Deep reinforcement learning (DRL) has achieved remarkable success across various domains, such as video games, robotics, and, recently, large language models. However, the computational costs and memory requirements of DRL models often limit their deployment in resource-constrained environments. The challenge underscores the urgent need to explore neural network compression methods to make RDL models more practical and broadly applicable. Our study investigates the impact of two prominent compression methods, quantization and pruning on DRL models. We examine how these techniques influence four performance factors: average return, memory, inference time, and battery utilization across various DRL algorithms and environments. Despite the decrease in model size, we identify that these compression techniques generally do not improve the energy efficiency of DRL models, but the model size decreases. We provide insights into the trade-offs between model compression and DRL performance, offering guidelines for deploying efficient DRL models in resource-constrained settings.
In real applications of Reinforcement Learning (RL), such as robotics, low latency and energy efficient inference is very desired. The use of sparsity and pruning for optimizing Neural Network inference, and particularly to improve energy and latency efficiency, is a standard technique. In this work, we perform a systematic investigation of applying these optimization techniques for different RL algorithms in different RL environments, yielding up to a 400-fold reduction in the size of neural networks.
The development of model compression is continuously motivated by the evolution of various neural network accelerators with ASIC or FPGA. On the algorithm side, the ultimate goal of quantization or pruning is accelerating the expensive DNN computations on low-power hardware. However, such a "design-and-deploy" workflow faces under-explored challenges in the current hardware-algorithm co-design community. First, although the state-of-the-art quantization algorithm can achieve low precision with negligible degradation of accuracy, the latest deep learning framework (e.g., PyTorch) can only support non-customizable 8-bit precision, data format, and parameter extraction. Secondly, the objective of quantization is to enable the computation with low-precision data. However, the current SoTA algorithm treats the quantized integer as an intermediate result, while the final output of the quantizer is the "discretized" floating-point values, ignoring the practical needs and adding additional workload to hardware designers for integer parameter extraction and layer fusion. Finally, the compression toolkits designed by the industry are constrained to their in-house product or a handful of algorithms. The limited degree of freedom in the current toolkit and the under-explored customization hinder the prototype ASIC or FPGA-based accelerator design. To resolve these challenges, we propose Torch2Chip, an open-sourced, fully customizable, and high-performance toolkit that supports user-designed compression followed by automatic model fusion and parameter extraction. Torch2Chip incorporates the hierarchical design workflow, and the user-customized compression algorithm will be directly packed into the deployment-ready format for prototype chip verification with either CNN or vision transformer (ViT). The code is available at https://github.com/SeoLabCornell/torch2chip.
Despite the impressive performance of deep neural networks (DNNs), their computational complexity and storage space consumption have led to the concept of network compression. While DNN compression techniques such as pruning and low-rank decomposition have been extensively studied, there has been insufficient attention paid to their theoretical explanation. In this paper, we propose a novel theoretical framework that leverages a probabilistic latent space of DNN weights and explains the optimal network sparsity by using the information-theoretic divergence measures. We introduce new analogous projected patterns (AP2) and analogous-in-probability projected patterns (AP3) notions for DNNs and prove that there exists a relationship between AP3/AP2 property of layers in the network and its performance. Further, we provide a theoretical analysis that explains the training process of the compressed network. The theoretical results are empirically validated through experiments conducted on standard pre-trained benchmarks, including AlexNet, ResNet50, and VGG16, using CIFAR10 and CIFAR100 datasets. Through our experiments, we highlight the relationship of AP3 and AP2 properties with fine-tuning pruned DNNs and sparsity levels.
Representing the Neural Radiance Field (NeRF) with the explicit voxel grid (EVG) is a promising direction for improving NeRFs. However, the EVG representation is not efficient for storage and transmission because of the terrific memory cost. Current methods for compressing EVG mainly inherit the methods designed for neural network compression, such as pruning and quantization, which do not take full advantage of the spatial correlation of voxels. Inspired by prosperous digital image compression techniques, this paper proposes SPC-NeRF, a novel framework applying spatial predictive coding in EVG compression. The proposed framework can remove spatial redundancy efficiently for better compression performance.Moreover, we model the bitrate and design a novel form of the loss function, where we can jointly optimize compression ratio and distortion to achieve higher coding efficiency. Extensive experiments demonstrate that our method can achieve 32% bit saving compared to the state-of-the-art method VQRF on multiple representative test datasets, with comparable training time.
