Philip Heng Wai Leong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Philip Heng Wai Leong has a total of 128 co-authors. Here are the top co-authors by number of shared publications:
- David Boland
The University of Sydney, Sydney, Australia
13 shared publications - Anthony Mihirana De Silva
University of Sydney
7 shared publications - Binglei Lou
School of Electrical and Computer Engineering, the University of Sydney, Australia
5 shared publications - Carol Jingyi Li
Faculty of Engineering, School of Electrical and Computer Engineering, The University of Sydney, Sydney, NSW, Australia
4 shared publications - Duncan J. M. Moss
School of Electrical and Information Engineering, The University of Sydney, Sydney, NSW, Australia
4 shared publications - Craig Jin
School of Electrical and Information Engineering, The University of Sydney, Sydney, NSW, Australia
3 shared publications
Latest Publications
- Open Access
- 18 July 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Reconfigurable Technology and Systems
A Flexible FPGA-based Butterfly Engine for Accelerating Signal Processing and Machine Learning
Field-programmable gate arrays (FPGAs) have emerged as efficient accelerators for both neural network (NN) inference and digital signal processing (DSP) tasks, particularly on resource-constrained edge devices. While previous NN accelerators based on butterfly operations demonstrate significant acceleration for inference, they are inadequately suited for long sequences and lack support for bit-reversed access patterns, limiting their applicability to variable-length DSP workloads. Based on prior butterfly accelerators, this paper presents flexible butterfly engine (FlexBE), together with a co-designed NN architecture, Butterfly-based Signal Processing Net (BSPNet). The proposed system jointly supports signal pre-processing and butterfly linear (BL)-based NN inference under stringent resource constraints. FlexBE incorporates novel adaptive data switching networks, dynamic access control mechanisms, and an efficient bit-reversal module, enabling runtime reconfiguration of sequence lengths and degrees of parallelism. Implemented on an AMD ZCU104 FPGA running at \(300\) MHz, FlexBE computes four \(32\) k-point fast Fourier transforms (FFTs) in approximately \(15,360\) clock cycles. On challenging automatic modulation classification (AMC) datasets, BSPNet achieves accuracy comparable to GPU baselines. For single-batch inference, BSPNet with FlexBE is \(2.2\sim 3.1\times\) faster than prior butterfly-based accelerators; on the ZCU104, the end-to-end latency achieves speedups of up to \(4.92\times\) and \(2.89\times\) compared to an Intel Core i9 CPU and an NVIDIA RTX 3090 GPU, respectively.- 18 July 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Reconfigurable Technology and Systems
Open AccessA Flexible FPGA-based Butterfly Engine for Accelerating Signal Processing and Machine Learning
Field-programmable gate arrays (FPGAs) have emerged as efficient accelerators for both neural network (NN) inference and digital signal processing (DSP) tasks, particularly on resource-constrained edge devices. While previous NN accelerators based on butterfly operations demonstrate significant acceleration for inference, they are inadequately suited for long sequences and lack support for bit-reversed access patterns, limiting their applicability to variable-length DSP workloads. Based on prior butterfly accelerators, this paper presents flexible butterfly engine (FlexBE), together with a co-designed NN architecture, Butterfly-based Signal Processing Net (BSPNet). The proposed system jointly supports signal pre-processing and butterfly linear (BL)-based NN inference under stringent resource constraints. FlexBE incorporates novel adaptive data switching networks, dynamic access control mechanisms, and an efficient bit-reversal module, enabling runtime reconfiguration of sequence lengths and degrees of parallelism. Implemented on an AMD ZCU104 FPGA running at \(300\) MHz, FlexBE computes four \(32\) k-point fast Fourier transforms (FFTs) in approximately \(15,360\) clock cycles. On challenging automatic modulation classification (AMC) datasets, BSPNet achieves accuracy comparable to GPU baselines. For single-batch inference, BSPNet with FlexBE is \(2.2\sim 3.1\times\) faster than prior butterfly-based accelerators; on the ZCU104, the end-to-end latency achieves speedups of up to \(4.92\times\) and \(2.89\times\) compared to an Intel Core i9 CPU and an NVIDIA RTX 3090 GPU, respectively. - 15 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Control Systems Letters
- Vol. 10, 1651-1656
Quantization Robustness of Monotone Operator Equilibrium Networks
Monotone operator equilibrium networks (MonDEQs) are implicit models whose well-posedness and convergence rely on a single scalar margin computed from the weight matrix, which has made them attractive for learned controllers with closed-loop stability guarantees. When such models are deployed on low-precision hardware, weights are quantized, and these structural guarantees can be destroyed. We provide an explicit quantizationpreservation certificate: a single computable check on the weight perturbation determines whether the quantized network retains existence, uniqueness, linear convergence, and a bounded displacement of its equilibrium. The same threshold certifies the backward solve as well as the forward solve, so it governs both training and inference. MNIST experiments confirm a phase transition at the predicted threshold: three- and four-bit post-training quantization diverge, while five-bit and above converge; quantization-aware training recovers provable convergence at four bits.
