The rapid development of large-scale deep learning models questions the affordability of hardware platforms, which necessitates the pruning to reduce their computational and memory footprints. Sparse neural networks as the product, have demonstrated numerous favorable benefits like low complexity, undamaged generalization, etc. Most of the prominent pruning strategies are invented from a model-centric perspective, focusing on searching and preserving crucial weights by analyzing network topologies. However, the role of data and its interplay with model-centric pruning has remained relatively unexplored. In this research, we introduce a novel data-model co-design perspective: to promote superior weight sparsity by learning important model topology and adequate input data in a synergetic manner. Specifically, customized Visual Prompts are mounted to upgrade neural Network sparsification in our proposed VPNs framework. As a pioneering effort, this paper conducts systematic investigations about the impact of different visual prompts on model pruning and suggests an effective joint optimization approach. Extensive experiments with 3 network architectures and 8 datasets evidence the substantial performance improvements from VPNs over existing start-of-the-art pruning algorithms. Furthermore, we find that subnetworks discovered by VPNs from pre-trained models enjoy better transferability across diverse downstream scenarios. These insights shed light on new promising possibilities of data-model co-designs for vision model sparsification.

This paper investigates the under-explored area of low-rank weight training for large-scale Conformer-based speech recognition models from scratch. Our study demonstrates the viability of this training paradigm for such models, yielding several notable findings. Firstly, we discover that applying a low-rank structure exclusively to the attention modules can unexpectedly enhance performance, even with a significant rank reduction of 12%. In contrast, feed-forward layers present greater challenges, as they begin to exhibit performance degradation with a moderate 50% rank reduction. Furthermore, we find that both initialization and layer-wise rank assignment play critical roles in successful low-rank training. Specifically, employing SVD initialization and linear layer-wise rank mapping significantly boosts the efficacy of low-rank weight training. Building on these insights, we introduce the Low-Rank Speech Model from Scratch (LR-SMS), an approach that achieves performance parity with full-rank training while delivering substantial reductions in parameters count (by at least 2×), and training time speedups (by 1.3× for ASR and 1.15× for AVSR).

Large Language Models (LLMs) have demonstrated exceptional performance across diverse tasks, yet their training remains highly resource intensive and susceptible to critical challenges such as training instability. A predominant source of this instability stems from gradient and loss spikes, which disrupt the learning process, often leading to costly interventions like checkpoint recovery and experiment restarts, further amplifying inefficiencies. This paper presents a comprehensive investigation into gradient spikes observed during LLM training, revealing their prevalence across multiple architectures and datasets. Our analysis shows that these spikes can be up to 1000× larger than typical gradients, substantially deteriorating model performance. To address this issue, we propose Spike-Aware Adam with Momentum Reset (SPAM), a novel optimizer designed to counteract gradient spikes through momentum reset and spike-aware gradient clipping. Extensive experiments, including both pre-training and fine-tuning, demonstrate that SPAM consistently surpasses Adam and its variants across a range of model scales. Additionally, SPAM facilitates memory-efficient training by enabling sparse momentum, where only a subset of momentum terms are maintained and updated. When operating under memory constraints, SPAM outperforms state-of-the-art memory-efficient optimizers such as GaLore and Adam-Mini. Our work underscores the importance of mitigating gradient spikes in LLM training and introduces an effective optimization strategy that enhances both training stability and resource efficiency at scale. Code is submitted.

Test-time interventions for language models can enhance factual accuracy, mitigate harmful outputs, and improve model efficiency without costly retraining. But despite a flood of new methods, different types of interventions are largely developing independently. In practice, multiple interventions must be applied sequentially to the same model, yet we lack standardized ways to study how interventions interact. We fill this gap by introducing composable interventions, a framework to study the effects of using multiple interventions on the same language models, featuring new metrics and a unified codebase. Using our framework, we conduct extensive experiments and compose popular methods from three emerging intervention categories---knowledge editing, model compression, and machine unlearning. Our results over 417 different compositions uncover meaningful interactions: compression hinders editing and unlearning, composing interventions hinges on their order of application, and popular general-purpose metrics are inadequate for assessing composability. Taken together, our findings showcase clear gaps in composability, suggesting a need for new multi-objective interventions.

Large Language Models (LLMs) have achieved remarkable success, yet recent findings reveal that their deeper layers often contribute minimally and can be pruned without affecting overall performance. While some view this as an opportunity for model compression, we identify it as a training shortfall rooted in the widespread use of Pre-Layer Normalization (Pre-LN). We demonstrate that Pre-LN, commonly employed in models like GPT and LLaMA, leads to diminished gradient norms in its deeper layers, reducing their effectiveness. In contrast, Post-Layer Normalization (Post-LN) preserves larger gradient norms in deeper layers but suffers from vanishing gradients in earlier layers. To address this, we introduce Mix-LN, a novel normalization technique that combines the strengths of Pre-LN and Post-LN within the same model. Mix-LN applies Post-LN to the earlier layers and Pre-LN to the deeper layers, ensuring more uniform gradient norms across layers. This allows all parts of the network—both shallow and deep layers—to contribute effectively to training. Extensive experiments with various model sizes demonstrate that Mix-LN consistently outperforms both Pre-LN and Post-LN, promoting more balanced, healthier gradient norms throughout the network, and enhancing the overall quality of LLM pre-training. Furthermore, we demonstrate that models pre-trained with Mix-LN learn better compared to those using Pre-LN or Post-LN during supervised fine-tuning, highlighting the critical importance of high-quality deep layers. By effectively addressing the inefficiencies of deep layers in current LLMs, Mix-LN unlocks their potential, enhancing model capacity without increasing model size. Our code is available at https://github.com/pixeli99/MixLN.

Large Language Models (LLMs) matrices can often be expressed in low-rank format with potential to relax memory and compute resource requirements. Unlike previous works which pivot around developing novel matrix decomposition algorithms, in this work we focus to study the emerging non-uniform low-rank properties across weight matrices in LLMs through the lens of stabilizing gradient subspace. \textit{Firstly,} we provide a theoretical framework to understand the stabilization of gradient subspaces through Hessian analysis. \textit{Secondly,} we empirically establish a consequential relationship between the gradient dynamics and low-rank expressiveness of weight matrices. Our findings reveal that different LLM components exhibit varying levels of converged low-rank structure, necessitating a non-uniform rank reduction across them to minimize performance drop due to compression. In view of that, we present \textit{Weight Low-Rank Projection} \textbf{(WeLore)} that unifies weight compression and memory-efficient fine-tuning as ONE, in a data-agnostic and one-shot way. Going beyond only as a compression technique, WeLore categorizes weight matrices into Low-rank Components (LRCs) and Non-Low-rank Components (N-LRCs) based on their ability to express themselves as low-rank. Our gradient dynamics perspective illustrate that \textit{LRCs tend to have better finetuning capabilities} and their standalone finetuning can closely mimic (sometimes outperform) the training loss trajectory and performance of full-finetuning with notable memory and compute footprint reduction. All codes and checkpoints will be released.

