Author name cluster

Pengfei Gu

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

3 papers

2 author rows

ICLR Conference 2023 Conference Paper

Searching Lottery Tickets in Graph Neural Networks: A Dual Perspective

Kun Wang 0056
Yuxuan Liang 0002
Pengkun Wang 0001
Xu Wang 0029
Pengfei Gu
Junfeng Fang
Yang Wang 0015

Graph Neural Networks (GNNs) have shown great promise in various graph learning tasks. However, the computational overheads of fitting GNNs to large-scale graphs grow rapidly, posing obstacles to GNNs from scaling up to real-world applications. To tackle this issue, Graph Lottery Ticket (GLT) hypothesis articulates that there always exists a sparse subnetwork/subgraph with admirable performance in GNNs with random initialization. Such a pair of core subgraph and sparse subnetwork (called graph lottery tickets) can be uncovered by iteratively applying a novel sparsification method. While GLT provides new insights for GNN compression, it requires a full pretraining process to obtain graph lottery tickets, which is not universal and friendly to real-world applications. Moreover, the graph sparsification in GLT utilizes sampling techniques, which may result in massive information loss and aggregation failure. In this paper, we explore the searching of graph lottery tickets from a complementary perspective -- transforming a random ticket into a graph lottery ticket, which allows us to more comprehensively explore the relationships between the original network/graph and their sparse counterpart. To achieve this, we propose regularization-based network pruning and hierarchical graph sparsification, leading to our Dual Graph Lottery Ticket (DGLT) framework for a joint sparsification of network and graph. Compared to GLT, our DGLT helps achieve a triple-win situation of graph lottery tickets with high sparsity, admirable performance, and good explainability. More importantly, we rigorously prove that our model can eliminate noise and maintain reliable information in substructures using the graph information bottleneck theory. Extensive experimental results on various graph-related tasks validate the effectiveness of our framework.

Details

IROS Conference 2022 Conference Paper

Real-Time Visual Inertial Odometry with a Resource-Efficient Harris Corner Detection Accelerator on FPGA Platform

Pengfei Gu
Ziyang Meng 0001
Pengkun Zhou

Visual Inertial Odometry (VIO) is a widely studied localization technique in robotics. State-of-the-art VIO algorithms are composed of two parts: a frontend which performs visual perception and inertial measurement pre-processing, and a backend which fuses vision and inertial measurements to estimate the robot's pose. Both image processing in the frontend and sensor fusion in the backend are computationally expensive, making it very challenging to run the VIO algorithm, especially the optimization-based VIO algorithm in real time on embedded platforms with limited power budget. In this paper, a real-time optimization-based monocular VIO algorithm is proposed based on algorithm-and-hardware co-design and successfully implemented on an embedded platform with only 2. 6W processor power consumption. In particular, the time-consuming Harris corner detection (HCD) is accelerated on Field Programmable Gate Array (FPGA), achieving an average 16 × processing time reduction compared with the ARM implementation. Compared with the state-of-the-art HCD accelerator provided by Xilinx, the hardware resource required of our accelerator is largely reduced without any compromise in speed, thanks to the proposed dedicated pruning and paral-lelization techniques. Finally, experiment on the public dataset demonstrates that the proposed real-time VIO algorithm on the FPGA-based platform has comparable accuracy with respect to the existing state-of-the-art VIO algorithm on the desktop, and 3 × faster frontend processing speed over the ARM-based implementation.

Details

TCS Journal 2021 Journal Article

Approximate set union via approximate randomization

Bin Fu
Pengfei Gu
Yuming Zhao

We develop a randomized approximation algorithm for the size of set union problem | A 1 ∪ A 2 ∪. .. ∪ A m |, which is given a list of sets A 1, .. ., A m with approximate set size m i for A i with m i ∈ ( ( 1 − β L ) | A i |, ( 1 + β R ) | A i | ), and biased random generators with probability Prob ( x = RandomElement ( A i ) ) ∈ [ 1 − α L | A i |, 1 + α R | A i | ] for each input set A i and element x ∈ A i, where i = 1, 2, .. ., m and α L, α R, β L, β R ∈ ( 0, 1 ). The approximation ratio for | A 1 ∪ A 2 ∪. .. ∪ A m | is in the range [ ( 1 − ϵ ) ( 1 − α L ) ( 1 − β L ), ( 1 + ϵ ) ( 1 + α R ) ( 1 + β R ) ] for any ϵ ∈ ( 0, 1 ). The complexity of the algorithm is measured by both time complexity and round complexity. One round of the algorithm has non-adaptive accesses to those RandomElement ( A i ) functions 1 ≤ i ≤ m, and membership queries ( x ∈ A i?) to input sets A i with 1 ≤ i ≤ m. Our algorithm gives an approximation scheme with O ( m ⋅ ( log ⁡ m ) 7 ) running time and O ( log ⁡ m ) rounds in contrast to the existing algorithm [1] that needs Ω ( m ) rounds in the worst case with O ( ( 1 + ϵ ) m / ϵ 2 ) running time, where m is the number of sets. Our algorithm gives a flexible tradeoff with time complexity O ( m 1 + ξ ) and round complexity O ( 1 ξ ) for any ξ ∈ ( 0, 1 ). Our algorithm runs sublinear in time under certain condition that each element in A 1 ∪ A 2 ∪. .. ∪ A m belongs to m a sets for any fixed a > 0, to our best knowledge, we have not seen any sublinear results about this problem. Our algorithm can handle input sets that can generate random elements with bias, and its approximation ratio depends on the bias. We prove that it is #P-hard to count the number of lattice points in a set of balls, and we also show that there is no polynomial time algorithm to approximate the number of lattice points in the intersection of n-dimensional balls unless P=NP. As applications of our algorithm, we propose approximation algorithms for counting the number of lattice points in a union of high dimensional balls and for the maximal coverage problem with balls.

Details DOI