- 15 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Control Systems Letters
- Vol. 10, 1651-1656
Quantization Robustness of Monotone Operator Equilibrium Networks
Monotone operator equilibrium networks (MonDEQs) are implicit models whose well-posedness and convergence rely on a single scalar margin computed from the weight matrix, which has made them attractive for learned controllers with closed-loop stability guarantees. When such models are deployed on low-precision hardware, weights are quantized, and these structural guarantees can be destroyed. We provide an explicit quantizationpreservation certificate: a single computable check on the weight perturbation determines whether the quantized network retains existence, uniqueness, linear convergence, and a bounded displacement of its equilibrium. The same threshold certifies the backward solve as well as the forward solve, so it governs both training and inference. MNIST experiments confirm a phase transition at the predicted threshold: three- and four-bit post-training quantization diverge, while five-bit and above converge; quantization-aware training recovers provable convergence at four bits.
- Open Access
- 19 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64  (01962892) , 1-6
Segmentation of the F2 Leading Edge in a Backscatter Sounder Ionogram
Backscatter sounder (BSS) systems colocated with over-the-horizon radar (OTHR) can be leveraged to provide a near-real-time characterization of the ionospheric propagation conditions to aid the OTHR surveillance mission. Ionogram inversion can be performed using features derived from a backscatter ionogram to infer the optimum OTHR operational parameters and to enable the geolocation of the detected targets. In this article, we demonstrate the application of a deep learning segmentation model on a backscatter ionogram dataset obtained from the Jindalee Operational Radar Network (JORN) frequency management system (FMS) to extract the F2 leading edge (LE) feature. Our results show that deep learning is a viable method for LE feature identification and extraction in backscatter ionogram imagery.
- 19 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64  (01962892) , 1-6
Open AccessSegmentation of the F2 Leading Edge in a Backscatter Sounder Ionogram
Backscatter sounder (BSS) systems colocated with over-the-horizon radar (OTHR) can be leveraged to provide a near-real-time characterization of the ionospheric propagation conditions to aid the OTHR surveillance mission. Ionogram inversion can be performed using features derived from a backscatter ionogram to infer the optimum OTHR operational parameters and to enable the geolocation of the detected targets. In this article, we demonstrate the application of a deep learning segmentation model on a backscatter ionogram dataset obtained from the Jindalee Operational Radar Network (JORN) frequency management system (FMS) to extract the F2 leading edge (LE) feature. Our results show that deep learning is a viable method for LE feature identification and extraction in backscatter ionogram imagery.
- 15 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. PP  (02780070) , 1
Enhancing LUT-based Deep Neural Networks Inference through Architecture and Connectivity Optimization
Deploying deep neural networks (DNNs) on resource-constrained edge devices such as FPGAs requires a careful balance among latency, power, and hardware resource usage, while maintaining high accuracy. Existing Lookup Table (LUT)-based DNNs—such as LogicNets, PolyLUT, and NeuraLUT—face two critical challenges: the exponential growth of LUT size and inefficient random sparse connectivity. This paper presents SparseLUT, a comprehensive framework that addresses these challenges through two orthogonal optimizations. First, we propose an architectural enhancement that aggregates multiple PolyLUT sub-neurons via an adder, significantly reducing LUT consumption by 2.0×–13.9× and lowering inference latency by 1.2×–1.6×, all while maintaining comparable accuracy. Building upon this foundation, we further introduce a non-greedy training algorithm that optimizes neuron connectivity by selectively pruning less significant inputs and strategically regrowing more effective ones. This training optimization, which incurs no additional area and latency overhead, delivers consistent accuracy improvements across benchmarks—achieving up to a 2.13% gain on MNIST and 0.94% on Jet Substructure Classification compared to existing LUT-DNN approaches.