Recent studies have shown that supervised fine-tuning of LLMs on a small number of high-quality datasets can yield strong reasoning capabilities. However, full fine-tuning (Full FT), while powerful, is computationally expensive and susceptible to overfitting and catastrophic forgetting, particularly when data is limited. Sparse fine-tuning, which previously achieved notable success by updating only a small subset of model parameters, offers a promising trade-off between efficiency and effectiveness. Yet, it has lagged behind in the LLM era due to the difficulty of identifying parameters truly critical for reasoning. In this work, we state that weights with the largest magnitude after low-rank approximation are critical weights for fine-tuning, which we call *Principal Weights*. Surprisingly, while magnitude-based sparse fine-tuning performs poorly as a baseline on LLM fine-tuning, it becomes highly effective after rank reduction. These insights motivate our method: **L**ow-rank **I**nformed Sparse **F**ine-**T**uning ($\texttt{LIFT}$). $\texttt{LIFT}$ only updates the top 5% *Principal Weights* throughout training and consistently achieves better performance on reasoning tasks than Full FT, while maintaining memory efficiency on par with popular parameter-efficient fine-tuning methods. In addition to strong performance on target domains such as arithmetic reasoning, $\texttt{LIFT}$ also retains up to 20% more source-domain knowledge, compared to Full FT and LoRA. Our code is available at: https://github.com/zihanghliu/LIFT.

Large Language Models (LLMs) are discovered to suffer from accurately retrieving key information. To address this, we propose Mask-Enhanced Autoregressive Prediction (MEAP), a simple yet effective training paradigm that seamlessly integrates Masked Language Modeling (MLM) into Next-Token Prediction (NTP) to enhance the latter’s in-context retrieval capabilities. Specifically, MEAP first randomly masks a small fraction of input tokens and then directly performs the standard next-token prediction autoregressive using a decoder-only Transformer. MEAP eliminates the need for bidirectional attention or encoder-decoder architectures for MLM, incurring no additional computational overhead during pre-training or inference. Intensive experiments demonstrate that MEAP substantially outperforms NTP on key information retrieval and long-context reasoning tasks, while performing on par or better on commonsense reasoning tasks. The benefits of MEAP also extend to supervised fine-tuning, where it shows remarkable advantages in lost-in-the-middle scenarios, outperforming NTP by 11.77% percentage points. Our analysis indicates that MEAP’s effectiveness arises from its ability to promote more distinguishable attention scores by concentrating on a reduced set of non-masked tokens. This mechanism improves the model’s focus on task-relevant signals while mitigating the influence of peripheral context. These findings position MEAP as a promising training paradigm for large language models. Code has been submitted.

Modern Large Language Models, such as the LLaMA, Qwen and DeepSeek series, predominantly adopt the Pre-LayerNorm (Pre-LN) Transformer architecture. While being stable during pretraining and scalable to large model sizes, Pre-LN suffers from an exponential growth in activation variance across layers, causing the shortcut to dominate over sub-layer outputs in the residual connection and limiting the learning capacity of deeper layers. To mitigate this issue, we propose Gradient-Preserving Activation Scaling (GPAS), a simple technique that can be used in combination with existing approaches. GPAS works by scaling down the intermediate activations while keeping their gradients unchanged. This leaves information in the activations intact, and avoids the gradient vanishing problem associated with gradient downscaling. Extensive experiments across various model sizes from 71M to 1B show that GPAS achieves consistent performance gains. Beyond enhancing Pre-LN Transformers, GPAS also shows promise in improving alternative architectures such as Sandwich-LN and DeepNorm, demonstrating its versatility and potential for improving training dynamics in a wide range of settings. Our code is available at https://github.com/dandingsky/GPAS.

Weight decay is a standard regularization technique for training large language models (LLMs). While it is common to assign a uniform decay rate to every layer, this approach overlooks the structural diversity of LLMs and the varying spectral properties across modules. In this paper, we introduce AlphaDecay, a simple yet effective method that adaptively assigns different weight decay strengths to each module of an LLM. Our approach is guided by Heavy-Tailed Self-Regularization (HT-SR) theory, which analyzes the empirical spectral density (ESD) of weight correlation matrices to quantify “heavy-tailedness.” Modules exhibiting more pronounced heavy-tailed ESDs, reflecting stronger feature learning, are assigned weaker decay, while modules with lighter-tailed spectra receive stronger decay. Our method leverages tailored weight decay assignments to balance the module-wise differences in spectral properties, leading to improved performance. Extensive pre-training tasks with various model sizes from 60M to 1B demonstrate that AlphaDecay achieves better perplexity and generalization than conventional uniform decay and other adaptive decay baselines. The code is available at https://github.com/hed-ucas/AlphaDecay.

In this paper, we re-introduce the Curse of Depth, a concept that re-introduces, explains, and addresses the recent observation in modern Large Language Models (LLMs) where deeper layers are much less effective than expected. We first confirm the wide existence of this phenomenon across the most popular families of LLMs, such as Llama, Mistral, DeepSeek, and Qwen. Our analysis, theoretically and empirically, identifies that the underlying reason for the ineffectiveness of deep layers in LLMs is the widespread usage of Pre-Layer Normalization (Pre-LN). While Pre-LN stabilizes the training of Transformer LLMs, its output variance exponentially grows with the model depth, which undesirably causes the derivative of the deep Transformer blocks to be an identity matrix, and therefore barely contributes to the training. To resolve this training pitfall, we propose LayerNorm Scaling, which scales the variance of output of the layer normalization inversely by the square root of its depth. This simple modification mitigates the output variance explosion of deeper Transformer layers, improving their contribution. Our experimental results, spanning model sizes from 130M to 7B, demonstrate that \ours significantly enhances LLM pre-training performance compared to Pre-LN. Moreover, this improvement seamlessly carries over to supervised fine-tuning. All these gains can be attributed to the fact that LayerNorm Scaling enables deeper layers to contribute more effectively during training.

