Peter Y. K. Cheung

Peter Y. K. Cheung

Auto-generated profile, based on publicly available information

Peter Y. K. Cheung

Imperial College London, United Kingdom

Research Focus & Metrics

Academic Career Length

21Years (2006-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
6 March 2020 in IEEE Transactions on Computers
Verified
Journal Article
Closed access
EWErwei Wang
Erwei Wang
JDJames J. Davis
James J. Davis
PCPeter Y. K. Cheung
Peter Y. K. Cheung
GCGeorge A. Constantinides
George A. Constantinides
Research has shown that deep neural networks contain significant redundancy, and thus that high classification accuracy can be achieved even when weights and activations are quantised down to binary values. Network binarisation on FPGAs greatly increases area efficiency by replacing resource-hungry multipliers with lightweight XNOR gates. However, an FPGA's fundamental building block, the K-LUT, is capable of implementing far more than an XNOR: it can perform any K-input Boolean operation. Inspired by this observation, we propose LUTNet, an end-to-end hardware-software framework for the construction of area-efficient FPGA-based neural network accelerators using the native LUTs as inference operators. We describe the realisation of both unrolled and tiled LUTNet architectures, with the latter facilitating smaller, less power-hungry deployment over the former while sacrificing area and energy efficiency along with throughput. For both varieties, we demonstrate that the exploitation of LUT flexibility allows for far heavier pruning than possible in prior works, resulting in significant area savings while achieving comparable accuracy. Against the state-of-the-art binarised neural network implementation, we achieve up to twice the area efficiency for several standard network models when inferencing popular datasets. We also demonstrate that even greater energy efficiency improvements are obtainable.

21 January 2021 in PLOS Computational Biology
Verified
Journal Article
Open access
PDF file
MBMichelangelo Bin
Michelangelo Bin
PCPeter Y. K. Cheung
Peter Y. K. Cheung
ECEmanuele Crisostomi
Emanuele Crisostomi
PFPietro Ferraro
Pietro Ferraro
HLHugo Lhachemi
Hugo Lhachemi
LSLewi Stone
Lewi Stone
COVID-19 abatement strategies have risks and uncertainties which could lead to repeating waves of infection. We show—as proof of concept grounded on rigorous mathematical evidence—that periodic, high-frequency alternation of into, and out-of, lockdown effectively mitigates second-wave effects, while allowing continued, albeit reduced, economic activity. Periodicity confers (i) predictability, which is essential for economic sustainability, and (ii) robustness, since lockdown periods are not activated by uncertain measurements over short time scales. In turn—while not eliminating the virus—this fast switching policy is sustainable over time, and it mitigates the infection until a vaccine or treatment becomes available, while alleviating the social costs associated with long lockdowns. Typically, the policy might be in the form of 1-day of work followed by 6-days of lockdown every week (or perhaps 2 days working, 5 days off) and it can be modified at a slow-rate based on measurements filtered over longer time scales. Our results highlight the potential efficacy of high frequency switching interventions in post lockdown mitigation. All code is available on Github at https://github.com/V4p1d/FPSP_Covid19. A software tool has also been developed so that interested parties can explore the proof-of-concept system. Why? The design of post-lockdown mitigation policies while vaccines are still not available is pressing now as new secondary waves of the virus have emerged in many countries (for example, in Spain, France, UK, Italy, Israel, and others), and as several of these countries grapple with the reintroduction of full lockdown measures. What do we do and find? We propose efficacious and realisable methods based on control theory to tame the complex behaviour of COVID-19 in well mixed populations. We achieve this through a policy of fast intermittent lockdown intervals with regular period. We illustrate how our approach offers a fundamentally new perspective on ways to design COVID-19 exit strategies from policies of total lockdown. Our theoretical results are also very general and apply to a wide range of epidemiological models. What do these findings mean? Unlike many other proposed abatement strategies, which have risks and uncertainties possibly leading to multiple waves of infection, we demonstrate that our proposed policies have the potential to suppress the virus outbreak, while at the same time allowing continued economic activity. These policies, while of practical significance, are built on rigorous theoretical results, which are to the best of our knowledge, new in mathematical epidemiology. An extensive validation is carried out using a detailed epidemic model validated on real COVID-19 data from Italy and published very recently in Nature Medicine.

