As a natural disaster with high suddenness and great destructiveness, fire has long posed a major threat to human society and ecological environment. In recent years, with the rapid development of smart city and Internet of Things (IoT) technologies, fire detection systems based on deep learning have gradually become a key means to cope with fire hazards. However, existing fire detection models still have many challenges in terms of detection accuracy and real-time performance in complex contexts. To address these issues, we propose two key modules: EAConv (Efficient Attention Convolution) and EADown (Efficient Attention Downsampling). The EAConv module significantly improves the feature extraction efficiency by combining an efficient attention mechanism with depth-separable convolution, while the EADown module enhances the accuracy and efficiency of feature downsampling by utilizing spatial and channel attention mechanisms in combination with pooling operations. Based on these two modules, we design an efficient and lightweight flame detection model, EFA-YOLO (Efficient Feature Attention YOLO). Experimental results show that EFA-YOLO has a model parameter quantity of only 1.4M, GFLOPs of 4.6, and the inference time per image on the CPU is only 22.19 ms. Compared with existing mainstream models (e.g., YOLOv5, YOLOv8, YOLOv9, and YOLOv10), EFA-YOLO exhibits a significant enhancement in detection accuracy (mAP) and inference speed, with model parameter amount is reduced by 94.6 and the inference speed is improved by 88 times.
We propose a Mamba accelerator with reconfigurable architecture, MARCA.We propose three novel approaches in this paper. (1) Reduction alternative PE array architecture for both linear and element-wise operations. For linear operations, the reduction tree connected to PE arrays is enabled and executes the reduction operation. For element-wise operations, the reduction tree is disabled and the output bypasses. (2) Reusable nonlinear function unit based on the reconfigurable PE. We decompose the exponential function into element-wise operations and a shift operation by a fast biased exponential algorithm, and the activation function (SiLU) into a range detection and element-wise operations by a piecewise approximation algorithm. Thus, the reconfigurable PEs are reused to execute nonlinear functions with negligible accuracy loss.(3) Intra-operation and inter-operation buffer management strategy. We propose intra-operation buffer management strategy to maximize input data sharing for linear operations within operations, and inter-operation strategy for element-wise operations between operations. We conduct extensive experiments on Mamba model families with different sizes.MARCA achieves up to 463.22$\times$/11.66$\times$ speedup and up to 9761.42$\times$/242.52$\times$ energy efficiency compared to Intel Xeon 8358P CPU and NVIDIA Tesla A100 GPU implementations, respectively.
Face recognition in the wild is now advancing towards light-weight models, fast inference speed and resolution-adapted capability. In this paper, we propose a bridge distillation approach to turn a complex face model pretrained on private high-resolution faces into a light-weight one for low-resolution face recognition. In our approach, such a cross-dataset resolution-adapted knowledge transfer problem is solved via two-step distillation. In the first step, we conduct cross-dataset distillation to transfer the prior knowledge from private high-resolution faces to public high-resolution faces and generate compact and discriminative features. In the second step, the resolution-adapted distillation is conducted to further transfer the prior knowledge to synthetic low-resolution faces via multi-task learning. By learning low-resolution face representations and mimicking the adapted high-resolution knowledge, a light-weight student model can be constructed with high efficiency and promising accuracy in recognizing low-resolution faces. Experimental results show that the student model performs impressively in recognizing low-resolution faces with only 0.21M parameters and 0.057MB memory. Meanwhile, its speed reaches up to 14,705, ~934 and 763 faces per second on GPU, CPU and mobile phone, respectively.
Graph neural networks (GNNs) are a type of neural network capable of learning on graph-structured data. However, training GNNs on large-scale graphs is challenging due to iterative aggregations of high-dimensional features from neighboring vertices within sparse graph structures combined with neural network operations. The sparsity of graphs frequently results in suboptimal memory access patterns and longer training time. Graph reordering is an optimization strategy aiming to improve the graph data layout. It has shown to be effective to speed up graph analytics workloads, but its effect on the performance of GNN training has not been investigated yet. The generalization of reordering to GNN performance is nontrivial, as multiple aspects must be considered: GNN hyper-parameters such as the number of layers, the number of hidden dimensions, and the feature size used in the GNN model, neural network operations, large intermediate vertex states, and GPU acceleration. In our work, we close this gap by performing an empirical evaluation of 12 reordering strategies in two state-of-the-art GNN systems, PyTorch Geometric and Deep Graph Library. Our results show that graph reordering is effective in reducing training time for CPU- and GPU-based training, respectively. Further, we find that GNN hyper-parameters influence the effectiveness of reordering, that reordering metrics play an important role in selecting a reordering strategy, that lightweight reordering performs better for GPU-based than for CPU-based training, and that invested reordering time can in many cases be amortized.