Structural pruning of neural networks conventionally relies on identifying and discarding less important neurons, a practice often resulting in significant accuracy loss that necessitates subsequent fine-tuning efforts. This paper introduces a novel approach named Intra-Fusion, challenging this prevailing pruning paradigm. Unlike existing methods that focus on designing meaningful neuron importance metrics, Intra-Fusion redefines the overlying pruning procedure. Through utilizing the concepts of model fusion and Optimal Transport, we leverage an agnostically given importance metric to arrive at a more effective sparse model representation. Notably, our approach achieves substantial accuracy recovery without the need for resource-intensive fine-tuning, making it an efficient and promising tool for neural network compression. Additionally, we explore how fusion can be added to the pruning process to significantly decrease the training time while maintaining competitive performance. We benchmark our results for various networks on commonly used datasets such as CIFAR-10, CIFAR-100, and ImageNet. More broadly, we hope that the proposed Intra-Fusion approach invigorates exploration into a fresh alternative to the predominant compression approaches. Our code is available here: https://github.com/alexandertheus/Intra-Fusion.
Neural network compression techniques, such as knowledge distillation (KD) and network pruning, have received increasing attention. Recent work `Prune, then Distill' reveals that a pruned student-friendly teacher network can benefit the performance of KD. However, the conventional teacher-student pipeline, which entails cumbersome pre-training of the teacher and complicated compression steps, makes pruning with KD less efficient. In addition to compressing models, recent compression techniques also emphasize the aspect of efficiency. Early pruning demands significantly less computational cost in comparison to the conventional pruning methods as it does not require a large pre-trained model. Likewise, a special case of KD, known as self-distillation (SD), is more efficient since it requires no pre-training or student-teacher pair selection. This inspires us to collaborate early pruning with SD for efficient model compression. In this work, we propose the framework named Early Pruning with Self-Distillation (EPSD), which identifies and preserves distillable weights in early pruning for a given SD task. EPSD efficiently combines early pruning and self-distillation in a two-step process, maintaining the pruned network's trainability for compression. Instead of a simple combination of pruning and SD, EPSD enables the pruned network to favor SD by keeping more distillable weights before training to ensure better distillation of the pruned network. We demonstrated that EPSD improves the training of pruned networks, supported by visual and quantitative analyses. Our evaluation covered diverse benchmarks (CIFAR-10/100, Tiny-ImageNet, full ImageNet, CUB-200-2011, and Pascal VOC), with EPSD outperforming advanced pruning and SD techniques.
While Convolutional Neural Networks (CNNs) excel at learning complex latent-space representations, their over-parameterization can lead to overfitting and reduced performance, particularly with limited data. This, alongside their high computational and memory demands, limits the applicability of CNNs for edge deployment. Low-rank matrix approximation has emerged as a promising approach to reduce CNN parameters, but its application presents challenges including rank selection and performance loss. To address these issues, we propose an efficient training method for CNN compression via dynamic parameter rank pruning. Our approach integrates efficient matrix factorization and novel regularization techniques, forming a robust framework for dynamic rank reduction and model compression. We use Singular Value Decomposition (SVD) to model low-rank convolutional filters and dense weight matrices and we achieve model compression by training the SVD factors with back-propagation in an end-to-end way. We evaluate our method on an array of modern CNNs, including ResNet-18, ResNet-20, and ResNet-32, and datasets like CIFAR-10, CIFAR-100, and ImageNet (2012), showcasing its applicability in computer vision. Our experiments show that the proposed method can yield substantial storage savings while maintaining or even enhancing classification performance.
