Training neural networks which are robust to adversarial attacks remains an important problem in deep learning, especially as heavily overparameterized models are adopted in safety-critical settings. Drawing from recent work which reformulates the training problems for two-layer ReLU and polynomial activation networks as convex programs, we devise a convex semidefinite program (SDP) for adversarial training of two-layer polynomial activation networks and prove that the convex SDP achieves the same globally optimal solution as its nonconvex counterpart. The convex SDP is observed to improve robust test accuracy against $\ell_\infty$ attacks relative to the original convex training formulation on multiple datasets. Additionally, we present scalable implementations of adversarial training for two-layer polynomial and ReLU networks which are compatible with standard machine learning libraries and GPU acceleration. Leveraging these implementations, we retrain the final two fully connected layers of a Pre-Activation ResNet-18 model on the CIFAR-10 dataset with both polynomial and ReLU activations. The two `robustified' models achieve significantly higher robust test accuracies against $\ell_\infty$ attacks than a Pre-Activation ResNet-18 model trained with sharpness-aware minimization, demonstrating the practical utility of convex adversarial training on large-scale problems.
Training large Deep Neural Network (DNN) models requires thousands of GPUs over the course of several days or weeks. At this scale, failures are frequent and can have a big impact on training throughput. Utilizing spare GPU servers to mitigate performance loss becomes increasingly costly as model sizes grow. ReCycle is a system designed for efficient DNN training in the presence of failures, without relying on spare servers. It exploits the inherent functional redundancy in distributed training systems -- where servers across data-parallel groups store the same model parameters -- and pipeline schedule bubbles within each data-parallel group. When servers fails, ReCycle dynamically re-routes micro-batches to data-parallel peers, allowing for uninterrupted training despite multiple failures. However, this re-routing can create imbalances across pipeline stages, leading to reduced training throughput. To address this, ReCycle introduces two key optimizations that ensure re-routed micro-batches are processed within the original pipeline schedule's bubbles. First, it decouples the backward pass into two phases: one for computing gradients for the input and another for calculating gradients for the parameters. Second, it avoids synchronization across pipeline stages by staggering the optimizer step. Together, these optimizations enable adaptive pipeline schedules that minimize or even eliminate training throughput degradation during failures. We describe a prototype for ReCycle and show that it achieves high training throughput under multiple failures, outperforming recent proposals for fault-tolerant training such as Oobleck and Bamboo by up to $1.46\times$ and $1.64\times$, respectively.
The advent of Transformers marked a significant breakthrough in sequence modelling, providing a highly performant architecture capable of leveraging GPU parallelism. However, Transformers are computationally expensive at inference time, limiting their applications, particularly in low-resource settings (e.g., mobile and embedded devices). Addressing this, we (1) begin by showing that attention can be viewed as a special Recurrent Neural Network (RNN) with the ability to compute its \textit{many-to-one} RNN output efficiently. We then (2) show that popular attention-based models such as Transformers can be viewed as RNN variants. However, unlike traditional RNNs (e.g., LSTMs), these models cannot be updated efficiently with new tokens, an important property in sequence modelling. Tackling this, we (3) introduce a new efficient method of computing attention's \textit{many-to-many} RNN output based on the parallel prefix scan algorithm. Building on the new attention formulation, we (4) introduce \textbf{Aaren}, an attention-based module that can not only (i) be trained in parallel (like Transformers) but also (ii) be updated efficiently with new tokens, requiring only constant memory for inferences (like traditional RNNs). Empirically, we show Aarens achieve comparable performance to Transformers on $38$ datasets spread across four popular sequential problem settings: reinforcement learning, event forecasting, time series classification, and time series forecasting tasks while being more time and memory-efficient.
In this paper, we present a low-budget and high-authenticity bidirectional telepresence system, Tele-Aloha, targeting peer-to-peer communication scenarios. Compared to previous systems, Tele-Aloha utilizes only four sparse RGB cameras, one consumer-grade GPU, and one autostereoscopic screen to achieve high-resolution (2048x2048), real-time (30 fps), low-latency (less than 150ms) and robust distant communication. As the core of Tele-Aloha, we propose an efficient novel view synthesis algorithm for upper-body. Firstly, we design a cascaded disparity estimator for obtaining a robust geometry cue. Additionally a neural rasterizer via Gaussian Splatting is introduced to project latent features onto target view and to decode them into a reduced resolution. Further, given the high-quality captured data, we leverage weighted blending mechanism to refine the decoded image into the final resolution of 2K. Exploiting world-leading autostereoscopic display and low-latency iris tracking, users are able to experience a strong three-dimensional sense even without any wearable head-mounted display device. Altogether, our telepresence system demonstrates the sense of co-presence in real-life experiments, inspiring the next generation of communication.