- 15 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. PP  (02780070) , 1
Enhancing LUT-based Deep Neural Networks Inference through Architecture and Connectivity Optimization
Deploying deep neural networks (DNNs) on resource-constrained edge devices such as FPGAs requires a careful balance among latency, power, and hardware resource usage, while maintaining high accuracy. Existing Lookup Table (LUT)-based DNNs—such as LogicNets, PolyLUT, and NeuraLUT—face two critical challenges: the exponential growth of LUT size and inefficient random sparse connectivity. This paper presents SparseLUT, a comprehensive framework that addresses these challenges through two orthogonal optimizations. First, we propose an architectural enhancement that aggregates multiple PolyLUT sub-neurons via an adder, significantly reducing LUT consumption by 2.0×–13.9× and lowering inference latency by 1.2×–1.6×, all while maintaining comparable accuracy. Building upon this foundation, we further introduce a non-greedy training algorithm that optimizes neuron connectivity by selectively pruning less significant inputs and strategically regrowing more effective ones. This training optimization, which incurs no additional area and latency overhead, delivers consistent accuracy improvements across benchmarks—achieving up to a 2.13% gain on MNIST and 0.94% on Jet Substructure Classification compared to existing LUT-DNN approaches.
- Open Access
Enhancing LUT-based Deep Neural Networks Inference through Architecture and Connectivity Optimization
Deploying deep neural networks (DNNs) on resource-constrained edge devices such as FPGAs requires a careful balance among latency, power, and hardware resource usage, while maintaining high accuracy. Existing Lookup Table (LUT)-based DNNs -- such as LogicNets, PolyLUT, and NeuraLUT -- face two critical challenges: the exponential growth of LUT size and inefficient random sparse connectivity. This paper presents SparseLUT, a comprehensive framework that addresses these challenges through two orthogonal optimizations. First, we propose an architectural enhancement that aggregates multiple PolyLUT sub-neurons via an adder, significantly reducing LUT consumption by 2.0x-13.9x and lowering inference latency by 1.2x-1.6x, all while maintaining comparable accuracy. Building upon this foundation, we further introduce a non-greedy training algorithm that optimizes neuron connectivity by selectively pruning less significant inputs and strategically regrowing more effective ones. This training optimization, which incurs no additional area and latency overhead, delivers consistent accuracy improvements across benchmarks -- achieving up to a 2.13% gain on MNIST and 0.94% on Jet Substructure Classification compared to existing LUT-DNN approaches.
Open AccessEnhancing LUT-based Deep Neural Networks Inference through Architecture and Connectivity Optimization
Deploying deep neural networks (DNNs) on resource-constrained edge devices such as FPGAs requires a careful balance among latency, power, and hardware resource usage, while maintaining high accuracy. Existing Lookup Table (LUT)-based DNNs -- such as LogicNets, PolyLUT, and NeuraLUT -- face two critical challenges: the exponential growth of LUT size and inefficient random sparse connectivity. This paper presents SparseLUT, a comprehensive framework that addresses these challenges through two orthogonal optimizations. First, we propose an architectural enhancement that aggregates multiple PolyLUT sub-neurons via an adder, significantly reducing LUT consumption by 2.0x-13.9x and lowering inference latency by 1.2x-1.6x, all while maintaining comparable accuracy. Building upon this foundation, we further introduce a non-greedy training algorithm that optimizes neuron connectivity by selectively pruning less significant inputs and strategically regrowing more effective ones. This training optimization, which incurs no additional area and latency overhead, delivers consistent accuracy improvements across benchmarks -- achieving up to a 2.13% gain on MNIST and 0.94% on Jet Substructure Classification compared to existing LUT-DNN approaches.
- 18 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Radio Frequency Identification
- Vol. 10, 90-100
Space Sign Language for Spacewalks: Sign Profiling and Edge Computing Approach
Long journeys for space exploration demand innovative solutions to address hazards where conventional communication systems may fail due to electromagnetic (EM) disruptions or environmental extremes. Effective search and rescue strategies are vital for spacewalks and unforeseen EM instability. We propose a research direction involving ad-hoc, direct communication protocols to enhance survivability under harsh space conditions. It provides a pathway for real-time communication that is perceptible to humans, computationally efficient, and resilient to EM interference. Furthermore, it can take advantage of upcoming advancements in wearable sensors and non-terrestrial edge computing. Our proposed methods include sign profiling via analysis of visual cues from sign language for the deaf. Profiles can be achieved by detecting critical pose landmarks through a body area network of wearable sensors. We also recommend an embedded artificial intelligence approach using edge computing to achieve real-time performance with small size, weight, power and cost. Our work may lead to new developments in spacesuit design and new search and rescue practices. We also propose related research problems concerning variations in sign languages across communities to foster seamless spoken and unspoken exchanges.