Graph Neural Networks (GNNs) have emerged as powerful tools for analyzing graph-structured data, but their widespread adoption in critical applications is hindered by inherent biases related to sensitive attributes such as gender and race. While existing debiasing approaches typically focus on either modifying input graphs or incorporating fairness constraints into model objectives, we propose Fair Sparse GNN (FS-GNN), a novel framework that simultaneously enhances fairness and efficiency through joint sparsification of both input graphs and model architectures. Our approach iteratively identifies and removes less informative edges from input graphs while pruning redundant weights from the GNN model, guided by carefully designed fairness-aware objective functions. Through extensive experiments on real-world datasets, we demonstrate that FS-GNN achieves superior fairness metrics (reducing Statistical Parity from 7.94 to 0.6) while maintaining competitive prediction accuracy compared to state-of-the-art methods. Additionally, our theoretical analysis reveals distinct fairness implications of graph versus architecture sparsification, providing insights for future fairness-aware GNN designs. The proposed method not only advances fairness in GNNs but also offers substantial computational benefits through reduced model complexity, with FLOPs reductions ranging from 24% to 67%.

The goal of continual learning (CL) is to learn from a series of continuously arriving new tasks without forgetting previously learned old tasks. To avoid catastrophic forgetting of old tasks, orthogonal gradient projection (OGP) based CL methods constrain the gradients of new tasks to be orthogonal to the space spanned by old tasks. This strict gradient constraint will limit the learning ability of new tasks, resulting in lower performance on new tasks. In this paper, we first establish a unified framework for OGP-based CL methods. We then revisit OGP-based CL methods from a new perspective on the loss landscape, where we find that when relaxing projection constraints to improve performance on new tasks, the unflatness of the loss landscape can lead to catastrophic forgetting of old tasks. Based on our findings, we propose a new Dual Flatness-aware OGD framework that optimizes the flatness of the loss landscape from both data and weight levels. Our framework consists of three modules: data and weight perturbation, flatness-aware optimization, and gradient projection. Specifically, we first perform perturbations on the task's data and current model weights to make the task's loss reach the worst-case. Next, we optimize the loss and loss landscape on the original data and the worst-case perturbed data to obtain a flatness-aware gradient. Finally, the flatness-aware gradient will update the network in directions orthogonal to the space spanned by the old tasks. Extensive experiments on four benchmark datasets show that the framework improves the flatness of the loss landscape and performance on new tasks, and achieves state-of-the-art (SOTA) performance on average accuracy across all tasks.

We are the first to explore the intricacies of LLM compression techniques in the context of text-to-music generation, focusing on the MusicGen Transformer model. We implement and evaluate multiple compression methods, including pruning techniques and low-rank approximations to the MusicGen architecture. Additionally, we analyze the model’s sensitivity to calibration data and assess the impact of these pruning techniques on different music styles. Our results show that the SparseGPT pruning method is more robust than the Wanda method when using limited calibration music samples. Finally, we delve into transformer interpretability, investigating how specific layers within the MusicGen Transformer influence music style generation. We conclude that existing pruning methods can effectively reduce the MusicGen Transformer’s size, making it feasible for deployment in complex music generation models that require handling long sequences of notes.

Abhinav Bandari, Lu Yin, Cheng-Yu Hsieh, Ajay Kumar Jaiswal, Tianlong Chen, Li Shen, Ranjay Krishna, Shiwei Liu. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Autoregressive Large Language Models (e.g., LLaMa, GPTs) are omnipresent achieving remarkable success in language understanding and generation.However, such impressive capability typically comes with a substantial model size, which presents significant challenges for autoregressive token-by-token generation.To mitigate computation overload incurred during generation, several early-exit and layer-dropping strategies have been proposed.Despite some promising success due to the redundancy across LLMs layers on metrics like Rough-L/BLUE, our careful knowledgeintensive evaluation unveils issues such as generation collapse, hallucination, and noticeable performance drop even at the trivial exit ratio of ∼ 10-15% of layers.We attribute these errors primarily to ineffective handling of the KV cache through state copying during early exit.In this work, we observe the saturation of computationally expensive feed-forward blocks of LLM layers and propose FFN-SkipLLM, which is a novel fine-grained skip strategy for autoregressive LLMs.FFN-SkipLLM leverages an input-adaptive feed-forward skipping approach that can skip ∼ 25-30% of FFN blocks of LLMs with marginal change in performance on knowledge-intensive generation tasks without any requirement to handle the KV cache.Our extensive experiments and ablation studies across benchmarks like MT-Bench, Factoid-QA, and variable-length text summarization illustrate how our simple and easy-touse method can facilitate faster autoregressive decoding.

The ever-increasing large language models (LLMs), though opening a potential path for the upcoming artificial general intelligence, sadly drops a daunting obstacle on the way towards their on-device deployment. As one of the most well-established pre-LLMs approaches in reducing model complexity, network pruning appears to lag behind in the era of LLMs, due mostly to its costly fine-tuning (or re-training) necessity under the massive volumes of model parameter and training data. To close this industry-academia gap, we introduce Dynamic Sparse No Training ($\texttt{DSNT}$), a training-free fine-tuning approach that slightly updates sparse LLMs without the expensive backpropagation and any weight updates. Inspired by the Dynamic Sparse Training, $\texttt{DSNT}$ minimizes the reconstruction error between the dense and sparse LLMs, in the fashion of performing iterative weight pruning-and-growing on top of sparse LLMs. To accomplish this purpose, $\texttt{DSNT}$ particularly takes into account the anticipated reduction in reconstruction error for pruning and growing, as well as the variance w.r.t. different input data for growing each weight. This practice can be executed efficiently in linear time since its obviates the need of backpropagation for fine-tuning LLMs. Extensive experiments on LLaMA-V1/V2, Vicuna, and OPT across various benchmarks demonstrate the effectiveness of $\texttt{DSNT}$ in enhancing the performance of sparse LLMs, especially at high sparsity levels. For instance, $\texttt{DSNT}$ is able to outperform the state-of-the-art Wanda by 26.79 perplexity at 70% sparsity with LLaMA-7B. Our paper offers fresh insights into how to fine-tune sparse LLMs in an efficient training-free manner and open new venues to scale the great potential of sparsity to LLMs. Codes are available at https://github.com/zyxxmu/DSnoT.

Dynamic Sparse Training (DST) employs a greedy search mechanism to identify an optimal sparse subnetwork by periodically pruning and growing network connections during training. To guarantee effectiveness, DST algorithms rely on high search frequency, which consequently, requires large learning rate and batch size to enforce stable neuron learning. Such settings demand extreme memory consumption, as well as generating significant system overheads that limit the wide deployment of deep learning-based applications on resource-constraint platforms. To reconcile such, we propose $\underline{Neur}$on $\underline{Rev}$italization framework for DST (NeurRev), based on an innovative finding that dormant neurons exist with the presence of weight sparsity, and cannot be revitalized (i.e., activated for learning) even with high sparse mask search frequency. These dormant neurons produce a large quantity of zeros during training, which contribute relatively little to the outputs of succeeding layers or to the final results. Different from most existing DST algorithms that spare no effort designing weight growing criteria, NeurRev focuses on optimizing the long-neglected pruning part, which awakes dormant neurons by pruning and incurs no additional computation costs. As such, NeurRev advances more effective neuron learning, which not only achieves outperforming accuracy in a variety of networks and datasets, but also promoting a low-cost dynamism at system-level. Systematical evaluations on training speed and system overhead are conducted on the mobile devices, where the proposed NeurRev framework consistently outperforms representative state-of-the-arts. Code will be released.