Parameter-efficient fine-tuning (PEFT) has emerged as a popular solution for adapting pre-trained Vision Transformer (ViT) models to downstream applications. While current PEFT methods have achieved parameter efficiency, they overlook the efficiency of computation and GPU memory during both fine-tuning and inference, falling short of practical requirements. In this paper, we propose \textbf{Sparse-Tuning}, a novel PEFT method that accounts for the information redundancy in images and videos to boost the above efficiency. By sparsely preserving the semantic-relevant tokens and merging irrelevant ones, Sparse-Tuning minimizes the quantity of tokens processed at each layer, leading to a quadratic reduction in computational and memory overhead. To align our token sparsification strategy suitably with fine-tuning purposes, we further design Dense Adapters that establish dense connections from shallow layers to deeper layers. These Dense Adapters integrate multi-level local features to enrich the current tokens, improving both token preservation and model adaptation. Empirical results on VTAB-1K, three image datasets, and two video datasets show that our Sparse-Tuning reduces GFLOPs to \textbf{62\%-70\%} of the original ViT-B while achieving state-of-the-art performance. Source code is available at \url{https://github.com/liuting20/Sparse-Tuning}.
Advances in latent diffusion models (LDMs) have revolutionized high-resolution image generation, but the design space of the autoencoder that is central to these systems remains underexplored. In this paper, we introduce LiteVAE, a new autoencoder design for LDMs, which leverages the 2D discrete wavelet transform to enhance scalability and computational efficiency over standard variational autoencoders (VAEs) with no sacrifice in output quality. We investigate the training methodologies and the decoder architecture of LiteVAE and propose several enhancements that improve the training dynamics and reconstruction quality. Our base LiteVAE model matches the quality of the established VAEs in current LDMs with a six-fold reduction in encoder parameters, leading to faster training and lower GPU memory requirements, while our larger model outperforms VAEs of comparable complexity across all evaluated metrics (rFID, LPIPS, PSNR, and SSIM).
Point cloud videos can faithfully capture real-world spatial geometries and temporal dynamics, which are essential for enabling intelligent agents to understand the dynamically changing world. However, designing an effective 4D backbone remains challenging, mainly due to the irregular and unordered distribution of points and temporal inconsistencies across frames. Also, recent transformer-based 4D backbones commonly suffer from large computational costs due to their quadratic complexity, particularly for long video sequences.To address these challenges, we propose a novel point cloud video understanding backbone purely based on the State Space Models (SSMs). Specifically, we first disentangle space and time in 4D video sequences and then establish the spatio-temporal correlation with our designed Mamba blocks. The Intra-frame Spatial Mamba module is developed to encode locally similar geometric structures within a certain temporal stride. Subsequently, locally correlated tokens are delivered to the Inter-frame Temporal Mamba module, which integrates long-term point features across the entire video with linear complexity. Our proposed Mamba4d achieves competitive performance on the MSR-Action3D action recognition (+10.4% accuracy), HOI4D action segmentation (+0.7 F1 Score), and Synthia4D semantic segmentation (+0.19 mIoU) datasets. Especially, for long video sequences, our method has a significant efficiency improvement with 87.5% GPU memory reduction and 5.36 times speed-up.
One of the guiding principles for designing AI-based weather forecasting systems is to embed physical constraints as inductive priors in the neural network architecture. A popular prior is locality, where the atmospheric data is processed with local neural interactions, like 3D convolutions or 3D local attention windows as in Pangu-Weather. On the other hand, some works have shown great success in weather forecasting without this locality principle, at the cost of a much higher parameter count. In this paper, we show that the 3D local processing in Pangu-Weather is computationally sub-optimal. We design ArchesWeather, a transformer model that combines 2D attention with a column-wise attention-based feature interaction module, and demonstrate that this design improves forecasting skill. ArchesWeather is trained at 1.5{\deg} resolution and 24h lead time, with a training budget of a few GPU-days and a lower inference cost than competing methods. An ensemble of four of our models shows better RMSE scores than the IFS HRES and is competitive with the 1.4{\deg} 50-members NeuralGCM ensemble for one to three days ahead forecasting. Our code and models are publicly available at https://github.com/gcouairon/ArchesWeather.