- 18 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Radio Frequency Identification
- Vol. 10, 90-100
Space Sign Language for Spacewalks: Sign Profiling and Edge Computing Approach
Long journeys for space exploration demand innovative solutions to address hazards where conventional communication systems may fail due to electromagnetic (EM) disruptions or environmental extremes. Effective search and rescue strategies are vital for spacewalks and unforeseen EM instability. We propose a research direction involving ad-hoc, direct communication protocols to enhance survivability under harsh space conditions. It provides a pathway for real-time communication that is perceptible to humans, computationally efficient, and resilient to EM interference. Furthermore, it can take advantage of upcoming advancements in wearable sensors and non-terrestrial edge computing. Our proposed methods include sign profiling via analysis of visual cues from sign language for the deaf. Profiles can be achieved by detecting critical pose landmarks through a body area network of wearable sensors. We also recommend an embedded artificial intelligence approach using edge computing to achieve real-time performance with small size, weight, power and cost. Our work may lead to new developments in spacesuit design and new search and rescue practices. We also propose related research problems concerning variations in sign languages across communities to foster seamless spoken and unspoken exchanges.
- Open Access
- 7 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 14  (1) , 49-68
Ionogram Scaling for Vertical, Oblique, And Backscatter Sounders: A review of challenges and opportunities
The process of feature identification in ionogram images and subsequent derivation of ionospheric parameters, known as ionogram scaling, enables the characterization of the ionospheric propagation conditions of radio waves in the high-frequency (HF) band. Manual ionogram scaling is a time-intensive and laborious activity; hence, numerous techniques have been developed by many researchers to automate this task. This article presents a review of the various automatic scaling approaches applicable to vertical, oblique, and backscatter ionograms. In addition, we provide a set of recommendations to guide and stimulate further research in this field.
- 7 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 14  (1) , 49-68
Open AccessIonogram Scaling for Vertical, Oblique, And Backscatter Sounders: A review of challenges and opportunities
The process of feature identification in ionogram images and subsequent derivation of ionospheric parameters, known as ionogram scaling, enables the characterization of the ionospheric propagation conditions of radio waves in the high-frequency (HF) band. Manual ionogram scaling is a time-intensive and laborious activity; hence, numerous techniques have been developed by many researchers to automate this task. This article presents a review of the various automatic scaling approaches applicable to vertical, oblique, and backscatter ionograms. In addition, we provide a set of recommendations to guide and stimulate further research in this field.
- Open Access
- 6 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Imaging
- Vol. 11  (25730436) , 888-900
A Physics-Inspired Deep Learning Framework With Polar Coordinate Attention for Ptychographic Imaging
Ptychographic imaging confronts inherent challenges in applying deep learning for phase retrieval from diffraction patterns. Conventional neural architectures, both convolutional neural networks and Transformer-based methods, are optimized for natural images with Euclidean spatial neighborhood-based inductive biases that exhibit geometric mismatch with the concentric coherent patterns characteristic of diffraction data in reciprocal space. In this paper, we present PPN, a physics-inspired deep learning network with Polar Coordinate Attention (PoCA) for ptychographic imaging, that aligns neural inductive biases with diffraction physics through a dual-branch architecture separating local feature extraction from non-local coherence modeling. It consists of a PoCA mechanism that replaces Euclidean spatial priors with physically consistent radial-angular correlations. PPN outperforms existing end-to-end models, with spectral and spatial analysis confirming its greater preservation of high-frequency details. Notably, PPN maintains robust performance compared to iterative methods even at low overlap ratios — well-suited for high-throughput imaging in real-world acquisition scenarios for samples with consistent structural characteristics.
- 6 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Imaging
- Vol. 11  (25730436) , 888-900
Open AccessA Physics-Inspired Deep Learning Framework With Polar Coordinate Attention for Ptychographic Imaging
Ptychographic imaging confronts inherent challenges in applying deep learning for phase retrieval from diffraction patterns. Conventional neural architectures, both convolutional neural networks and Transformer-based methods, are optimized for natural images with Euclidean spatial neighborhood-based inductive biases that exhibit geometric mismatch with the concentric coherent patterns characteristic of diffraction data in reciprocal space. In this paper, we present PPN, a physics-inspired deep learning network with Polar Coordinate Attention (PoCA) for ptychographic imaging, that aligns neural inductive biases with diffraction physics through a dual-branch architecture separating local feature extraction from non-local coherence modeling. It consists of a PoCA mechanism that replaces Euclidean spatial priors with physically consistent radial-angular correlations. PPN outperforms existing end-to-end models, with spectral and spatial analysis confirming its greater preservation of high-frequency details. Notably, PPN maintains robust performance compared to iterative methods even at low overlap ratios — well-suited for high-throughput imaging in real-world acquisition scenarios for samples with consistent structural characteristics.