Multi-task learning (MTL) aims to empower a model to tackle multiple tasks simultaneously. A recent development known as task arithmetic has revealed that several models, each fine-tuned for distinct tasks, can be directly merged into a single model to execute MTL without necessitating a retraining process using the initial training data. Nevertheless, this direct addition of models often leads to a significant deterioration in the overall performance of the merged model. This decline occurs due to potential conflicts and intricate correlations among the multiple tasks. Consequently, the challenge emerges of how to merge pre-trained models more effectively without using their original training data. This paper introduces an innovative technique called Adaptive Model Merging (AdaMerging). This approach aims to autonomously learn the coefficients for model merging, either in a task-wise or layer-wise manner, without relying on the original training data. Specifically, our AdaMerging method operates as an automatic, unsupervised task arithmetic scheme. It leverages entropy minimization on unlabeled test samples from the multi-task setup as a surrogate objective function to iteratively refine the merging coefficients of the multiple models. Our experimental findings across eight tasks demonstrate the efficacy of the AdaMerging scheme we put forth. Compared to the current state-of-the-art (SOTA) task arithmetic merging scheme, AdaMerging showcases a remarkable 11\% improvement in performance. Notably, AdaMerging also exhibits superior generalization capabilities when applied to unseen downstream tasks. Furthermore, it displays a significantly enhanced robustness to data distribution shifts that may occur during the testing phase.

Despite the exceptional performance of Large Language Models (LLMs), the substantial volume of key-value (KV) pairs cached during inference presents a barrier to their efficient deployment. To ameliorate this, recent works have aimed to selectively eliminate these caches, informed by the attention scores of associated tokens. However, such cache eviction invariably leads to output perturbation, regardless of the token choice. This perturbation escalates with the compression ratio, which can precipitate a marked deterioration in LLM inference performance. This paper introduces Cache Merging (CaM) as a solution to mitigate this challenge. CaM adaptively merges to-be-evicted caches into the remaining ones, employing a novel sampling strategy governed by the prominence of attention scores within discarded locations. In this manner, CaM enables memory-efficient LLMs to preserve critical token information, even obviating the need to maintain their corresponding caches. Extensive experiments utilizing LLaMA, OPT, and GPT-NeoX across various benchmarks corroborate CaM’s proficiency in bolstering the performance of memory-efficient LLMs. Code is released at https://github.com/zyxxmu/cam.

Large Language Models (LLMs), renowned for their remarkable performance across diverse domains, present a challenge due to their colossal model size when it comes to practical deployment. In response to this challenge, efforts have been directed toward the application of traditional network pruning techniques to LLMs, uncovering a massive number of parameters can be pruned in one-shot without hurting performance. Building upon insights gained from pre-LLM models, particularly BERT-level language models, prevailing LLM pruning strategies have consistently adhered to the practice of uniformly pruning all layers at equivalent sparsity levels, resulting in robust performance. However, this observation stands in contrast to the prevailing trends observed in the field of vision models, where non-uniform layerwise sparsity typically yields substantially improved results. To elucidate the underlying reasons for this disparity, we conduct a comprehensive analysis of the distribution of token features within LLMs. In doing so, we discover a strong correlation with the emergence of outliers, defined as features exhibiting significantly greater magnitudes compared to their counterparts in feature dimensions. Inspired by this finding, we introduce a novel LLM pruning methodology that incorporates a tailored set of **non-uniform layerwise sparsity ratios** specifically designed for LLM pruning, termed as **O**utlier **W**eighed **L**ayerwise sparsity (**OWL**). The sparsity ratio of OWL is directly proportional to the outlier ratio observed within each layer, facilitating a more effective alignment between layerwise weight sparsity and outlier ratios. Our empirical evaluation, conducted across the LLaMA-V1/V2, Vicuna, OPT, and Mistral, spanning various benchmarks, demonstrates the distinct advantages offered by OWL over previous methods. For instance, OWL exhibits a remarkable performance gain, surpassing the state-of-the-art Wanda and SparseGPT by **61.22** and **6.80** perplexity at a high sparsity level of 70%, respectively, while delivering **2.6$\times$** end-to-end inference speed-up in the DeepSparse inference engine. Code is available at https://github.com/luuyin/OWL.git.

Dynamic Sparse Training (DST) is an effective approach for addressing the substantial training resource requirements posed by the ever-increasing size of the Deep Neural Networks (DNNs). Characterized by its dynamic "train-prune-grow” schedule during training, DST implicitly develops a bi-level structure for training the weights while discovering a subnetwork topology. However, such a structure is consistently overlooked by the current DST algorithms for further optimization opportunities, and these algorithms, on the other hand, solely optimize the weights while determining masks heuristically. In this paper, we extensively study DST algorithms and argue that the training scheme of DST naturally forms a bi-level problem in which the updating of weight and mask is interdependent. Based on this observation, we introduce a novel efficient training framework called BiDST, which for the first time, introduces bi-level optimization methodology into dynamic sparse training domain. Unlike traditional partial-heuristic DST schemes, which suffer from sub-optimal search efficiency for masks and miss the opportunity to fully explore the topological space of neural networks, BiDST excels at discovering excellent sparse patterns by optimizing mask and weight simultaneously, resulting in maximum 2.62% higher accuracy, 2.1$\times$ faster execution speed, and 25$\times$ reduced overhead. Code available at https://github.com/jjsrf/BiDST-ICML2024.

Sparse Neural Networks (SNNs) have received voluminous attention for mitigating the explosion in computational costs and memory footprints of modern deep neural networks. Despite their popularity, most state-of-the-art training approaches seek to find a single high-quality sparse subnetwork with a preset sparsity pattern and ratio, making them inadequate to satiate platform and resource variability. Recently proposed approaches attempt to jointly train multiple subnetworks (we term as “sparse co-training") with a fixed sparsity pattern, to allow switching sparsity ratios subject to resource requirements. In this work, we take one more step forward and expand the scope of sparse co-training to cover diverse sparsity patterns and multiple sparsity ratios at once. We introduce Sparse Cocktail, the first sparse co-training framework that co-trains a suite of sparsity patterns simultaneously, loaded with multiple sparsity ratios which facilitate harmonious switch across various sparsity patterns and ratios at inference depending on the hardware availability. More specifically, Sparse Cocktail alternatively trains subnetworks generated from different sparsity patterns with a gradual increase in sparsity ratios across patterns and relies on an unified mask generation process and the Dense Pivot Co-training to ensure the subnetworks of different patterns orchestrate their shared parameters without canceling each other’s performance. Experiment results on image classification, object detection, and instance segmentation illustrate the favorable effectiveness and flexibility of Sparse Cocktail, pointing to a promising direction for sparse co-training. Codes will be released.