Training large AI models such as LLMs and DLRMs costs massive GPUs and computing time. The high training cost has become only affordable to big tech companies, meanwhile also causing increasing concerns about the environmental impact. This paper presents CoMERA, a Computing- and Memory-Efficient training method via Rank-Adaptive tensor optimization. CoMERA achieves rank-adaptive tensor-compressed (pre)-training via a multi-objective optimization formulation and improves the training to provide both a high compression ratio and excellent accuracy in the training process. Our optimized numerical computation (e.g., optimized tensorized embedding and tensor-network contractions) and GPU implementation eliminate part of the run-time overhead in the tensorized training on GPU. This leads to, for the first time, $2-3\times$ speedup per training epoch compared with standard training. CoMERA also outperforms the recent GaLore in terms of both memory and computing efficiency. Specifically, CoMERA is $2\times$ faster per training epoch and $9\times$ more memory-efficient than GaLore on a tested six-encoder transformer with single-batch training. Our method also shows $\sim 2\times$ speedup than standard pre-training on a BERT-like code-generation LLM while achieving $4.23\times$ compression ratio in pre-training. With further HPC optimization, CoMERA may reduce the pre-training cost of many other LLMs. An implementation of CoMERA is available at https://github.com/ziyangjoy/CoMERA.
KV cache stores key and value states from previous tokens to avoid re-computation, yet it demands substantial storage space, especially for long sequences. Adaptive KV cache compression seeks to discern the saliency of tokens, preserving vital information while aggressively compressing those of less importance. However, previous methods of this approach exhibit significant performance degradation at high compression ratios due to inaccuracies in identifying salient tokens. In this paper, we present ZipCache, an accurate and efficient KV cache quantization method for LLMs. First, we construct a strong baseline for quantizing KV cache. Through the proposed channel-separable tokenwise quantization scheme, the memory overhead of quantization parameters are substantially reduced compared to fine-grained groupwise quantization. To enhance the compression ratio, we propose normalized attention score as an effective metric for identifying salient tokens by considering the lower triangle characteristics of the attention matrix. Moreover, we develop an efficient approximation method that decouples the saliency metric from full attention scores, enabling compatibility with fast attention implementations like FlashAttention. Extensive experiments demonstrate that ZipCache achieves superior compression ratios, fast generation speed and minimal performance losses compared with previous KV cache compression methods. For instance, when evaluating Mistral-7B model on GSM8k dataset, ZipCache is capable of compressing the KV cache by $4.98\times$, with only a $0.38\%$ drop in accuracy. In terms of efficiency, ZipCache also showcases a $37.3\%$ reduction in prefill-phase latency, a $56.9\%$ reduction in decoding-phase latency, and a $19.8\%$ reduction in GPU memory usage when evaluating LLaMA3-8B model with a input length of $4096$.
Large language models (LLMs) are trained on a vast amount of human-written data, but data providers often remain uncredited. In response to this issue, data valuation (or data attribution), which quantifies the contribution or value of each data to the model output, has been discussed as a potential solution. Nevertheless, applying existing data valuation methods to recent LLMs and their vast training datasets has been largely limited by prohibitive compute and memory costs. In this work, we focus on influence functions, a popular gradient-based data valuation method, and significantly improve its scalability with an efficient gradient projection strategy called LoGra that leverages the gradient structure in backpropagation. We then provide a theoretical motivation of gradient projection approaches to influence functions to promote trust in the data valuation process. Lastly, we lower the barrier to implementing data valuation systems by introducing LogIX, a software package that can transform existing training code into data valuation code with minimal effort. In our data valuation experiments, LoGra achieves competitive accuracy against more expensive baselines while showing up to 6,500x improvement in throughput and 5x reduction in GPU memory usage when applied to Llama3-8B-Instruct and the 1B-token dataset.
