Anshumali Shrivastava

Zichang Liu, Jue Wang, Tri Dao et al. · eth-zurich

Large language models (LLMs) with hundreds of billions of parameters have sparked a new wave of exciting AI applications. However, they are computationally expensive at inference time. Sparsity is a natural approach to reduce this cost, but existing methods either require costly retraining, have to forgo LLM's in-context learning ability, or do not yield wall-clock time speedup on modern hardware. We hypothesize that contextual sparsity, which are small, input-dependent sets of attention heads and MLP parameters that yield approximately the same output as the dense model for a given input, can address these issues. We show that contextual sparsity exists, that it can be accurately predicted, and that we can exploit it to speed up LLM inference in wall-clock time without compromising LLM's quality or in-context learning ability. Based on these insights, we propose DejaVu, a system that uses a low-cost algorithm to predict contextual sparsity on the fly given inputs to each layer, along with an asynchronous and hardware-aware implementation that speeds up LLM inference. We validate that DejaVu can reduce the inference latency of OPT-175B by over 2X compared to the state-of-the-art FasterTransformer, and over 6X compared to the widely used Hugging Face implementation, without compromising model quality. The code is available at https://github.com/FMInference/DejaVu.

Aditya Desai, Benjamin Meisburger, Zichang Liu et al.

Recommendation systems (RS) for items (e.g., movies, books) and ads are widely used to tailor content to users on various internet platforms. Traditionally, recommendation models are trained on a central server. However, due to rising concerns for data privacy and regulations like the GDPR, federated learning is an increasingly popular paradigm in which data never leaves the client device. Applying federated learning to recommendation models is non-trivial due to large embedding tables, which often exceed the memory constraints of most user devices. To include data from all devices in federated learning, we must enable collective training of embedding tables on devices with heterogeneous memory capacities. Current solutions to heterogeneous federated learning can only accommodate a small range of capacities and thus limit the number of devices that can participate in training. We present Federated Averaging in Random subspaces (FAIR), which allows arbitrary compression of embedding tables based on device capacity and ensures the participation of all devices in training. FAIR uses what we call consistent and collapsible subspaces defined by hashing-based random projections to jointly train large embedding tables while using varying amounts of compression on user devices. We evaluate FAIR on Neural Collaborative Filtering tasks with multiple datasets and verify that FAIR can gather and share information from a wide range of devices with varying capacities, allowing for seamless collaboration. We prove the convergence of FAIR in the homogeneous setting with non-i.i.d data distribution. Our code is open source at {https://github.com/apd10/FLCF}

Zichang Liu, Zhiqiang Tang, Xingjian Shi et al.

The ability to jointly learn from multiple modalities, such as text, audio, and visual data, is a defining feature of intelligent systems. While there have been promising advances in designing neural networks to harness multimodal data, the enormous success of data augmentation currently remains limited to single-modality tasks like image classification. Indeed, it is particularly difficult to augment each modality while preserving the overall semantic structure of the data; for example, a caption may no longer be a good description of an image after standard augmentations have been applied, such as translation. Moreover, it is challenging to specify reasonable transformations that are not tailored to a particular modality. In this paper, we introduce LeMDA, Learning Multimodal Data Augmentation, an easy-to-use method that automatically learns to jointly augment multimodal data in feature space, with no constraints on the identities of the modalities or the relationship between modalities. We show that LeMDA can (1) profoundly improve the performance of multimodal deep learning architectures, (2) apply to combinations of modalities that have not been previously considered, and (3) achieve state-of-the-art results on a wide range of applications comprised of image, text, and tabular data.