The recent success of Automatic Speech Recognition (ASR) is largely attributed to the ever-growing amount of training data. However, this trend has made model training prohibitively costly and imposed computational demands. While data pruning has been proposed to mitigate this issue by identifying a small subset of relevant data, its application in ASR has been barely explored, and existing works often entail significant overhead to achieve meaningful results. To fill this gap, this paper presents the first investigation of dynamic data pruning for ASR, finding that we can reach the full-data performance by dynamically selecting 70% of data. Furthermore, we introduce Dynamic Data Pruning for ASR (DDP-ASR), which offers several fine-grained pruning granularities specifically tailored for speech-related datasets, going beyond the conventional pruning of entire time sequences. Our intensive experiments show that DDP-ASR can save up to 1.6x training time with negligible performance loss.

This paper aims to overcome the ``lost-in-the-middle'' challenge of large language models (LLMs). While recent advancements have successfully enabled LLMs to perform stable language modeling with up to 4 million tokens, the persistent difficulty faced by most LLMs in identifying relevant information situated in the middle of the context has not been adequately tackled. To address this problem, this paper introduces Multi-scale Positional Encoding (Ms-PoE) which is a simple yet effective plug-and-play approach to enhance the capacity of LLMs to handle the relevant information located in the middle of the context, without fine-tuning or introducing any additional overhead. Ms-PoE leverages the position indice rescaling to relieve the long-term decay effect introduced by RoPE, while meticulously assigning distinct scaling ratios to different attention heads to preserve essential knowledge learned during the pre-training step, forming a multi-scale context fusion from short to long distance. Extensive experiments with a wide range of LLMs demonstrate the efficacy of our approach. Notably, Ms-PoE achieves an average accuracy gain of up to 3.8 on the Zero-SCROLLS benchmark over the original LLMs. Code will be made public upon acceptence.

Recent work on pruning large language models (LLMs) has shown that one can eliminate a large number of parameters without compromising performance, making pruning a promising strategy to reduce LLM model size. Existing LLM pruning strategies typically assign uniform pruning ratios across layers, limiting overall pruning ability; and recent work on layerwise pruning of LLMs is often based on heuristics that can easily lead to suboptimal performance. In this paper, we leverage Heavy-Tailed Self-Regularization (HT-SR) Theory, in particular the shape of empirical spectral densities (ESDs) of weight matrices, to design improved layerwise pruning ratios for LLMs. Our analysis reveals a wide variability in how well-trained, and thus relatedly how prunable, different layers of an LLM are. Based on this, we propose AlphaPruning, which uses shape metrics to allocate layerwise sparsity ratios in a more theoretically-principled manner. AlphaPruning can be used in conjunction with multiple existing LLM pruning methods. Our empirical results show that AlphaPruning prunes LLaMA-7B to 80% sparsity while maintaining reasonable perplexity, marking a first in the literature on LLMs.

Deep neural networks have evolved as the leading approach in 3D medical image segmentation due to their outstanding performance. However, the ever-increasing model size and computational cost of deep neural networks have become the primary barriers to deploying them on real-world, resource-limited hardware. To achieve both segmentation accuracy and efficiency, we propose a 3D medical image segmentation model called Efficient to Efficient Network (E2ENet), which incorporates two parametrically and computationally efficient designs. i. Dynamic sparse feature fusion (DSFF) mechanism: it adaptively learns to fuse informative multi-scale features while reducing redundancy. ii. Restricted depth-shift in 3D convolution: it leverages the 3D spatial information while keeping the model and computational complexity as 2D-based methods. We conduct extensive experiments on AMOS, Brain Tumor Segmentation and BTCV Challenge, demonstrating that E2ENet consistently achieves a superior trade-off between accuracy and efficiency than prior arts across various resource constraints. %In particular, with a single model and single scale, E2ENet achieves comparable accuracy on the large-scale challenge AMOS-CT, while saving over 69% parameter count and 27% FLOPs in the inference phase, compared with the previous best-performing method. Our code has been made available at: https://github.com/boqian333/E2ENet-Medical.

Lottery tickets (LTs) is able to discover accurate and sparse subnetworks that could be trained in isolation to match the performance of dense networks. Ensemble, in parallel, is one of the oldest time-proven tricks in machine learning to improve performance by combining the output of multiple independent models. However, the benefits of ensemble in the context of LTs will be diluted since ensemble does not directly lead to stronger sparse subnetworks, but leverages their predictions for a better decision. In this work, we first observe that directly averaging the weights of the adjacent learned subnetworks significantly boosts the performance of LTs. Encouraged by this observation, we further propose an alternative way to perform an "ensemble'' over the subnetworks identified by iterative magnitude pruning via a simple interpolating strategy. We call our method Lottery Pools. In contrast to the naive ensemble which brings no performance gains to each single subnetwork, Lottery Pools yields much stronger sparse subnetworks than the original LTs without requiring any extra training or inference cost. Across various modern architectures on CIFAR-10/100 and ImageNet, we show that our method achieves significant performance gains in both, in-distribution and out-of-distribution scenarios. Impressively, evaluated with VGG-16 and ResNet-18, the produced sparse subnetworks outperform the original LTs by up to 1.88% on CIFAR-100 and 2.36% on CIFAR-100-C; the resulting dense network surpasses the pre-trained dense-model up to 2.22% on CIFAR-100 and 2.38% on CIFAR-100-C. Our source code can be found at https://github.com/luuyin/Lottery-pools.

The goal of continual learning (CL) is to continuously learn new tasks without forgetting previously learned old tasks. To alleviate catastrophic forgetting, gradient projection based CL methods require that the gradient updates of new tasks are orthogonal to the subspace spanned by old tasks. This limits the learning process and leads to poor performance on the new task due to the projection constraint being too strong. In this paper, we first revisit the gradient projection method from the perspective of flatness of loss surface, and find that unflatness of the loss surface leads to catastrophic forgetting of the old tasks when the projection constraint is reduced to improve the performance of new tasks. Based on our findings, we propose a Data Augmented Flatness-aware Gradient Projection (DFGP) method to solve the problem, which consists of three modules: data and weight perturbation, flatness-aware optimization, and gradient projection. Specifically, we first perform a flatness-aware perturbation on the task data and current weights to find the case that makes the task loss worst. Next, flatness-aware optimization optimizes both the loss and the flatness of the loss surface on raw and worst-case perturbed data to obtain a flatness-aware gradient. Finally, gradient projection updates the network with the flatness-aware gradient along directions orthogonal to the subspace of the old tasks. Extensive experiments on four datasets show that our method improves the flatness of loss surface and the performance of new tasks, and achieves state-of-the-art (SOTA) performance in the average accuracy of all tasks.

