Transformers are the cornerstone of modern large language models, but their quadratic computational complexity limits efficiency in long-sequence processing. Recent advancements in Mamba, a state space model (SSM) with linear complexity, offer promising efficiency gains but suffer from unstable contextual learning and multitask generalization. Some works conduct layer-level hybrid structures that combine Transformer and Mamba layers, aiming to make full use of both advantages. This paper proposes TransMamba, a novel sequence-level hybrid framework that unifies Transformer and Mamba through shared parameter matrices (QKV and CBx), and thus could dynamically switch between attention and SSM mechanisms at different token lengths and layers. We design the Memory Converter to bridge Transformer and Mamba by converting attention outputs into SSM-compatible states, ensuring seamless information flow at TransPoints where the transformation happens. The TransPoint scheduling is also thoroughly explored for balancing effectiveness and efficiency. We conducted extensive experiments demonstrating that TransMamba achieves superior training efficiency and performance compared to single and hybrid baselines, and validated the deeper consistency between Transformer and Mamba paradigms at sequence level, offering a scalable solution for next-generation language modeling.

Low-precision training is considered an effective strategy for reducing both training and downstream inference costs. Previous scaling laws for precision mainly focus on integer quantization, which pay less attention to the constituents in floating-point (FP) quantization, and thus cannot well fit the LLM losses in this scenario. In contrast, while FP quantization training is more commonly implemented in production, it's research has been relatively superficial. In this paper, we thoroughly explore the effects of FP quantization targets, exponent bits, mantissa bits, and the calculation granularity of the scaling factor in FP quantization training performance of LLM models. In addition to an accurate FP quantization unified scaling law, we also provide valuable suggestions for the community: (1) Exponent bits contribute slightly more to the model performance than mantissa bits. We provide the optimal exponent-mantissa bit ratio for different bit numbers, which is available for future reference by hardware manufacturers; (2) We discover the formation of the critical data size in low-precision LLM training. Too much training data exceeding the critical data size will inversely bring in degradation of LLM performance; (3) The optimal FP quantization precision is directly proportional to the computational power, but within a wide computational power range. We estimate that the best cost-performance precision should lie between 4-8 bits.

Action understanding has attracted long-term attention. It can be formed as the mapping from the physical space to the semantic space. Typically, researchers built datasets according to idiosyncratic choices to define classes and push the envelope of benchmarks respectively. Datasets are incompatible with each other like “Isolated Islands” due to semantic gaps and various class granularities, e.g., do housework in dataset A and wash plate in dataset B. We argue that we need a more principled semantic space to concentrate the community efforts and use all datasets together to pursue generalizable action learning. To this end, we design a structured action semantic space in view of verb taxonomy hierarchy and covering massive actions. By aligning the classes of previous datasets to our semantic space, we gather (image/video/skeleton/McCap] datasets into a unified database in a unified label system, i.e., bridging “isolated islands” into a “Pangea”. Accordingly, we propose a novel model mapping from the physical space to semantic space to fully use Pangea. In extensive experiments, our new system shows significant superiority, especially in transfer learning. Our code and data will be made public at https://mvig-rhos.com/pangea.

Thermal issue is a major concern in 3D integrated circuit (IC) design. Thermal optimization of 3D IC often requires massive expensive PDE simulations. Neural network-based thermal prediction models can perform real-time prediction for many unseen new designs. However, existing works either solve 2D temperature fields only or do not generalize well to new designs with unseen design configurations (e.g., heat sources and boundary conditions). In this paper, for the first time, we propose DeepOHeat, a physics-aware operator learning framework to predict the temperature field of a family of heat equations with multiple parametric or non-parametric design configurations. This framework learns a functional map from the function space of multiple key PDE configurations (e.g., boundary conditions, power maps, heat transfer coefficients) to the function space of the corresponding solution (i.e., temperature fields), enabling fast thermal analysis and optimization by changing key design configurations (rather than just some parameters). We test DeepOHeat on some industrial design cases and compare it against Celsius 3D from Cadence Design Systems. Our results show that, for the unseen testing cases, a well-trained DeepOHeat can produce accurate results with 1000× to 300000× speedup.