Advancements in deep learning and voice-activated technologies have driven the development of human-vehicle interaction. Distributed microphone arrays are widely used in in-car scenarios because they can accurately capture the voices of passengers from different speech zones. However, the increase in the number of audio channels, coupled with the limited computational resources and low latency requirements of in-car systems, presents challenges for in-car multi-channel speech separation. To migrate the problems, we propose a lightweight framework that cascades digital signal processing (DSP) and neural networks (NN). We utilize fixed beamforming (BF) to reduce computational costs and independent vector analysis (IVA) to provide spatial prior. We employ dual encoders for dual-branch modeling, with spatial encoder capturing spatial cues and spectral encoder preserving spectral information, facilitating spatial-spectral fusion. Our proposed system supports both streaming and non-streaming modes. Experimental results demonstrate the superiority of the proposed system across various metrics. With only 0.83M parameters and 0.39 real-time factor (RTF) on an Intel Core i7 (2.6GHz) CPU, it effectively separates speech into distinct speech zones. Our demos are available at https://honee-w.github.io/DualSep/.
Transformer-based Large Language Models (LLMs) have become increasingly important. However, due to the quadratic time complexity of attention computation, scaling LLMs to longer contexts incurs extremely slow inference speed and high GPU memory consumption for caching key-value (KV) vectors. This paper proposes RetrievalAttention, a training-free approach to both accelerate attention computation and reduce GPU memory consumption. By leveraging the dynamic sparsity of attention mechanism, RetrievalAttention proposes to build approximate nearest neighbor search (ANNS) indexes for KV vectors in CPU memory and retrieve the most relevant ones through vector search during generation. Unfortunately, we observe that the off-the-shelf ANNS indexes are often ineffective for such retrieval tasks due to the out-of-distribution (OOD) between query vectors and key vectors in the attention mechanism. RetrievalAttention addresses the OOD challenge by designing an attention-aware vector search algorithm that can adapt to the distribution of query vectors. Our evaluation demonstrates that RetrievalAttention achieves near full attention accuracy while only requiring access to 1--3% of the data. This leads to a significant reduction in the inference cost of long-context LLMs, with a much lower GPU memory footprint. In particular, RetrievalAttention only needs a single NVIDIA RTX4090 (24GB) to serve 128K tokens for LLMs with 8B parameters, which is capable of generating one token in 0.188 seconds.
Training Large Language Models (LLMs) requires immense computational power and vast amounts of data. As a result, protecting the intellectual property of these models through fingerprinting is essential for ownership authentication. While adding fingerprints to LLMs through fine-tuning has been attempted, it remains costly and unscalable. In this paper, we introduce FP-VEC, a pilot study on using fingerprint vectors as an efficient fingerprinting method for LLMs. Our approach generates a fingerprint vector that represents a confidential signature embedded in the model, allowing the same fingerprint to be seamlessly incorporated into an unlimited number of LLMs via vector addition. Results on several LLMs show that FP-VEC is lightweight by running on CPU-only devices for fingerprinting, scalable with a single training and unlimited fingerprinting process, and preserves the model's normal behavior. The project page is available at https://fingerprintvector.github.io .
This paper delves into the challenging task of Active Speaker Detection (ASD), where the system needs to determine in real-time whether a person is speaking or not in a series of video frames. While previous works have made significant strides in improving network architectures and learning effective representations for ASD, a critical gap exists in the exploration of real-time system deployment. Existing models often suffer from high latency and memory usage, rendering them impractical for immediate applications. To bridge this gap, we present two scenarios that address the key challenges posed by real-time constraints. First, we introduce a method to limit the number of future context frames utilized by the ASD model. By doing so, we alleviate the need for processing the entire sequence of future frames before a decision is made, significantly reducing latency. Second, we propose a more stringent constraint that limits the total number of past frames the model can access during inference. This tackles the persistent memory issues associated with running streaming ASD systems. Beyond these theoretical frameworks, we conduct extensive experiments to validate our approach. Our results demonstrate that constrained transformer models can achieve performance comparable to or even better than state-of-the-art recurrent models, such as uni-directional GRUs, with a significantly reduced number of context frames. Moreover, we shed light on the temporal memory requirements of ASD systems, revealing that larger past context has a more profound impact on accuracy than future context. When profiling on a CPU we find that our efficient architecture is memory bound by the amount of past context it can use and that the compute cost is negligible as compared to the memory cost.
