Jun Chu

Shuai Lu, Jun Chu, Xu T. Liu

Convolution is the most expensive operation among neural network operations, thus its performance is critical to the overall performance of neural networks. Commonly used convolution approaches, including general matrix multiplication (GEMM)-based convolution and direct convolution, rely on im2col for data transformation or do not use data transformation at all, respectively. However, the im2col data transformation can lead to at least 2$\times$ memory footprint compared to not using data transformation at all, thus limiting the size of neural network models running on memory-limited systems. Meanwhile, not using data transformation usually performs poorly due to nonconsecutive memory access although it consumes less memory. To solve those problems, we propose a new memory-efficient data transformation algorithm, called im2win. This algorithm refactorizes a row of square or rectangle dot product windows of the input image and flattens unique elements within these windows into a row in the output tensor, which enables consecutive memory access and data reuse, and thus greatly reduces the memory overhead. Furthermore, we propose a high-performance im2win-based convolution algorithm with various optimizations, including vectorization, loop reordering, etc. Our experimental results show that our algorithm reduces the memory overhead by average to 41.6% compared to the PyTorch's convolution implementation based on im2col, and achieves average to 3.6$\times$ and 5.3$\times$ speedup in performance compared to the im2col-based convolution and not using data transformation, respectively.

Shuai Lu, Jun Chu, Luanzheng Guo et al.

Convolution is the most time-consuming operation in deep neural network operations, so its performance is critical to the overall performance of the neural network. The commonly used methods for convolution on GPU include the general matrix multiplication (GEMM)-based convolution and the direct convolution. GEMM-based convolution relies on the im2col algorithm, which results in a large memory footprint and reduced performance. Direct convolution does not have the large memory footprint problem, but the performance is not on par with GEMM-based approach because of the discontinuous memory access. This paper proposes a window-order-based convolution paradigm on GPU, called im2win, which not only reduces memory footprint but also offers continuous memory accesses, resulting in improved performance. Furthermore, we apply a range of optimization techniques on the convolution CUDA kernel, including shared memory, tiling, micro-kernel, double buffer, and prefetching. We compare our implementation with the direct convolution, and PyTorch's GEMM-based convolution with cuBLAS and six cuDNN-based convolution implementations, with twelve state-of-the-art DNN benchmarks. The experimental results show that our implementation 1) uses less memory footprint by 23.1% and achieves 3.5$\times$ TFLOPS compared with cuBLAS, 2) uses less memory footprint by 32.8% and achieves up to 1.8$\times$ TFLOPS compared with the best performant convolutions in cuDNN, and 3) achieves up to 155$\times$ TFLOPS compared with the direct convolution. We further perform an ablation study on the applied optimization techniques and find that the micro-kernel has the greatest positive impact on performance.

Maolin Wang, Jun Chu, Sicong Xie et al.

In the era of mobile computing, deploying efficient Natural Language Processing (NLP) models in resource-restricted edge settings presents significant challenges, particularly in environments requiring strict privacy compliance, real-time responsiveness, and diverse multi-tasking capabilities. These challenges create a fundamental need for ultra-compact models that maintain strong performance across various NLP tasks while adhering to stringent memory constraints. To this end, we introduce Edge ultra-lIte BERT framework (EI-BERT) with a novel cross-distillation method. EI-BERT efficiently compresses models through a comprehensive pipeline including hard token pruning, cross-distillation and parameter quantization. Specifically, the cross-distillation method uniquely positions the teacher model to understand the student model's perspective, ensuring efficient knowledge transfer through parameter integration and the mutual interplay between models. Through extensive experiments, we achieve a remarkably compact BERT-based model of only 1.91 MB - the smallest to date for Natural Language Understanding (NLU) tasks. This ultra-compact model has been successfully deployed across multiple scenarios within the Alipay ecosystem, demonstrating significant improvements in real-world applications. For example, it has been integrated into Alipay's live Edge Recommendation system since January 2024, currently serving the app's recommendation traffic across \textbf{8.4 million daily active devices}.

Tiehang Duan, Mihir Chauhan, Mohammad Abuzar Shaikh et al.

The pattern of Electroencephalogram (EEG) signal differs significantly across different subjects, and poses challenge for EEG classifiers in terms of 1) effectively adapting a learned classifier onto a new subject, 2) retaining knowledge of known subjects after the adaptation. We propose an efficient transfer learning method, named Meta UPdate Strategy (MUPS-EEG), for continuous EEG classification across different subjects. The model learns effective representations with meta update which accelerates adaptation on new subject and mitigate forgetting of knowledge on previous subjects at the same time. The proposed mechanism originates from meta learning and works to 1) find feature representation that is broadly suitable for different subjects, 2) maximizes sensitivity of loss function for fast adaptation on new subject. The method can be applied to all deep learning oriented models. Extensive experiments on two public datasets demonstrate the effectiveness of the proposed model, outperforming current state of the arts by a large margin in terms of both adapting on new subject and retain knowledge of learned subjects.

Mohammad Abuzar Shaikh, Mihir Chauhan, Jun Chu et al.

We propose an effective Hybrid Deep Learning (HDL) architecture for the task of determining the probability that a questioned handwritten word has been written by a known writer. HDL is an amalgamation of Auto-Learned Features (ALF) and Human-Engineered Features (HEF). To extract auto-learned features we use two methods: First, Two Channel Convolutional Neural Network (TC-CNN); Second, Two Channel Autoencoder (TC-AE). Furthermore, human-engineered features are extracted by using two methods: First, Gradient Structural Concavity (GSC); Second, Scale Invariant Feature Transform (SIFT). Experiments are performed by complementing one of the HEF methods with one ALF method on 150000 pairs of samples of the word "AND" cropped from handwritten notes written by 1500 writers. Our results indicate that HDL architecture with AE-GSC achieves 99.7% accuracy on seen writer dataset and 92.16% accuracy on shuffled writer dataset which out performs CEDAR-FOX, as for unseen writer dataset, AE-SIFT performs comparable to this sophisticated handwriting comparison tool.

Luanzheng Guo, Jun Chu

An important yet challenging problem in understanding indoor scene is recovering indoor frame structure from a monocular image. It is more difficult when occlusions and illumination vary, and object boundaries are weak. To overcome these difficulties, a new approach based on line segment refinement with two constraints is proposed. First, the line segments are refined by four consecutive operations, i.e., reclassifying, connecting, fitting, and voting. Specifically, misclassified line segments are revised by the reclassifying operation, some short line segments are joined by the connecting operation, the undetected key line segments are recovered by the fitting operation with the help of the vanishing points, the line segments converging on the frame are selected by the voting operation. Second, we construct four frame models according to four classes of possible shooting angles of the monocular image, the natures of all frame models are introduced via enforcing the cross ratio and depth constraints. The indoor frame is then constructed by fitting those refined line segments with related frame model under the two constraints, which jointly advance the accuracy of the frame. Experimental results on a collection of over 300 indoor images indicate that our algorithm has the capability of recovering the frame from complex indoor scenes.