With the rapid growth of the Internet of Things ecosystem, Automatic Modulation Classification (AMC) has become increasingly paramount. However, extended signal lengths offer a bounty of information, yet impede the model's adaptability, introduce more noise interference, extend the training and inference time, and increase storage overhead. To bridge the gap between these requisites, we propose a novel AMC framework, designated as the Mamba-based Automatic Modulation ClassificAtion (MAMCA). Our method adeptly addresses the accuracy and efficiency requirements for long-sequence AMC. Specifically, we introduce the Selective State Space Model as the backbone, enhancing the model efficiency by reducing the dimensions of the state matrices and diminishing the frequency of information exchange across GPU memories. We design a denoising-capable unit to elevate the network's performance under low signal-to-noise radio. Rigorous experimental evaluations on the publicly available dataset RML2016.10, along with our synthetic dataset within multiple quadrature amplitude modulations and lengths, affirm that MAMCA delivers superior recognition accuracy while necessitating minimal computational time and memory occupancy. Codes are available on https://github.com/ZhangYezhuo/MAMCA.
Large Language Models (LLMs) have shown remarkable comprehension abilities but face challenges in GPU memory usage during inference, hindering their scalability for real-time applications like chatbots. To accelerate inference, we store computed keys and values (KV cache) in the GPU memory. Existing methods study the KV cache compression to reduce memory by pruning the pre-computed KV cache. However, they neglect the inter-layer dependency between layers and huge memory consumption in pre-computation. To explore these deficiencies, we find that the number of crucial keys and values that influence future generations decreases layer by layer and we can extract them by the consistency in attention weights. Based on the findings, we propose PyramidInfer, a method that compresses the KV cache by layer-wise retaining crucial context. PyramidInfer saves significant memory by computing fewer keys and values without sacrificing performance. Experimental results show PyramidInfer improves 2.2x throughput compared to Accelerate with over 54% GPU memory reduction in KV cache.
We present personalized residuals and localized attention-guided sampling for efficient concept-driven generation using text-to-image diffusion models. Our method first represents concepts by freezing the weights of a pretrained text-conditioned diffusion model and learning low-rank residuals for a small subset of the model's layers. The residual-based approach then directly enables application of our proposed sampling technique, which applies the learned residuals only in areas where the concept is localized via cross-attention and applies the original diffusion weights in all other regions. Localized sampling therefore combines the learned identity of the concept with the existing generative prior of the underlying diffusion model. We show that personalized residuals effectively capture the identity of a concept in ~3 minutes on a single GPU without the use of regularization images and with fewer parameters than previous models, and localized sampling allows using the original model as strong prior for large parts of the image.
The K-Means clustering using LLoyd's algorithm is an iterative approach to partition the given dataset into K different clusters. The algorithm assigns each point to the cluster based on the following objective function \[\ \min \Sigma_{i=1}^{n}||x_i-\mu_{x_i}||^2\] The serial algorithm involves iterative steps where we compute the distance of each datapoint from the centroids and assign the datapoint to the nearest centroid. This approach is essentially known as the expectation-maximization step. Clustering involves extensive computations to calculate distances at each iteration, which increases as the number of data points increases. This provides scope for parallelism. However, we must ensure that in a parallel process, each thread has access to the updated centroid value and no racing condition exists on any centroid values. We will compare two different approaches in this project. The first approach is an OpenMP flat synchronous method where all processes are run in parallel, and we use synchronization to ensure safe updates of clusters. The second approach we adopt is a GPU based parallelization approach using OpenACC wherein we will try to make use of GPU architecture to parallelize chunks of the algorithm to observe decreased computation time. We will analyze metrics such as speed up, efficiency,time taken with varying data points, and number of processes to compare the two approaches and understand the relative performance improvement we can get.
Transformers have revolutionized deep learning across various tasks, including audio representation learning, due to their powerful modeling capabilities. However, they often suffer from quadratic complexity in both GPU memory usage and computational inference time, affecting their efficiency. Recently, state space models (SSMs) like Mamba have emerged as a promising alternative, offering a more efficient approach by avoiding these complexities. Given these advantages, we explore the potential of SSM-based models in audio tasks. In this paper, we introduce Self-Supervised Audio Mamba (SSAMBA), the first self-supervised, attention-free, and SSM-based model for audio representation learning. SSAMBA leverages the bidirectional Mamba to capture complex audio patterns effectively. We incorporate a self-supervised pretraining framework that optimizes both discriminative and generative objectives, enabling the model to learn robust audio representations from large-scale, unlabeled datasets. We evaluated SSAMBA on various tasks such as audio classification, keyword spotting, and speaker identification. Our results demonstrate that SSAMBA outperforms the Self-Supervised Audio Spectrogram Transformer (SSAST) in most tasks. Notably, SSAMBA is approximately 92.7% faster in batch inference speed and 95.4% more memory-efficient than SSAST for the tiny model size with an input token size of 22k. These efficiency gains, combined with superior performance, underscore the effectiveness of SSAMBA's architectural innovation, making it a compelling choice for a wide range of audio processing applications.