Singing Voice Conversion (SVC) has emerged as a significant subfield of Voice Conversion (VC), enabling the transformation of one singer's voice into another while preserving musical elements such as melody, rhythm, and timbre. Traditional SVC methods have limitations in terms of audio quality, data requirements, and computational complexity. In this paper, we propose LHQ-SVC, a lightweight, CPU-compatible model based on the SVC framework and diffusion model, designed to reduce model size and computational demand without sacrificing performance. We incorporate features to improve inference quality, and optimize for CPU execution by using performance tuning tools and parallel computing frameworks. Our experiments demonstrate that LHQ-SVC maintains competitive performance, with significant improvements in processing speed and efficiency across different devices. The results suggest that LHQ-SVC can meet
Wearable Internet of Things (IoT) devices are gaining ground for continuous physiological data acquisition and health monitoring. These physiological signals can be used for security applications to achieve continuous authentication and user convenience due to passive data acquisition. This paper investigates an electrocardiogram (ECG) based biometric user authentication system using features derived from the Convolutional Neural Network (CNN) and self-supervised contrastive learning. Contrastive learning enables us to use large unlabeled datasets to train the model and establish its generalizability. We propose approaches enabling the CNN encoder to extract appropriate features that distinguish the user from other subjects. When evaluated using the PTB ECG database with 290 subjects, the proposed technique achieved an authentication accuracy of 99.15%. To test its generalizability, we applied the model to two new datasets, the MIT-BIH Arrhythmia Database and the ECG-ID Database, achieving over 98.5% accuracy without any modifications. Furthermore, we show that repeating the authentication step three times can increase accuracy to nearly 100% for both PTBDB and ECGIDDB. This paper also presents model optimizations for embedded device deployment, which makes the system more relevant to real-world scenarios. To deploy our model in IoT edge sensors, we optimized the model complexity by applying quantization and pruning. The optimized model achieves 98.67% accuracy on PTBDB, with 0.48% accuracy loss and 62.6% CPU cycles compared to the unoptimized model. An accuracy-vs-time-complexity tradeoff analysis is performed, and results are presented for different optimization levels.
We present recent progress in the design and development of DEPLOYERS, an agent-based macroeconomics modeling (ABM) framework, capable to deploy and simulate a full economic system (individual workers, goods and services firms, government, central and private banks, financial market, external sectors) whose structure and activity analysis reproduce the desired calibration data, that can be, for example a Social Accounting Matrix (SAM) or a Supply-Use Table (SUT) or an Input-Output Table (IOT).Here we extend our previous work to a multi-country version and show an example using data from a 46-countries 64-sectors FIGARO Inter-Country IOT. The simulation of each country runs on a separate thread or CPU core to simulate the activity of one step (month, week, or day) and then interacts (updates imports, exports, transfer) with that country's foreign partners, and proceeds to the next step. This interaction can be chosen to be aggregated (a single row and column IO account) or disaggregated (64 rows and columns) with each partner. A typical run simulates thousands of individuals and firms engaged in their monthly activity and then records the results, much like a survey of the country's economic system. This data can then be subjected to, for example, an Input-Output analysis to find out the sources of observed stylized effects as a function of time in the detailed and realistic modeling environment that can be easily implemented in an ABM framework.