This paper presents Systolic-CNN, an OpenCLdefined scalable, run-time-flexible FPGA accelerator architecture, optimized for performing the low-latency, energy-efficient inference of various convolutional neural networks (CNNs) in the context of multi-tenancy cloud/edge computing. Systolic-CNN adopts a highly pipelined and parallelized 1-D systolic array architecture, which efficiently explores both spatial and temporal parallelism for accelerating CNN inference on FPGAs. SystolicCNN is highly scalable and parameterized, which can be easily adapted by users to achieve 100% utilization of the coarsegrained computation resources (i.e., DSP blocks) for a given FPGA. In addition, Systolic-CNN is run-time-flexible, which can be time-shared, in the context of multi-tenancy cloud or edge computing, to accelerate a variety of CNN models at run time without the need of recompiling the FPGA kernel hardware nor reprogramming the FPGA. The experiment results based on an Intel Arria 10 GX FPGA Development board show that Systolic-CNN, when mapped with a single-precision data format, can achieve 100% utilization of the DSP block resource and an average inference latency of 10ms, 84ms, 1615ms, and 990ms per image for accelerating AlexNet, ResNet-50, RetinaNet, and Light-weight RetinaNet, respectively. The peak computational throughput is measured at 80–170 GFLOPS/s across the acceleration of different CNN models. Codes are available at https://github.com/PSCLab-ASU/SystolicCNN.

A new approach for real-time scene text recognition is proposed in this paper. A novel binary convolutional encoder-decoder network (B-CEDNet) together with a bidirectional recurrent neural network (Bi-RNN). The B-CEDNet is engaged as a visual front-end to provide elaborated character detection, and a back-end Bi-RNN performs character-level sequential correction and classification based on learned contextual knowledge. The front-end B-CEDNet can process multiple regions containing characters using a one-off forward operation, and is trained under binary constraints with significant compression. Hence it leads to both remarkable inference run-time speedup as well as memory usage reduction. With the elaborated character detection, the back-end Bi-RNN merely processes a low dimension feature sequence with category and spatial information of extracted characters for sequence correction and classification. By training with over 1,000,000 synthetic scene text images, the B-CEDNet achieves a recall rate of 0.86, precision of 0.88 and F-score of 0.87 on ICDAR-03 and ICDAR-13. With the correction and classification by Bi-RNN, the proposed real-time scene text recognition achieves state-of-the-art accuracy while only consumes less than 1-ms inference run-time. The flow processing flow is realized on GPU with a small network size of 1.01 MB for B-CEDNet and 3.23 MB for Bi-RNN, which is much faster and smaller than the existing solutions.

FPGA-based hardware accelerators for convolutional neural networks (CNNs) have received attention due to their higher energy efficiency than GPUs. However, it is challenging for FPGA-based solutions to achieve a higher throughput than GPU counterparts. In this article, we demonstrate that FPGA acceleration can be a superior solution in terms of both throughput and energy efficiency when a CNN is trained with binary constraints on weights and activations. Specifically, we propose an optimized fully mapped FPGA accelerator architecture tailored for bitwise convolution and normalization that features massive spatial parallelism with deep pipelines stages. A key advantage of the FPGA accelerator is that its performance is insensitive to data batch size, while the performance of GPU acceleration varies largely depending on the batch size of the data. Experiment results show that the proposed accelerator architecture for binary CNNs running on a Virtex-7 FPGA is 8.3× faster and 75× more energy-efficient than a Titan X GPU for processing online individual requests in small batch sizes. For processing static data in large batch sizes, the proposed solution is on a par with a Titan X GPU in terms of throughput while delivering 9.5× higher energy efficiency.

Compressive sensing (CS) is a promising technology for realizing energy-efficient wireless sensors for long-term health monitoring. However, conventional model-driven CS frameworks suffer from limited compression ratio and reconstruction quality when dealing with physiological signals due to inaccurate models and the overlook of individual variability. In this paper, we propose a data-driven CS framework that can learn signal characteristics and personalized features from any individual recording of physiologic signals to enhance CS performance with a minimized number of measurements. Such improvements are accomplished by a co-training approach that optimizes the sensing matrix and the dictionary towards improved restricted isometry property and signal sparsity, respectively. Experimental results upon ECG signals show that the proposed method, at a compression ratio of 10×, successfully reduces the isometry constant of the trained sensing matrices by 86% against random matrices and improves the overall reconstructed signal-to-noise ratio by 15dB over conventional model-driven approaches.

Wireless body area network (WBAN) is emerging in the mobile healthcare area to replace the traditional wire-connected monitoring devices. As wireless data transmission dominates power cost of sensor nodes, it is beneficial to reduce the data size without much information loss. Compressive sensing (CS) is a perfect candidate to achieve this goal compared to existing compression techniques. In this paper, we proposed a general framework that utilize CS and online dictionary learning (ODL) together. The learned dictionary carries individual characteristics of the original signal, under which the signal has an even sparser representation compared to pre-determined dictionaries. As a consequence, the compression ratio is effectively improved by 2-4× comparing to prior works. Besides, the proposed framework offloads pre-processing from sensor nodes to the server node prior to dictionary learning, providing further reduction in hardware costs. As it is data driven, the proposed framework has the potential to be used with a wide range of physiological signals.