With the increasing size of Large Vision-Language Models (LVLMs), network pruning techniques aimed at compressing models for deployment in resource-constrained environments have garnered significant attention. However, we observe that pruning often leads to a degradation in safety performance. To address this issue, we present a novel and lightweight approach, termed Hierarchical Safety Realignment (HSR). HSR operates by first quantifying the contribution of each attention head to safety, identifying the most critical ones, and then selectively restoring neurons directly within these attention heads that play a pivotal role in maintaining safety. This process hierarchically realigns the safety of pruned LVLMs, progressing from the attention head level to the neuron level. We validate HSR across various models and pruning strategies, consistently achieving notable improvements in safety performance. To our knowledge, this is the first work explicitly focused on restoring safety in LVLMs post-pruning.
Artificial neural network pruning is a method in which artificial neural network sizes can be reduced while attempting to preserve the predicting capabilities of the network. This is done to make the model smaller or faster during inference time. In this work we analyze the ability of a selection of artificial neural network pruning methods to generalize to a new cybersecurity dataset utilizing a simpler network type than was designed for. We analyze each method using a variety of pruning degrees to best understand how each algorithm responds to the new environment. This has allowed us to determine the most well fit pruning method of those we searched for the task. Unexpectedly, we have found that many of them do not generalize to the problem well, leaving only a few algorithms working to an acceptable degree.
This study explores a novel approach to neural network pruning using evolutionary computation, focusing on simultaneously pruning the encoder and decoder of an autoencoder. We introduce two new mutation operators that use layer activations to guide weight pruning. Our findings reveal that one of these activation-informed operators outperforms random pruning, resulting in more efficient autoencoders with comparable performance to canonically trained models. Prior work has established that autoencoder training is effective and scalable with a spatial coevolutionary algorithm that cooperatively coevolves a population of encoders with a population of decoders, rather than one autoencoder. We evaluate how the same activity-guided mutation operators transfer to this context. We find that random pruning is better than guided pruning, in the coevolutionary setting. This suggests activation-based guidance proves more effective in low-dimensional pruning environments, where constrained sample spaces can lead to deviations from true uniformity in randomization. Conversely, population-driven strategies enhance robustness by expanding the total pruning dimensionality, achieving statistically uniform randomness that better preserves system dynamics. We experiment with pruning according to different schedules and present best combinations of operator and schedule for the canonical and coevolving populations cases.
We introduce ReplaceMe, a generalized training-free depth pruning method that effectively replaces transformer blocks with a linear operation, while maintaining high performance for low compression ratios. In contrast to conventional pruning approaches that require additional training or fine-tuning, our approach requires only a small calibration dataset that is used to estimate a linear transformation to approximate the pruned blocks. This estimated linear mapping can be seamlessly merged with the remaining transformer blocks, eliminating the need for any additional network parameters. Our experiments show that ReplaceMe consistently outperforms other training-free approaches and remains highly competitive with state-of-the-art pruning methods that involve extensive retraining/fine-tuning and architectural modifications. Applied to several large language models (LLMs), ReplaceMe achieves up to 25% pruning while retaining approximately 90% of the original model's performance on open benchmarks - without any training or healing steps, resulting in minimal computational overhead (see Fig.1). We provide an open-source library implementing ReplaceMe alongside several state-of-the-art depth pruning techniques, available at this repository.
The model is usually kept integral in the mainstream large language model (LLM) fine-tuning protocols. No works have questioned whether maintaining the integrity of the model is indispensable for performance. In this work, we introduce Mask Fine-Tuning (MFT), a brand-new LLM fine-tuning paradigm to show that properly breaking the integrity of the model can surprisingly lead to improved performance. Specifically, MFT learns a set of binary masks supervised by the typical LLM fine-tuning objective. Extensive experiments show that MFT gains a consistent performance boost across various domains and backbones (e.g., 1.95%/1.88% average gain in coding with LLaMA2-7B/3.1-8B). Detailed procedures are provided to study the proposed MFT from different hyperparameter perspectives for better insight. In particular, MFT naturally updates the current LLM training protocol by deploying it on a complete well-trained model. This study extends the functionality of mask learning from its conventional network pruning context for model compression to a more general scope.