This is the 1st part of the dissertation for my master degree and compares the power consumption using the default floating point (32bit) and Nvidia mixed precision (16bit and 32bit) while training a classification ML model. A custom PC with specific hardware was built to perform the experiments, and different ML hyper-parameters, such as batch size, neurons, and epochs, were chosen to build Deep Neural Networks (DNN). Additionally, various software was used during the experiments to collect the power consumption data in Watts from the Graphics Processing Unit (GPU), Central Processing Unit (CPU), Random Access Memory (RAM) and manually from a wattmeter connected to the wall. A benchmarking test with default hyper parameter values for the DNN was used as a reference, while the experiments used a combination of different settings. The results were recorded in Excel, and descriptive statistics were chosen to calculate the mean between the groups and compare them using graphs and tables. The outcome was positive when using mixed precision combined with specific hyper-parameters. Compared to the benchmarking, the optimisation for the classification reduced the power consumption between 7 and 11 Watts. Similarly, the carbon footprint is reduced because the calculation uses the same power consumption data. Still, a consideration is required when configuring hyper-parameters because it can negatively affect hardware performance. However, this research required inferential statistics, specifically ANOVA and T-test, to compare the relationship between the means. Furthermore, tests indicated no statistical significance of the relationship between the benchmarking and experiments. However, a more extensive implementation with a cluster of GPUs can increase the sample size significantly, as it is an essential factor and can change the outcome of the statistical analysis.
Monotonic alignment search (MAS), introduced by Glow-TTS, is one of the most popular algorithm in TTS to estimate unknown alignments between text and speech. Since this algorithm needs to search for the most probable alignment with dynamic programming by caching all paths, the time complexity of the algorithm is $O(T \times S)$. The authors of Glow-TTS run this algorithm on CPU, and while they mentioned it is difficult to parallelize, we found that MAS can be parallelized in text-length dimension and CPU execution consumes an inordinate amount of time for inter-device copy. Therefore, we implemented a Triton kernel and PyTorch JIT script to accelerate MAS on GPU without inter-device copy. As a result, Super-MAS Triton kernel is up to 72 times faster in the extreme-length case. The code is available at \url{https://github.com/supertone-inc/super-monotonic-align}.
The Audio Question Answering (AQA) task includes audio event classification, audio captioning, and open-ended reasoning. Recently, AQA has garnered attention due to the advent of Large Audio Language Models (LALMs). Current literature focuses on constructing LALMs by integrating audio encoders with text-only Large Language Models (LLMs) through a projection module. While LALMs excel in general audio understanding, they are limited in temporal reasoning, which may hinder their commercial applications and on-device deployment. This paper addresses these challenges and limitations in audio temporal reasoning. First, we introduce a data augmentation technique for generating reliable audio temporal questions and answers using an LLM. Second, we perform a further fine-tuning of an existing baseline using curriculum learning strategy to specialize in temporal reasoning without compromising performance on fine-tuned tasks. We demonstrate the performance of our model using state-of-the-art LALMs on public audio benchmark datasets. Third, we implement our AQA model on-device locally and investigate its CPU inference for edge applications.
This report evaluates the performance impact of enabling Trusted Execution Environments (TEE) on NVIDIA Hopper GPUs for large language model (LLM) inference tasks. We benchmark the overhead introduced by TEE mode across various LLMs and token lengths, with a particular focus on the bottleneck caused by CPU-GPU data transfers via PCIe. Our results indicate that while there is minimal computational overhead within the GPU, the overall performance penalty is primarily attributable to data transfer. For the majority of typical LLM queries, the overhead remains below 7%, with larger models and longer sequences experiencing nearly zero overhead.
This paper investigates the RWKV model's efficacy in content moderation through targeted experimentation. We introduce a novel dataset specifically designed for distillation into smaller models, enhancing content moderation practices. This comprehensive dataset encompasses images, videos, sounds, and text data that present societal challenges. Leveraging advanced Large Language Models (LLMs), we generated an extensive set of responses -- 558,958 for text and 83,625 for images -- to train and refine content moderation systems. Our core experimentation involved fine-tuning the RWKV model, capitalizing on its CPU-efficient architecture to address large-scale content moderation tasks. By highlighting the dataset's potential for knowledge distillation, this study not only demonstrates RWKV's capability in improving the accuracy and efficiency of content moderation systems but also paves the way for developing more compact, resource-efficient models in this domain. Datasets and models can be found in HuggingFace: https://huggingface.co/modrwkv
Research in efficient vision backbones is evolving into models that are a mixture of convolutions and transformer blocks. A smart combination of both, architecture-wise and component-wise is mandatory to excel in the speedaccuracy trade-off. Most publications focus on maximizing accuracy and utilize MACs (multiply accumulate operations) as an efficiency metric. The latter however often do not measure accurately how fast a model actually is due to factors like memory access cost and degree of parallelism. We analyzed common modules and architectural design choices for backbones not in terms of MACs, but rather in actual throughput and latency, as the combination of the latter two is a better representation of the efficiency of models in real applications. We applied the conclusions taken from that analysis to create a recipe for increasing hardware-efficiency in macro design. Additionally we introduce a simple slimmed-down version of MultiHead Self-Attention, that aligns with our analysis. We combine both macro and micro design to create a new family of hardware-efficient backbone networks called LowFormer. LowFormer achieves a remarkable speedup in terms of throughput and latency, while achieving similar or better accuracy than current state-of-the-art efficient backbones. In order to prove the generalizability of our hardware-efficient design, we evaluate our method on GPU, mobile GPU and ARM CPU. We further show that the downstream tasks object detection and semantic segmentation profit from our hardware-efficient architecture. Code and models are available at https://github.com/ altair199797/LowFormer.