Weight-sharing plays a significant role in the success of many deep neural networks, by increasing memory efficiency and incorporating useful inductive priors about the problem into the network. But understanding how weight-sharing can be used effectively in general is a topic that has not been studied extensively. Chen et al. [2015] proposed HashedNets, which augments a multi-layer perceptron with a hash table, as a method for neural network compression. We generalize this method into a framework (ArbNets) that allows for efficient arbitrary weight-sharing, and use it to study the role of weight-sharing in neural networks. We show that common neural networks can be expressed as ArbNets with different hash functions. We also present two novel hash functions, the Dirichlet hash and the Neighborhood hash, and use them to demonstrate experimentally that balanced and deterministic weight-sharing helps with the performance of a neural network.
Graph Neural Networks (GNNs) have proven to be quite versatile for a variety of applications, including recommendation systems, fake news detection, drug discovery, and even computer vision. Due to the expanding size of graph-structured data, GNN models have also increased in complexity, leading to substantial latency issues. This is primarily attributed to the irregular structure of graph data and its access pattern into memory. The natural solution to reduce latency is to compress large GNNs into small GNNs. One way to do this is via knowledge distillation (KD). However, most KD approaches for GNNs only consider the outputs of the last layers and do not consider the outputs of the intermediate layers of the GNNs; these layers may contain important inductive biases indicated by the graph structure. To address this shortcoming, we propose a novel KD approach to GNN compression that we call Attention-Based Knowledge Distillation (ABKD). ABKD is a KD approach that uses attention to identify important intermediate teacher-student layer pairs and focuses on aligning their outputs. ABKD enables higher compression of GNNs with a smaller accuracy dropoff compared to existing KD approaches. On average, we achieve a 1.79% increase in accuracy with a 32.3x compression ratio on OGBN-Mag, a large graph dataset, compared to state-of-the-art approaches.
We attribute grokking, the phenomenon where generalization is much delayed after memorization, to compression. To do so, we define linear mapping number (LMN) to measure network complexity, which is a generalized version of linear region number for ReLU networks. LMN can nicely characterize neural network compression before generalization. Although the $L_2$ norm has been a popular choice for characterizing model complexity, we argue in favor of LMN for a number of reasons: (1) LMN can be naturally interpreted as information/computation, while $L_2$ cannot. (2) In the compression phase, LMN has linear relations with test losses, while $L_2$ is correlated with test losses in a complicated nonlinear way. (3) LMN also reveals an intriguing phenomenon of the XOR network switching between two generalization solutions, while $L_2$ does not. Besides explaining grokking, we argue that LMN is a promising candidate as the neural network version of the Kolmogorov complexity since it explicitly considers local or conditioned linear computations aligned with the nature of modern artificial neural networks.
Model quantization, which aims to compress deep neural networks and accelerate inference speed, has greatly facilitated the development of cumbersome models on mobile and edge devices. There is a common assumption in quantization methods from prior works that training data is available. In practice, however, this assumption cannot always be fulfilled due to reasons of privacy and security, rendering these methods inapplicable in real-life situations. Thus, data-free network quantization has recently received significant attention in neural network compression. Causal reasoning provides an intuitive way to model causal relationships to eliminate data-driven correlations, making causality an essential component of analyzing data-free problems. However, causal formulations of data-free quantization are inadequate in the literature. To bridge this gap, we construct a causal graph to model the data generation and discrepancy reduction between the pre-trained and quantized models. Inspired by the causal understanding, we propose the Causality-guided Data-free Network Quantization method, Causal-DFQ, to eliminate the reliance on data via approaching an equilibrium of causality-driven intervened distributions. Specifically, we design a content-style-decoupled generator, synthesizing images conditioned on the relevant and irrelevant factors; then we propose a discrepancy reduction loss to align the intervened distributions of the pre-trained and quantized models. It is worth noting that our work is the first attempt towards introducing causality to data-free quantization problem. Extensive experiments demonstrate the efficacy of Causal-DFQ. The code is available at https://github.com/42Shawn/Causal-DFQ.