- 2 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. PP  (10638210) , 1-14
Toward High-Performance Network Coding: FPGA Acceleration With Bounded-Value Generators
The network coding enhances performance in network communications and distributed storage by increasing throughput and robustness while reducing latency. Batched sparse (BATS) codes are a class of capacity-achieving network codes, but their practical applications are hindered by their structure, computational intensity, and power demands of finite field (FF) operations. Most literature focuses on algorithmic-level techniques to improve the coding efficiency. Optimization with an algorithm/hardware co-designing approach has long been neglected. Leveraging the unique structure of BATS codes, we first present cyclic-shift BATS (CS-BATS), a hardware-friendly variant. Next, we propose a simple but effective bounded-value (BV) generator, to reduce the size of a finite field multiplier by up to 70%. Finally, we report on a scalable and resource-efficient field-programmable gate array (FPGA)-based network coding accelerator that achieves a throughput of 27 Gb/s, a speedup of more than 300 over software.
- 2 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Very Large Scale Integration (VLSI) Systems
- Vol. PP  (10638210) , 1-14
Toward High-Performance Network Coding: FPGA Acceleration With Bounded-Value Generators
The network coding enhances performance in network communications and distributed storage by increasing throughput and robustness while reducing latency. Batched sparse (BATS) codes are a class of capacity-achieving network codes, but their practical applications are hindered by their structure, computational intensity, and power demands of finite field (FF) operations. Most literature focuses on algorithmic-level techniques to improve the coding efficiency. Optimization with an algorithm/hardware co-designing approach has long been neglected. Leveraging the unique structure of BATS codes, we first present cyclic-shift BATS (CS-BATS), a hardware-friendly variant. Next, we propose a simple but effective bounded-value (BV) generator, to reduce the size of a finite field multiplier by up to 70%. Finally, we report on a scalable and resource-efficient field-programmable gate array (FPGA)-based network coding accelerator that achieves a throughput of 27 Gb/s, a speedup of more than 300 over software.
- Open Access
- 21 March 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Reconfigurable Technology and Systems
- Vol. 18  (2) , 1-23
FPGA-based Block Minifloat Training Accelerator for a Time Series Prediction Network
Time series forecasting is the problem of predicting future data samples from historical information and recent deep neural network (DNNs) based techniques have achieved excellent results compared with conventional statistical approaches. Many applications at the edge can utilize this technology and most implementations have focused on inference, an ability to train at the edge would enable the DNN to adapt to changing conditions. Unfortunately, training requires approximately three times more memory and computation than inference. Moreover, edge applications are often constrained by energy efficiency. In this work, we implement a block minifloat (BM) training accelerator for a time series prediction network, N-BEATS. Our architecture involves a mixed-precision GEMM accelerator that utilizes BM arithmetic. We use a 4-bit DSP packing scheme to optimize the implementation further, achieving a throughput of 779 Gops. The resulting power efficiency is 42.4 Gops/W, 3.1 \(\times\) better than a graphics processing unit in a similar technology.- 21 March 2025
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Reconfigurable Technology and Systems
- Vol. 18  (2) , 1-23
Open AccessFPGA-based Block Minifloat Training Accelerator for a Time Series Prediction Network
Time series forecasting is the problem of predicting future data samples from historical information and recent deep neural network (DNNs) based techniques have achieved excellent results compared with conventional statistical approaches. Many applications at the edge can utilize this technology and most implementations have focused on inference, an ability to train at the edge would enable the DNN to adapt to changing conditions. Unfortunately, training requires approximately three times more memory and computation than inference. Moreover, edge applications are often constrained by energy efficiency. In this work, we implement a block minifloat (BM) training accelerator for a time series prediction network, N-BEATS. Our architecture involves a mixed-precision GEMM accelerator that utilizes BM arithmetic. We use a 4-bit DSP packing scheme to optimize the implementation further, achieving a throughput of 779 Gops. The resulting power efficiency is 42.4 Gops/W, 3.1 \(\times\) better than a graphics processing unit in a similar technology.