Recent advancements in large language models (LLMs) have significantly enhanced their capacity to aggregate and process information across multiple modalities, enabling them to perform a wide range of tasks such as multimodal data querying, tool usage, web interactions, and handling long documents. These capabilities pave the way for transforming LLMs from mere chatbots into general-purpose agents capable of interacting with the real world. This paper explores the concept of using a language model as the core component of an operating system (OS), effectively acting as a CPU that processes data stored in a context window, which functions as RAM. A key challenge in realizing such an LM OS is managing the life-long context and ensuring statefulness across sessions, a feature limited by the current session-based interaction paradigm due to context window size limit. To address this, we introduce compressor-retriever, a model-agnostic architecture designed for life-long context management. Unlike other long-context solutions such as retrieval-augmented generation, our approach exclusively uses the base model's forward function to compress and retrieve context, ensuring end-to-end differentiability. Preliminary experiments demonstrate the effectiveness of this architecture in in-context learning tasks, marking a step towards the development of a fully stateful LLM OS. Project repo available at: https://github.com/gblackout/LM-OS
Out-of-distribution (OOD) detectors can act as safety monitors in embedded cyber-physical systems by identifying samples outside a machine learning model's training distribution to prevent potentially unsafe actions. However, OOD detectors are often implemented using deep neural networks, which makes it difficult to meet real-time deadlines on embedded systems with memory and power constraints. We consider the class of variational autoencoder (VAE) based OOD detectors where OOD detection is performed in latent space, and apply quantization, pruning, and knowledge distillation. These techniques have been explored for other deep models, but no work has considered their combined effect on latent space OOD detection. While these techniques increase the VAE's test loss, this does not correspond to a proportional decrease in OOD detection performance and we leverage this to develop lean OOD detectors capable of real-time inference on embedded CPUs and GPUs. We propose a design methodology that combines all three compression techniques and yields a significant decrease in memory and execution time while maintaining AUROC for a given OOD detector. We demonstrate this methodology with two existing OOD detectors on a Jetson Nano and reduce GPU and CPU inference time by 20% and 28% respectively while keeping AUROC within 5% of the baseline.
We exploit the widening margin in tensor-core performance between [FP64/FP32/FP16/INT8,FP64/FP32/FP16/FP8/INT8] on NVIDIA [Ampere,Hopper] GPUs to boost the performance of output accuracy-preserving mixed-precision computation of Genome-Wide Association Studies (GWAS) of 305K patients from the UK BioBank, the largest-ever GWAS cohort studied for genetic epistasis using a multivariate approach. Tile-centric adaptive-precision linear algebraic techniques motivated by reducing data motion gain enhanced significance with low-precision GPU arithmetic. At the core of Kernel Ridge Regression (KRR) techniques for GWAS lie compute-bound cubic-complexity matrix operations that inhibit scaling to aspirational dimensions of the population, genotypes, and phenotypes. We accelerate KRR matrix generation by redesigning the computation for Euclidean distances to engage INT8 tensor cores while exploiting symmetry.We accelerate solution of the regularized KRR systems by deploying a new four-precision Cholesky-based solver, which, at 1.805 mixed-precision ExaOp/s on a nearly full Alps system, outperforms the state-of-the-art CPU-only REGENIE GWAS software by five orders of magnitude.