4 October 2007 in Journal of Signal Processing Systems
Verified
Journal Article
Closed access
PDF file
MAMaria E. Angelopoulou
Maria E. Angelopoulou
KMKonstantinos Masselos
Konstantinos Masselos
PCPeter Y. K. Cheung
Peter Y. K. Cheung
YAYiannis Andreopoulos
Yiannis Andreopoulos

1 December 2007
Verified
Conference Paper
Closed access
JWJustin S. J. Wong
Justin S. J. Wong
PSPete Sedcole
Pete Sedcole
PCPeter Y. K. Cheung
Peter Y. K. Cheung
This paper proposes a built-in self-test (BIST) method to measure accurately the combinatorial circuit delays on an FPGA. The flexibility of the on-chip clock generation capability found in modern FPGAs is employed to step through a range of frequencies until timing failure in the combinatorial circuit is detected. In this way, the delay of any combinatorial circuit can be determined with a timing resolution of 1 ps or lower. A parallel implementation of the method for self-characterization of the delay of all the LUTs on an FPGA is also proposed. The method was applied to an Altera Cyclone-II FPGA (EP2C35). A complete self-characterization was achieved in 3 seconds, utilizing only 13 kbit of block RAM to store the results. This self-characterization method paves the way for matching timing requirements in designs to FPGAs as a means of combating the problem of process variations.

1 August 2007
Verified
Conference Paper
Closed access
YLYang Liu
Yang Liu
CBChristos-Savvas Bouganis
Christos-Savvas Bouganis
PCPeter Y. K. Cheung
Peter Y. K. Cheung
The Kalman filter is widely used as an estimator in many modern applications. In the case where its implementation in hardware is required, the computational complexity of the algorithm dictates the use of many resources. This paper presents an approximation of the conventional Kalman filter by using Taylor expansion and matrix calculus in order to remove the hardware expensive part of the algorithm. The Bierman-Thornton algorithm, as the exact counterpart of our proposed Approximate Kalman filter algorithm, is also implemented for comparison purposes. Comparing to the Bierman-Thornton algorithm, the FPGA implementation results demonstrate that our proposed Approximate Kalman filter implementation achieves one order of magnitude higher throughput using less hardware resources, obtaining similar convergence rate and accuracy.

9 November 2023 in ACM Transactions on Embedded Computing Systems
Verified
Journal Article
Open access
EWErwei Wang
Erwei Wang
JDJames J. Davis
James J. Davis
DMDaniele Moro
Daniele Moro
PZPiotr Zielinski
Piotr Zielinski
JLJia Jie Lim
Jia Jie Lim
GCGeorge A. Constantinides
George A. Constantinides
The ever-growing computational demands of increasingly complex machine learning models frequently necessitate the use of powerful cloud-based infrastructure for their training. Binary neural networks are known to be promising candidates for on-device inference due to their extreme compute and memory savings over higher-precision alternatives. However, their existing training methods require the concurrent storage of high-precision activations for all layers, generally making learning on memory-constrained devices infeasible. In this article, we demonstrate that the backward propagation operations needed for binary neural network training are strongly robust to quantization, thereby making on-the-edge learning with modern models a practical proposition. We introduce a low-cost binary neural network training strategy exhibiting sizable memory footprint reductions while inducing little to no accuracy loss vs Courbariaux & Bengio’s standard approach. These decreases are primarily enabled through the retention of activations exclusively in binary format. Against the latter algorithm, our drop-in replacement sees memory requirement reductions of 3–5×, while reaching similar test accuracy (± 2 pp) in comparable time, across a range of small-scale models trained to classify popular datasets. We also demonstrate from-scratch ImageNet training of binarized ResNet-18, achieving a 3.78× memory reduction. Our work is open-source, and includes the Raspberry Pi-targeted prototype we used to verify our modeled memory decreases and capture the associated energy drops. Such savings will allow for unnecessary cloud offloading to be avoided, reducing latency, increasing energy efficiency, and safeguarding end-user privacy.