Accurate 6-DoF pose estimation of surgical instruments during minimally invasive surgeries can substantially improve treatment strategies and eventual surgical outcome. Existing deep learning methods have achieved accurate results, but they require custom approaches for each object and laborious setup and training environments often stretching to extensive simulations, whilst lacking real-time computation. We propose a general-purpose approach of data acquisition for 6-DoF pose estimation tasks in X-ray systems, a novel and general purpose YOLOv5-6D pose architecture for accurate and fast object pose estimation and a complete method for surgical screw pose estimation under acquisition geometry consideration from a monocular cone-beam X-ray image. The proposed YOLOv5-6D pose model achieves competitive results on public benchmarks whilst being considerably faster at 42 FPS on GPU. In addition, the method generalizes across varying X-ray acquisition geometry and semantic image complexity to enable accurate pose estimation over different domains. Finally, the proposed approach is tested for bone-screw pose estimation for computer-aided guidance during spine surgeries. The model achieves a 92.41% by the 0.1 ADD-S metric, demonstrating a promising approach for enhancing surgical precision and patient outcomes. The code for YOLOv5-6D is publicly available at https://github.com/cviviers/YOLOv5-6D-Pose
Given a Large Language Model (LLM) generation, how can we identify which training data led to this generation? In this paper, we proposed RapidIn, a scalable framework adapting to LLMs for estimating the influence of each training data. The proposed framework consists of two stages: caching and retrieval. First, we compress the gradient vectors by over 200,000x, allowing them to be cached on disk or in GPU/CPU memory. Then, given a generation, RapidIn efficiently traverses the cached gradients to estimate the influence within minutes, achieving over a 6,326x speedup. Moreover, RapidIn supports multi-GPU parallelization to substantially accelerate caching and retrieval. Our empirical result confirms the efficiency and effectiveness of RapidIn.
Transformer-based encoder-decoder models have demonstrated impressive results in chemical reaction prediction tasks. However, these models typically rely on pretraining using tens of millions of unlabelled molecules, which can be time-consuming and GPU-intensive. One of the central questions we aim to answer in this work is: Can FlanT5 and ByT5, the encode-decoder models pretrained solely on language data, be effectively specialised for organic reaction prediction through task-specific fine-tuning? We conduct a systematic empirical study on several key issues of the process, including tokenisation, the impact of (SMILES-oriented) pretraining, fine-tuning sample efficiency, and decoding algorithms at inference. Our key findings indicate that although being pretrained only on language tasks, FlanT5 and ByT5 provide a solid foundation to fine-tune for reaction prediction, and thus become `chemistry domain compatible' in the process. This suggests that GPU-intensive and expensive pretraining on a large dataset of unlabelled molecules may be useful yet not essential to leverage the power of language models for chemistry. All our models achieve comparable Top-1 and Top-5 accuracy although some variation across different models does exist. Notably, tokenisation and vocabulary trimming slightly affect final performance but can speed up training and inference; The most efficient greedy decoding strategy is very competitive while only marginal gains can be achieved from more sophisticated decoding algorithms. In summary, we evaluate FlanT5 and ByT5 across several dimensions and benchmark their impact on organic reaction prediction, which may guide more effective use of these state-of-the-art language models for chemistry-related tasks in the future.
The increasing size of language models necessitates a thorough analysis across multiple dimensions to assess trade-offs among crucial hardware metrics such as latency, energy consumption, GPU memory usage, and performance. Identifying optimal model configurations under specific hardware constraints is becoming essential but remains challenging due to the computational load of exhaustive training and evaluation on multiple devices. To address this, we introduce HW-GPT-Bench, a hardware-aware benchmark that utilizes surrogate predictions to approximate various hardware metrics across 13 devices of architectures in the GPT-2 family, with architectures containing up to 1.55B parameters. Our surrogates, via calibrated predictions and reliable uncertainty estimates, faithfully model the heteroscedastic noise inherent in the energy and latency measurements. To estimate perplexity, we employ weight-sharing techniques from Neural Architecture Search (NAS), inheriting pretrained weights from the largest GPT-2 model. Finally, we demonstrate the utility of HW-GPT-Bench by simulating optimization trajectories of various multi-objective optimization algorithms in just a few seconds.