This paper reveals a new appeal of the recently emerged large-kernel Convolutional Neural Networks (ConvNets): as the teacher in Knowledge Distillation (KD) for small-kernel ConvNets. While Transformers have led state-of-the-art (SOTA) performance in various fields with ever-larger models and labeled data, small-kernel ConvNets are considered more suitable for resource-limited applications due to the efficient convolution operation and compact weight sharing. KD is widely used to boost the performance of small-kernel ConvNets. However, previous research shows that it is not quite effective to distill knowledge (e.g., global information) from Transformers to small-kernel ConvNets, presumably due to their disparate architectures. We hereby carry out a first-of-its-kind study unveiling that modern large-kernel ConvNets, a compelling competitor to Vision Transformers, are remarkably more effective teachers for small-kernel ConvNets, due to more similar architectures. Our findings are backed up by extensive experiments on both logit-level and feature-level KD "out of the box", with no dedicated architectural nor training recipe modifications. Notably, we obtain the **best-ever pure ConvNet** under 30M parameters with 83.1% top-1 accuracy on ImageNet, outperforming current SOTA methods including ConvNeXt V2 and Swin V2. We also find that beneficial characteristics of large-kernel ConvNets, e.g., larger effective receptive fields, can be seamlessly transferred to students through this large-to-small kernel distillation. Code is available at: https://github.com/VITA-Group/SLaK.

Graphs are omnipresent and GNNs are a powerful family of neural networks for learning over graphs. Despite their popularity, scaling GNNs either by deepening or widening suffers from prevalent issues of $\textit{unhealthy gradients, over-smoothening, information squashing}$, which often lead to sub-standard performance. In this work, we are interested in exploring a principled way to scale GNNs capacity without deepening or widening, which can improve its performance across multiple small and large graphs. Motivated by the recent intriguing phenomenon of model soups, which suggest that fine-tuned weights of multiple large-language pre-trained models can be merged to a better minima, we argue to exploit the fundamentals of model soups to mitigate the aforementioned issues of memory bottleneck and trainability during GNNs scaling. More specifically, we propose not to deepen or widen current GNNs, but instead present $\textbf{first data-centric perspective}$ of model soups to build powerful GNNs by dividing giant graph data to build independently and parallelly trained multiple comparatively weaker GNNs without any intermediate communication, and $\textit{combining their strength}$ using a greedy interpolation soup procedure to achieve state-of-the-art performance. Moreover, we provide a wide variety of model soup preparation techniques by leveraging state-of-the-art graph sampling and graph partitioning approaches that can handle large graph data structures. Our extensive experiments across many real-world small and large graphs, illustrate the effectiveness of our approach and point towards a promising orthogonal direction for GNN scaling. Codes are available at: https://github.com/VITA-Group/graph_ladling

Large pre-trained transformers have been receiving explosive attention in the past few years, due to their acculturation for numerous downstream applications via fine-tuning, but their exponentially increasing parameter counts are becoming a primary hurdle to even just fine-tune them without industry-standard hardware. Recently, Lottery Ticket Hypothesis (LTH) and its variants, have been exploited to prune these large pre-trained models generating subnetworks which can achieve similar performance as their dense counterparts, but LTH pragmatism is enormously inhibited by repetitive full training and pruning routine of iterative magnitude pruning (IMP) which worsens with increasing model size. Motivated by the recent observations of model soups, which suggest that fine-tuned weights of multiple models can be merged to a better minima, we propose **Instant Soup Pruning (ISP)** to generate lottery ticket quality subnetworks, using a fraction of the original IMP cost by replacing the expensive intermediate pruning stages of IMP with computationally efficient weak mask generation and aggregation routine. More specifically, during the mask generation stage, ISP takes a small handful of iterations using varying training protocols and data subsets to generate many weak and noisy subnetworks, and superpose them to average out the noise creating a high-quality denoised subnetwork. Our extensive experiments and ablation on two popular large-scale pre-trained models: $\texttt{CLIP} (unexplored in pruning till date)$ and $\texttt{BERT}$ across multiple benchmark vision $\texttt{\{MNIST, SVHN, Cars, GTSRB, CIFAR-10, CIFAR-100\}}$ and language datasets $\texttt{\{MNLI, QNLI, QQP, SST, ...\}}$ validate the effectiveness of ISP compared to several state-of-the-art pruning methods. Additionally, we show that ISP can be easily modified with minimal overhead to produce benefits comparable to model soups, without the prerequisite to generate multiple candidates fine-tuned models. Codes are available at: https://github.com/VITA-Group/instant_soup.

Sparse training has received an upsurging interest in machine learning due to its tantalizing saving potential for both the entire training process as well as the inference. Dynamic sparse training (DST) as a leading approach can train deep neural networks at high sparsity from scratch to match the performance of their dense counterparts. However, most if not all DST prior arts demonstrate their effectiveness on unstructured sparsity with highly irregular sparse patterns, which receives limited support in common hardware. This limitation hinders the usage of DST in practice. In this paper, we propose Channel-aware dynamic sparse (Chase), that for the first time seamlessly translates the promise of unstructured dynamic sparsity to GPU-friendly channel-level sparsity (not fine-grained N:M or group sparsity) during one end-to-end training process, without any ad-hoc operations. The resulting small sparse networks can be directly accelerated by commodity hardware, without using any particularly sparsity-aware hardware accelerators. This appealing outcome is partially motivated by a hidden phenomenon of dynamic sparsity: off-the-shelf unstructured DST implicitly involves biased parameter reallocation across channels, with a large fraction of channels (up to 60%) being sparser than others. By progressively identifying and removing these channels during training, our approach transfers unstructured sparsity to channel-wise sparsity. Our experimental results demonstrate that Chase achieves 1.7x inference throughput speedup on common GPU devices without compromising accuracy with ResNet-50 on ImageNet. We release our code in https://github.com/luuyin/chase.