Modern deep neural networks, particularly recent large language models, come with massive model sizes that require significant computational and storage resources. To enable the deployment of modern models on resource-constrained environments and accelerate inference time, researchers have increasingly explored pruning techniques as a popular research direction in neural network compression. However, there is a dearth of up-to-date comprehensive review papers on pruning. To address this issue, in this survey, we provide a comprehensive review of existing research works on deep neural network pruning in a taxonomy of 1) universal/specific speedup, 2) when to prune, 3) how to prune, and 4) fusion of pruning and other compression techniques. We then provide a thorough comparative analysis of eight pairs of contrast settings for pruning and explore emerging topics, including pruning for large language models, large multimodal models, post-training pruning, and different supervision levels for pruning to shed light on the commonalities and differences of existing methods and lay the foundation for further method development. To facilitate future research, we build a curated collection of datasets, networks, and evaluations on different applications. Finally, we provide valuable recommendations on selecting pruning methods and prospect several promising research directions. We build a repository at https://github.com/hrcheng1066/awesome-pruning.
Tensor decomposition of convolutional and fully-connected layers is an effective way to reduce parameters and FLOP in neural networks. Due to memory and power consumption limitations of mobile or embedded devices, the quantization step is usually necessary when pre-trained models are deployed. A conventional post-training quantization approach applied to networks with decomposed weights yields a drop in accuracy. This motivated us to develop an algorithm that finds tensor approximation directly with quantized factors and thus benefit from both compression techniques while keeping the prediction quality of the model. Namely, we propose to use Alternating Direction Method of Multipliers (ADMM) for Canonical Polyadic (CP) decomposition with factors whose elements lie on a specified quantization grid. We compress neural network weights with a devised algorithm and evaluate it's prediction quality and performance. We compare our approach to state-of-the-art post-training quantization methods and demonstrate competitive results and high flexibility in achiving a desirable quality-performance tradeoff.
Deep neural networks (DNNs) are state-of-the-art techniques for solving most computer vision problems. DNNs require billions of parameters and operations to achieve state-of-the-art results. This requirement makes DNNs extremely compute, memory, and energy-hungry, and consequently difficult to deploy on small battery-powered Internet-of-Things (IoT) devices with limited computing resources. Deployment of DNNs on Internet-of-Things devices, such as traffic cameras, can improve public safety by enabling applications such as automatic accident detection and emergency response.Through this paper, we survey the recent advances in low-power and energy-efficient DNN implementations that improve the deployability of DNNs without significantly sacrificing accuracy. In general, these techniques either reduce the memory requirements, the number of arithmetic operations, or both. The techniques can be divided into three major categories: neural network compression, network architecture search and design, and compiler and graph optimizations. In this paper, we survey both low-power techniques for both convolutional and transformer DNNs, and summarize the advantages, disadvantages, and open research problems.
Over the past few years, extensive research has been devoted to enhancing YOLO object detectors. Since its introduction, eight major versions of YOLO have been introduced with the purpose of improving its accuracy and efficiency. While the evident merits of YOLO have yielded to its extensive use in many areas, deploying it on resource-limited devices poses challenges. To address this issue, various neural network compression methods have been developed, which fall under three main categories, namely network pruning, quantization, and knowledge distillation. The fruitful outcomes of utilizing model compression methods, such as lowering memory usage and inference time, make them favorable, if not necessary, for deploying large neural networks on hardware-constrained edge devices. In this review paper, our focus is on pruning and quantization due to their comparative modularity. We categorize them and analyze the practical results of applying those methods to YOLOv5. By doing so, we identify gaps in adapting pruning and quantization for compressing YOLOv5, and provide future directions in this area for further exploration. Among several versions of YOLO, we specifically choose YOLOv5 for its excellent trade-off between recency and popularity in literature. This is the first specific review paper that surveys pruning and quantization methods from an implementation point of view on YOLOv5. Our study is also extendable to newer versions of YOLO as implementing them on resource-limited devices poses the same challenges that persist even today. This paper targets those interested in the practical deployment of model compression methods on YOLOv5, and in exploring different compression techniques that can be used for subsequent versions of YOLO.