Bayesian Neural Networks (BNNs) offer robust uncertainty quantification in model predictions, but training them presents a significant computational challenge. This is mainly due to the problem of sampling multimodal posterior distributions using Markov Chain Monte Carlo (MCMC) sampling and variational inference algorithms. Moreover, the number of model parameters scales exponentially with additional hidden layers, neurons, and features in the dataset. Typically, a significant portion of these densely connected parameters are redundant and pruning a neural network not only improves portability but also has the potential for better generalisation capabilities. In this study, we address some of the challenges by leveraging MCMC sampling with network pruning to obtain compact probabilistic models having removed redundant parameters. We sample the posterior distribution of model parameters (weights and biases) and prune weights with low importance, resulting in a compact model. We ensure that the compact BNN retains its ability to estimate uncertainty via the posterior distribution while retaining the model training and generalisation performance accuracy by adapting post-pruning resampling. We evaluate the effectiveness of our MCMC pruning strategy on selected benchmark datasets for regression and classification problems through empirical result analysis. We also consider two coral reef drill-core lithology classification datasets to test the robustness of the pruning model in complex real-world datasets. We further investigate if refining compact BNN can retain any loss of performance. Our results demonstrate the feasibility of training and pruning BNNs using MCMC whilst retaining generalisation performance with over 75% reduction in network size. This paves the way for developing compact BNN models that provide uncertainty estimates for real-world applications.
We develop a pipeline to streamline neural architecture codesign for physics applications to reduce the need for ML expertise when designing models for novel tasks. Our method employs neural architecture search and network compression in a two-stage approach to discover hardware efficient models. This approach consists of a global search stage that explores a wide range of architectures while considering hardware constraints, followed by a local search stage that fine-tunes and compresses the most promising candidates. We exceed performance on various tasks and show further speedup through model compression techniques such as quantization-aware-training and neural network pruning. We synthesize the optimal models to high level synthesis code for FPGA deployment with the hls4ml library. Additionally, our hierarchical search space provides greater flexibility in optimization, which can easily extend to other tasks and domains. We demonstrate this with two case studies: Bragg peak finding in materials science and jet classification in high energy physics, achieving models with improved accuracy, smaller latencies, or reduced resource utilization relative to the baseline models.
Network pruning focuses on computational techniques that aim to reduce a given model's computational cost by removing a subset of its parameters while having minimal impact on performance. Throughout the last decade, the most widely used pruning paradigm has been pruning and re-training, which nowadays is inconvenient due to the vast amount of pre-trained models, which are in any case too expensive to re-train. In this paper, we exploit functional information from dense pre-trained models, i.e., their activations, to obtain sparse models that maximize the activations' alignment w.r.t. their corresponding dense models. Hence, we propose \textsc{NeuroAL}, a \emph{top-up} algorithm that can be used on top of any given pruning algorithm for LLMs, which modifies the block-wise and row-wise sparsity exploiting information from both the dense model and its sparse version to maximize the \emph{neuron alignment} among activations. Differently from existing methods, our approach adaptively selects the best hyperparameters for the block-wise and row-wise sparsity ratios w.r.t. the model and the desired sparsity, and requires \emph{no re-training}. We test our method over 276 cases combining four LLM families, three sparsity ratios, and ten language tasks (three language modeling and seven zero-shot datasets), showing how it consistently outperforms the latest state-of-the-art methods in terms of performance-runtime trade-off. The code is available at \href{https://github.com/eliacunegatti/NeuroAL}{https://github.com/eliacunegatti/NeuroAL}.
The evolving capabilities of large language models are accompanied by growing sizes and deployment costs, necessitating effective inference optimisation techniques. We propose a novel pruning method utilising centrality measures from graph theory, reducing both the computational requirements and the memory footprint of these models. Specifically, we devise a method for creating a weighted directed acyclical graph representation of multilayer perceptrons to which we apply a modified version of the weighted PageRank centrality measure to compute node importance scores. In combination with uniform pruning this leads to structured sparsity. We call this pruning method MLPRank. Furthermore we introduce an extension to decoder-only transformer models and call it LLMRank. For both variants we demonstrate a strong performance. With MLPRank on average leading to 6.09 % higher accuracy retention than three popular baselines and 13.42 % with LLMRank compared to two popular baselines. Code is available at https://github.com/amazon-science/llm-rank-pruning.