Recent years have witnessed significant progress in understanding the relationship between the connectivity of a deep network's architecture as a graph, and the network's performance. A few prior arts connected deep architectures to expander graphs or Ramanujan graphs, and particularly,[7] demonstrated the use of such graph connectivity measures with ranking and relative performance of various obtained sparse sub-networks (i.e. models with prune masks) without the need for training. However, no prior work explicitly explores the role of parameters in the graph's connectivity, making the graph-based understanding of prune masks and the magnitude/gradient-based pruning practice isolated from one another. This paper strives to fill in this gap, by analyzing the Weighted Spectral Gap of Ramanujan structures in sparse neural networks and investigates its correlation with final performance. We specifically examine the evolution of sparse structures under a popular dynamic sparse-to-sparse network training scheme, and intriguingly find that the generated random topologies inherently maximize Ramanujan graphs. We also identify a strong correlation between masks, performance, and the weighted spectral gap. Leveraging this observation, we propose to construct a new "full-spectrum coordinate'' aiming to comprehensively characterize a sparse neural network's promise. Concretely, it consists of the classical Ramanujan's gap (structure), our proposed weighted spectral gap (parameters), and the constituent nested regular graphs within. In this new coordinate system, a sparse subnetwork's L2-distance from its original initialization is found to have nearly linear correlated with its performance. Eventually, we apply this unified perspective to develop a new actionable pruning method, by sampling sparse masks to maximize the L2-coordinate distance. Our method can be augmented with the "pruning at initialization" (PaI) method, and significantly outperforms existing PaI methods. With only a few iterations of training (e.g 500 iterations), we can get LTH-comparable performance as that yielded via "pruning after training", significantly saving pre-training costs. Codes can be found at: https://github.com/VITA-Group/FullSpectrum-PAI.

Large pre-trained transformers are $\textit{show-stealer}$ in modern-day deep learning, and it becomes crucial to comprehend the parsimonious patterns that exist within them as they grow in scale. With exploding parameter counts, Lottery Ticket Hypothesis (LTH) and its variants, have lost their pragmatism in sparsifying them due to high computation and memory bottleneck of repetitive $\textit{train-prune-retrain}$ routine of iterative magnitude pruning (IMP) which worsens with increasing model size. In this paper, we comprehensively study $\textit{induced sparse patterns}$ across multiple large pre-trained vision and language transformers. We propose the existence of -- $\textbf{essential sparsity}$ defined with a $\textbf{sharp dropping point}$ beyond which the performance declines much faster w.r.t the rise of sparsity level, when we directly remove weights with the smallest magnitudes in $\textbf{one-shot}$. We also present an intriguing emerging phenomenon of $\textbf{abrupt sparsification}$ during the pre-training of BERT, i.e., BERT suddenly becomes heavily sparse in pre-training after certain iterations. Moreover, our observations also indicate a $\textbf{counter-intuitive}$ finding that BERT trained with a larger amount of pre-training data tends to have a better ability to condense knowledge in comparatively relatively fewer parameters. Lastly, we investigate the effect of the pre-training loss on essential sparsity and discover that self-supervised learning (SSL) objectives trigger stronger emergent sparsification properties than supervised learning (SL). All our codes will be publicly available.

Pruning at initialization (PaI) aims to remove weights of neural networks before training in pursuit of training efficiency besides the inference. While off-the-shelf PaI methods manage to find trainable subnetworks that outperform random pruning, their performance in terms of both accuracy and computational reduction is far from satisfactory compared to post-training pruning and the understanding of PaI is missing. For instance, recent studies show that existing PaI methods only able to find good layerwise sparsities not weights, as the discovered subnetworks are surprisingly resilient against layerwise random mask shuffling and weight re-initialization. In this paper, we study PaI from a brand-new perspective -- the topology of subnetworks. In particular, we propose a principled framework for analyzing the performance of Pruning and Initialization (PaI) methods with two quantities, namely, the number of effective paths and effective nodes. These quantities allow for a more comprehensive understanding of PaI methods, giving us an accurate assessment of different subnetworks at initialization. We systematically analyze the behavior of various PaI methods through our framework and observe a guiding principle for constructing effective subnetworks: *at a specific sparsity, the top-performing subnetwork always presents a good balance between the number of effective nodes and the number of effective paths.* Inspired by this observation, we present a novel data-agnostic pruning method by solving a multi-objective optimization problem. By conducting extensive experiments across different architectures and datasets, our results demonstrate that our approach outperforms state-of-the-art PaI methods while it is able to discover subnetworks that have much lower inference FLOPs (up to 3.4$\times$). Code will be fully released.

The receptive field (RF), which determines the region of time series to be “seen” and used, is critical to improve the performance for time series classification (TSC). However, the variation of signal scales across and within time series data, makes it challenging to decide on proper RF sizes for TSC. In this paper, we propose a dynamic sparse network (DSN) with sparse connections for TSC, which can learn to cover various RF without cumbersome hyper-parameters tuning. The kernels in each sparse layer are sparse and can be explored under the constraint regions by dynamic sparse training, which makes it possible to reduce the resource cost. The experimental results show that the proposed DSN model can achieve state-of-art performance on both univariate and multivariate TSC datasets with less than 50% computational cost compared with recent baseline methods, opening the path towards more accurate resource-aware methods for time series analyses. Our code is publicly available at: https://github.com/QiaoXiao7282/DSN.

Abstract Sparse neural networks attract increasing interest as they exhibit comparable performance to their dense counterparts while being computationally efficient. Pruning the dense neural networks is among the most widely used methods to obtain a sparse neural network. Driven by the high training cost of such methods that can be unaffordable for a low-resource device, training sparse neural networks sparsely from scratch has recently gained attention. However, existing sparse training algorithms suffer from various issues, including poor performance in high sparsity scenarios, computing dense gradient information during training, or pure random topology search. In this paper, inspired by the evolution of the biological brain and the Hebbian learning theory, we present a new sparse training approach that evolves sparse neural networks according to the behavior of neurons in the network. Concretely, by exploiting the cosine similarity metric to measure the importance of the connections, our proposed method, “Cosine similarity-based and random topology exploration (CTRE)”, evolves the topology of sparse neural networks by adding the most important connections to the network without calculating dense gradient in the backward. We carried out different experiments on eight datasets, including tabular, image, and text datasets, and demonstrate that our proposed method outperforms several state-of-the-art sparse training algorithms in extremely sparse neural networks by a large gap. The implementation code is available on Github.

Assigning meaning to parts of image data is the goal of semantic image segmentation. Machine learning methods, specifically supervised learning is commonly used in a variety of tasks formulated as semantic segmentation. One of the major challenges in the supervised learning approaches is expressing and collecting the rich knowledge that experts have with respect to the meaning present in the image data. Towards this, typically a fixed set of labels is specified and experts are tasked with annotating the pixels, patches or segments in the images with the given labels. In general, however, the set of classes does not fully capture the rich semantic information present in the images. For example, in medical imaging such as histology images, the different parts of cells could be grouped and sub-grouped based on the expertise of the pathologist. To achieve such a precise semantic representation of the concepts in the image, we need access to the full depth of knowledge of the annotator. In this work, we develop a novel approach to collect segmentation annotations from experts based on psychometric testing. Our method consists of psychometric testing procedure, active query selection, query enhancement, and a deep metric learning model to achieve a patch-level image embedding that allows for semantic segmentation of images. We show the merits of our method with evaluation on the synthetically generated image, aerial image and histology image.