Pedestrian Attribute Recognition (PAR) deals with the problem of identifying features in a pedestrian image. It has found interesting applications in person retrieval, suspect re-identification and soft biometrics. In the past few years, several Deep Neural Networks (DNNs) have been designed to solve the task; however, the developed DNNs predominantly suffer from over-parameterization and high computational complexity. These problems hinder them from being exploited in resource-constrained embedded devices with limited memory and computational capacity. By reducing a network's layers using effective compression techniques, such as tensor decomposition, neural network compression is an effective method to tackle these problems. We propose novel Lightweight Attribute Localizing Models (LWALM) for Pedestrian Attribute Recognition (PAR). LWALM is a compressed neural network obtained after effective layer-wise compression of the Attribute Localization Model (ALM) using the Canonical Polyadic Decomposition with Error Preserving Correction (CPD-EPC) algorithm.
Neural network compression has been an increasingly important subject, not only due to its practical relevance, but also due to its theoretical implications, as there is an explicit connection between compressibility and generalization error. Recent studies have shown that the choice of the hyperparameters of stochastic gradient descent (SGD) can have an effect on the compressibility of the learned parameter vector. These results, however, rely on unverifiable assumptions and the resulting theory does not provide a practical guideline due to its implicitness. In this study, we propose a simple modification for SGD, such that the outputs of the algorithm will be provably compressible without making any nontrivial assumptions. We consider a one-hidden-layer neural network trained with SGD, and show that if we inject additive heavy-tailed noise to the iterates at each iteration, for any compression rate, there exists a level of overparametrization such that the output of the algorithm will be compressible with high probability. To achieve this result, we make two main technical contributions: (i) we prove a 'propagation of chaos' result for a class of heavy-tailed stochastic differential equations, and (ii) we derive error estimates for their Euler discretization. Our experiments suggest that the proposed approach not only achieves increased compressibility with various models and datasets, but also leads to robust test performance under pruning, even in more realistic architectures that lie beyond our theoretical setting.
Quantization and pruning are two effective Deep Neural Networks model compression methods. In this paper, we propose Automatic Prune Binarization (APB), a novel compression technique combining quantization with pruning. APB enhances the representational capability of binary networks using a few full-precision weights. Our technique jointly maximizes the accuracy of the network while minimizing its memory impact by deciding whether each weight should be binarized or kept in full precision. We show how to efficiently perform a forward pass through layers compressed using APB by decomposing it into a binary and a sparse-dense matrix multiplication. Moreover, we design two novel efficient algorithms for extremely quantized matrix multiplication on CPU, leveraging highly efficient bitwise operations. The proposed algorithms are 6.9x and 1.5x faster than available state-of-the-art solutions. We extensively evaluate APB on two widely adopted model compression datasets, namely CIFAR10 and ImageNet. APB delivers better accuracy/memory trade-off compared to state-of-the-art methods based on i) quantization, ii) pruning, and iii) combination of pruning and quantization. APB outperforms quantization in the accuracy/efficiency trade-off, being up to 2x faster than the 2-bit quantized model with no loss in accuracy.
Network compression is now a mature sub-field of neural network research: over the last decade, significant progress has been made towards reducing the size of models and speeding up inference, while maintaining the classification accuracy. However, many works have observed that focusing on just the overall accuracy can be misguided. E.g., it has been shown that mismatches between the full and compressed models can be biased towards under-represented classes. This raises the important research question, can we achieve network compression while maintaining "semantic equivalence" with the original network? In this work, we study this question in the context of the "long tail" phenomenon in computer vision datasets observed by Feldman, et al. They argue that memorization of certain inputs (appropriately defined) is essential to achieving good generalization. As compression limits the capacity of a network (and hence also its ability to memorize), we study the question: are mismatches between the full and compressed models correlated with the memorized training data? We present positive evidence in this direction for image classification tasks, by considering different base architectures and compression schemes.