Network pruning has emerged as a potential solution to make LLMs cheaper to deploy. However, existing LLM pruning approaches universally rely on the C4 dataset as the calibration data for calculating pruning scores, leaving its optimality unexplored. In this study, we evaluate the choice of calibration data on LLM pruning, across a wide range of datasets that are most commonly used in LLM training and evaluation, including four pertaining datasets as well as three categories of downstream tasks encompassing nine datasets. Each downstream dataset is prompted with In-Context Learning (ICL) and Chain-of-Thought (CoT), respectively. Besides the already intriguing observation that the choice of calibration data significantly impacts the performance of pruned LLMs, our results also uncover several subtle and often unexpected findings, summarized as follows: (1) C4 is not the optimal choice for LLM pruning, even among commonly used pre-training datasets; (2) arithmetic datasets, when used as calibration data, performs on par or even better than pre-training datasets; (3) pruning with downstream datasets does not necessarily help the corresponding downstream task, compared to pre-training data; (4) ICL is widely beneficial to all data categories, whereas CoT is only useful on certain tasks. Our findings shed light on the importance of carefully selecting calibration data for LLM pruning and pave the way for more efficient deployment of these powerful models in real-world applications. We release our code at: https://github.com/abx393/llm-pruning-calibration-data.
Deep Neural Networks (DNNs) are often over-parameterized for their tasks and can be compressed quite drastically by removing weights, a process called pruning. We investigate the impact of different pruning techniques on the classification performance and interpretability of GoogLeNet. We systematically apply unstructured and structured pruning, as well as connection sparsity (pruning of input weights) methods to the network and analyze the outcomes regarding the network's performance on the validation set of ImageNet. We also compare different retraining strategies, such as iterative pruning and one-shot pruning. We find that with sufficient retraining epochs, the performance of the networks can approximate the performance of the default GoogLeNet - and even surpass it in some cases. To assess interpretability, we employ the Mechanistic Interpretability Score (MIS) developed by Zimmermann et al. . Our experiments reveal that there is no significant relationship between interpretability and pruning rate when using MIS as a measure. Additionally, we observe that networks with extremely low accuracy can still achieve high MIS scores, suggesting that the MIS may not always align with intuitive notions of interpretability, such as understanding the basis of correct decisions.
The recent paradigm shift to large-scale foundation models has brought about a new era for deep learning that, while has found great success in practice, has also been plagued by prohibitively expensive costs in terms of high memory consumption and compute. To mitigate these issues, there has been a concerted effort in post-hoc neural network pruning techniques that do not require costly retraining. Despite the considerable progress being made, existing methods often exhibit a steady drop in model performance as the compression increases. In this paper, we present a novel approach to compressing large transformers, coined OATS, that utilizes the second moment information in the input embeddings to decompose the model weights into a sum of sparse and low-rank matrices. Without any retraining, OATS achieves state-of-the-art performance when compressing models by up to $60\%$ on large language models such as Llama-3 and Phi-3 and vision transformers such as ViT and DINOv2 while delivering up to $1.37\times$ the CPU acceleration versus a model that was comparably pruned.
The colossal parameters and computational overhead of Large Language Models (LLMs) challenge their real-world applications. Network pruning, which targets unstructured or structured sparsity by removing redundant parameters, has recently been explored for LLM acceleration. Existing LLM pruning works focus on unstructured pruning, which typically requires special hardware support for a practical speed-up. In contrast, structured pruning can reduce latency on general devices. However, it remains a challenge to perform structured pruning efficiently and maintain performance, especially at high sparsity ratios. To this end, we introduce an efficient structured pruning framework named CFSP, which leverages both Coarse (interblock) and Fine-grained (intrablock) activation information as an importance criterion to guide pruning. The pruning is highly efficient, as it only requires one forward pass to compute feature activations. Specifically, we first allocate the sparsity budget across blocks based on their importance and then retain important weights within each block. In addition, we introduce a recovery fine-tuning strategy that adaptively allocates training overhead based on coarse-grained importance to further improve performance. Experimental results demonstrate that CFSP outperforms existing methods on diverse models across various sparsity budgets. Our code will be available at https://github.com/wyxscir/CFSP.
