predictive Data Analytics
- (16-17 S') ELE464 Embedded Computing
- (16-17) Senior Thesis Assistant, Electrical Engineering, Princeton University.
Details in my Linkedin profile.