Sparse neural networks have been widely applied to reduce the computational demands of training and deploying over-parameterized deep neural networks. For inference acceleration, methods that discover a sparse network from a pre-trained dense network (dense-to-sparse training) work effectively. Recently, dynamic sparse training (DST) has been proposed to train sparse neural networks without pre-training a dense model (sparse-to-sparse training), so that the training process can also be accelerated. However, previous sparse-to-sparse methods mainly focus on Multilayer Perceptron Networks (MLPs) and Convolutional Neural Networks (CNNs), failing to match the performance of dense-to-sparse methods in the Recurrent Neural Networks (RNNs) setting. In this paper, we propose an approach to train intrinsically sparse RNNs with a fixed parameter count in one single run, without compromising performance. During training, we allow RNN layers to have a non-uniform redistribution across cell gates for better regularization. Further, we propose SNT-ASGD, a novel variant of the averaged stochastic gradient optimizer, which significantly improves the performance of all sparse training methods for RNNs. Using these strategies, we achieve state-of-the-art sparse training results, better than the dense-to-sparse methods, with various types of RNNs on Penn TreeBank and Wikitext-2 datasets. Our codes are available at https://github.com/Shiweiliuiiiiiii/Selfish-RNN.

In this paper, we introduce a new perspective on training deep neural networks capable of state-of-the-art performance without the need for the expensive over-parameterization by proposing the concept of In-Time Over-Parameterization (ITOP) in sparse training. By starting from a random sparse network and continuously exploring sparse connectivities during training, we can perform an Over-Parameterization over the course of training, closing the gap in the expressibility between sparse training and dense training. We further use ITOP to understand the underlying mechanism of Dynamic Sparse Training (DST) and discover that the benefits of DST come from its ability to consider across time all possible parameters when searching for the optimal sparse connectivity. As long as sufficient parameters have been reliably explored, DST can outperform the dense neural network by a large margin. We present a series of experiments to support our conjecture and achieve the state-of-the-art sparse training performance with ResNet-50 on ImageNet. More impressively, ITOP achieves dominant performance over the overparameterization-based sparse methods at extreme sparsities. When trained with ResNet-34 on CIFAR-100, ITOP can match the performance of the dense model at an extreme sparsity 98%.

Works on lottery ticket hypothesis (LTH) and single-shot network pruning (SNIP) have raised a lot of attention currently on post-training pruning (iterative magnitude pruning), and before-training pruning (pruning at initialization). The former method suffers from an extremely large computation cost and the latter usually struggles with insufficient performance. In comparison, during-training pruning, a class of pruning methods that simultaneously enjoys the training/inference efficiency and the comparable performance, temporarily, has been less explored. To better understand during-training pruning, we quantitatively study the effect of pruning throughout training from the perspective of pruning plasticity (the ability of the pruned networks to recover the original performance). Pruning plasticity can help explain several other empirical observations about neural network pruning in literature. We further find that pruning plasticity can be substantially improved by injecting a brain-inspired mechanism called neuroregeneration, i.e., to regenerate the same number of connections as pruned. We design a novel gradual magnitude pruning (GMP) method, named gradual pruning with zero-cost neuroregeneration (GraNet), that advances state of the art. Perhaps most impressively, its sparse-to-sparse version for the first time boosts the sparse-to-sparse training performance over various dense-to-sparse methods with ResNet-50 on ImageNet without extending the training time. We release all codes in https://github.com/Shiweiliuiiiiiii/GraNet.

Abstract Artificial neural networks (ANNs) have emerged as hot topics in the research community. Despite the success of ANNs, it is challenging to train and deploy modern ANNs on commodity hardware due to the ever-increasing model size and the unprecedented growth in the data volumes. Particularly for microarray data, the very high dimensionality and the small number of samples make it difficult for machine learning techniques to handle. Furthermore, specialized hardware such as graphics processing unit (GPU) is expensive. Sparse neural networks are the leading approaches to address these challenges. However, off-the-shelf sparsity-inducing techniques either operate from a pretrained model or enforce the sparse structure via binary masks. The training efficiency of sparse neural networks cannot be obtained practically. In this paper, we introduce a technique allowing us to train truly sparse neural networks with fixed parameter count throughout training. Our experimental results demonstrate that our method can be applied directly to handle high-dimensional data, while achieving higher accuracy than the traditional two-phase approaches. Moreover, we have been able to create truly sparse multilayer perceptron models with over one million neurons and to train them on a typical laptop without GPU ( https://github.com/dcmocanu/sparse-evolutionary-artificial-neural-networks/tree/master/SET-MLP-Sparse-Python-Data-Structures ), this being way beyond what is possible with any state-of-the-art technique.

Abstract Recurrent neural networks (RNNs) have achieved state-of-the-art performances on various applications. However, RNNs are prone to be memory-bandwidth limited in practical applications and need both long periods of training and inference time. The aforementioned problems are at odds with training and deploying RNNs on resource-limited devices where the memory and floating-point operations (FLOPs) budget are strictly constrained. To address this problem, conventional model compression techniques usually focus on reducing inference costs, operating on a costly pre-trained model. Recently, dynamic sparse training has been proposed to accelerate the training process by directly training sparse neural networks from scratch. However, previous sparse training techniques are mainly designed for convolutional neural networks and multi-layer perceptron. In this paper, we introduce a method to train intrinsically sparse RNN models with a fixed number of parameters and floating-point operations (FLOPs) during training. We demonstrate state-of-the-art sparse performance with long short-term memory and recurrent highway networks on widely used tasks, language modeling, and text classification. We simply use the results to advocate that, contrary to the general belief that training a sparse neural network from scratch leads to worse performance than dense networks, sparse training with adaptive connectivity can usually achieve better performance than dense models for RNNs.

Deep neural networks perform well on test data when they are highly overparameterized, which, however, also leads to large cost to train and deploy them. As a leading approach to address this problem, sparse neural networks have been widely used to significantly reduce the size of networks, making them more efficient during training and deployment, without compromising performance. Recently, sparse neural networks, either compressed from a pre-trained model or obtained by training from scratch, have been observed to be able to generalize as well as or even better than their dense counterparts. However, conventional techniques to find well fitted sparse sub-networks are expensive and the mechanisms underlying this phenomenon are far from clear. To tackle these problems, this Ph.D. research aims to study the generalization of sparse neural networks, and to propose more efficient approaches that can yield sparse neural networks with generalization bounds.