Large Language Models (LLMs) are distinguished by their massive parameter counts, which typically result in significant redundancy. This work introduces MaskLLM, a learnable pruning method that establishes Semi-structured (or ``N:M'') Sparsity in LLMs, aimed at reducing computational overhead during inference. Instead of developing a new importance criterion, MaskLLM explicitly models N:M patterns as a learnable distribution through Gumbel Softmax sampling. This approach facilitates end-to-end training on large-scale datasets and offers two notable advantages: 1) High-quality Masks - our method effectively scales to large datasets and learns accurate masks; 2) Transferability - the probabilistic modeling of mask distribution enables the transfer learning of sparsity across domains or tasks. We assessed MaskLLM using 2:4 sparsity on various LLMs, including LLaMA-2, Nemotron-4, and GPT-3, with sizes ranging from 843M to 15B parameters, and our empirical results show substantial improvements over state-of-the-art methods. For instance, leading approaches achieve a perplexity (PPL) of 10 or greater on Wikitext compared to the dense model's 5.12 PPL, but MaskLLM achieves a significantly lower 6.72 PPL solely by learning the masks with frozen weights. Furthermore, MaskLLM's learnable nature allows customized masks for lossless application of 2:4 sparsity to downstream tasks or domains. Code is available at https://github.com/NVlabs/MaskLLM.
Recent work has proposed neural network pruning techniques to reduce the size of a network while preserving robustness against adversarial examples, i.e., well-crafted inputs inducing a misclassification. These methods, which we refer to as adversarial pruning methods, involve complex and articulated designs, making it difficult to analyze the differences and establish a fair and accurate comparison. In this work, we overcome these issues by surveying current adversarial pruning methods and proposing a novel taxonomy to categorize them based on two main dimensions: the pipeline, defining when to prune; and the specifics, defining how to prune. We then highlight the limitations of current empirical analyses and propose a novel, fair evaluation benchmark to address them. We finally conduct an empirical re-evaluation of current adversarial pruning methods and discuss the results, highlighting the shared traits of top-performing adversarial pruning methods, as well as common issues. We welcome contributions in our publicly-available benchmark at https://github.com/pralab/AdversarialPruningBenchmark
A point cloud is a crucial geometric data structure utilized in numerous applications. The adoption of deep neural networks referred to as Point Cloud Neural Networks (PC- NNs), for processing 3D point clouds, has significantly advanced fields that rely on 3D geometric data to enhance the efficiency of tasks. Expanding the size of both neural network models and 3D point clouds introduces significant challenges in minimizing computational and memory requirements. This is essential for meeting the demanding requirements of real-world applications, which prioritize minimal energy consumption and low latency. Therefore, investigating redundancy in PCNNs is crucial yet challenging due to their sensitivity to parameters. Additionally, traditional pruning methods face difficulties as these networks rely heavily on weights and points. Nonetheless, our research reveals a promising phenomenon that could refine standard PCNN pruning techniques. Our findings suggest that preserving only the top p% of the highest magnitude weights is crucial for accuracy preservation. For example, pruning 99% of the weights from the PointNet model still results in accuracy close to the base level. Specifically, in the ModelNet40 dataset, where the base accuracy with the PointNet model was 87. 5%, preserving only 1% of the weights still achieves an accuracy of 86.8%. Codes are available in: https://github.com/apurba-nsu-rnd-lab/PCNN_Pruning
In the past few years, neural networks have evolved from simple Feedforward Neural Networks to more complex neural networks, such as Convolutional Neural Networks and Recurrent Neural Networks. Where CNNs are a perfect fit for tasks where the sequence is not important such as image recognition, RNNs are useful when order is important such as machine translation. An increasing number of layers in a neural network is one way to improve its performance, but it also increases its complexity making it much more time and power-consuming to train. One way to tackle this problem is to introduce sparsity in the architecture of the neural network. Pruning is one of the many methods to make a neural network architecture sparse by clipping out weights below a certain threshold while keeping the performance near to the original. Another way is to generate arbitrary structures using random graphs and embed them between an input and output layer of an Artificial Neural Network. Many researchers in past years have focused on pruning mainly CNNs, while hardly any research is done for the same in RNNs. The same also holds in creating sparse architectures for RNNs by generating and embedding arbitrary structures. Therefore, this thesis focuses on investigating the effects of the before-mentioned two techniques on the performance of RNNs. We first describe the pruning of RNNs, its impact on the performance of RNNs, and the number of training epochs required to regain accuracy after the pruning is performed. Next, we continue with the creation and training of Sparse Recurrent Neural Networks and identify the relation between the performance and the graph properties of its underlying arbitrary structure. We perform these experiments on RNN with Tanh nonlinearity (RNN-Tanh), RNN with ReLU nonlinearity (RNN-ReLU), GRU, and LSTM. Finally, we analyze and discuss the results achieved from both the experiments.
Recent advances in neural network pruning have shown how it is possible to reduce the computational costs and memory demands of deep learning models before training. We focus on this framework and propose a new pruning at initialization algorithm that leverages the Neural Tangent Kernel (NTK) theory to align the training dynamics of the sparse network with that of the dense one. Specifically, we show how the usually neglected data-dependent component in the NTK's spectrum can be taken into account by providing an analytical upper bound to the NTK's trace obtained by decomposing neural networks into individual paths. This leads to our Path eXclusion (PX), a foresight pruning method designed to preserve the parameters that mostly influence the NTK's trace. PX is able to find lottery tickets (i.e. good paths) even at high sparsity levels and largely reduces the need for additional training. When applied to pre-trained models it extracts subnetworks directly usable for several downstream tasks, resulting in performance comparable to those of the dense counterpart but with substantial cost and computational savings. Code available at: https://github.com/iurada/px-ntk-pruning
Federated learning (FL) has been introduced to enable a large number of clients, possibly mobile devices, to collaborate on generating a generalized machine learning model thanks to utilizing a larger number of local samples without sharing to offer certain privacy to collaborating clients. However, due to the participation of a large number of clients, it is often difficult to profile and verify each client, which leads to a security threat that malicious participants may hamper the accuracy of the trained model by conveying poisoned models during the training. Hence, the aggregation framework at the parameter server also needs to minimize the detrimental effects of these malicious clients. A plethora of attack and defence strategies have been analyzed in the literature. However, often the Byzantine problem is analyzed solely from the outlier detection perspective, being oblivious to the topology of neural networks (NNs). In the scope of this work, we argue that by extracting certain side information specific to the NN topology, one can design stronger attacks. Hence, inspired by the sparse neural networks, we introduce a hybrid sparse Byzantine attack that is composed of two parts: one exhibiting a sparse nature and attacking only certain NN locations with higher sensitivity, and the other being more silent but accumulating over time, where each ideally targets a different type of defence mechanism, and together they form a strong but imperceptible attack. Finally, we show through extensive simulations that the proposed hybrid Byzantine attack is effective against 8 different defence methods.
Current techniques for deep neural network (DNN) pruning often involve intricate multi-step processes that require domain-specific expertise, making their widespread adoption challenging. To address the limitation, the Only-Train-Once (OTO) and OTOv2 are proposed to eliminate the need for additional fine-tuning steps by directly training and compressing a general DNN from scratch. Nevertheless, the static design of optimizers (in OTO) can lead to convergence issues of local optima. In this paper, we proposed the Auto-Train-Once (ATO), an innovative network pruning algorithm designed to automatically reduce the computational and storage costs of DNNs. During the model training phase, our approach not only trains the target model but also leverages a controller network as an architecture generator to guide the learning of target model weights. Furthermore, we developed a novel stochastic gradient algorithm that enhances the coordination between model training and controller network training, thereby improving pruning performance. We provide a comprehensive convergence analysis as well as extensive experiments, and the results show that our approach achieves state-of-the-art performance across various model architectures (including ResNet18, ResNet34, ResNet50, ResNet56, and MobileNetv2) on standard benchmark datasets (CIFAR-10, CIFAR-100, and ImageNet).