Tianyi Zhang, Anshumali Shrivastava

Large language models (LLMs) have shown immense potential across various domains, but their high memory requirements and inference costs remain critical challenges for deployment. Post-training quantization (PTQ) has emerged as a promising technique to reduce memory requirements and decoding latency. However, recent accurate quantization methods often depend on specialized computations or custom data formats to achieve better model quality, which limits their compatibility with popular frameworks, as they require dedicated inference kernels tailored to specific hardware and software platforms, hindering wider adoption. Furthermore, many competitive methods have high resource requirements and computational overhead for quantizing models, making it challenging to scale them to hundreds of billions of parameters. In response to these challenges, we propose LeanQuant (Loss-Error-Aware Network Quantization), a novel quantization method that is accurate, versatile, and scalable. In the existing popular iterative loss-error-based quantization framework, we identify a critical limitation in prior methods: the min-max affine quantization grid fails to preserve model quality due to outliers in inverse Hessian diagonals. To overcome this fundamental issue, we propose learning loss-error-aware grids, instead of using non-adaptive min-max affine grids. Our approach not only produces quantized models that are more accurate but also generalizes to a wider range of quantization types, including affine and non-uniform quantization, enhancing compatibility with more frameworks. Extensive experiments with recent LLMs demonstrate that LeanQuant is highly accurate, comparing favorably against competitive baselines in model quality, and scalable, achieving very accurate quantization of Llama-3.1 405B, one of the largest open-source LLMs to date, using two Quadro RTX 8000-48GB GPUs in 21 hours.

Anshumali Shrivastava, Zhao Song, Zhaozhuo Xu

Graph-based algorithms have demonstrated state-of-the-art performance in the nearest neighbor search (NN-Search) problem. These empirical successes urge the need for theoretical results that guarantee the search quality and efficiency of these algorithms. However, there exists a practice-to-theory gap in the graph-based NN-Search algorithms. Current theoretical literature focuses on greedy search on exact near neighbor graph while practitioners use approximate near neighbor graph (ANN-Graph) to reduce the preprocessing time. This work bridges this gap by presenting the theoretical guarantees of solving NN-Search via greedy search on ANN-Graph for low dimensional and dense vectors. To build this bridge, we leverage several novel tools from computational geometry. Our results provide quantification of the trade-offs associated with the approximation while building a near neighbor graph. We hope our results will open the door for more provable efficient graph-based NN-Search algorithms.

Aditya Desai, Anshumali Shrivastava

When considering a model architecture, there are several ways to reduce its memory footprint. Historically, popular approaches included selecting smaller architectures and creating sparse networks through pruning. More recently, randomized parameter-sharing (RPS) methods have gained traction for model compression at start of training. In this paper, we comprehensively assess the trade-off between memory and accuracy across RPS, pruning techniques, and building smaller models. Our findings demonstrate that RPS, which is both data and model-agnostic, consistently outperforms/matches smaller models and all moderately informed pruning strategies, such as MAG, SNIP, SYNFLOW, and GRASP, across the entire compression range. This advantage becomes particularly pronounced in higher compression scenarios. Notably, even when compared to highly informed pruning techniques like Lottery Ticket Rewinding (LTR), RPS exhibits superior performance in high compression settings. This points out inherent capacity advantage that RPS enjoys over sparse models. Theoretically, we establish RPS as a superior technique in terms of memory-efficient representation when compared to pruning for linear models. This paper argues in favor of paradigm shift towards RPS based models. During our rigorous evaluation of RPS, we identified issues in the state-of-the-art RPS technique ROAST, specifically regarding stability (ROAST's sensitivity to initialization hyperparameters, often leading to divergence) and Pareto-continuity (ROAST's inability to recover the accuracy of the original model at zero compression). We provably address both of these issues. We refer to the modified RPS, which incorporates our improvements, as STABLE-RPS.

Nicholas Meisburger, Vihan Lakshman, Benito Geordie et al.

Efficient large-scale neural network training and inference on commodity CPU hardware is of immense practical significance in democratizing deep learning (DL) capabilities. Presently, the process of training massive models consisting of hundreds of millions to billions of parameters requires the extensive use of specialized hardware accelerators, such as GPUs, which are only accessible to a limited number of institutions with considerable financial resources. Moreover, there is often an alarming carbon footprint associated with training and deploying these models. In this paper, we take a step towards addressing these challenges by introducing BOLT, a sparse deep learning library for training large-scale search and recommendation models on standard CPU hardware. BOLT provides a flexible, high-level API for constructing models that will be familiar to users of existing popular DL frameworks. By automatically tuning specialized hyperparameters, BOLT also abstracts away the algorithmic details of sparse network training. We evaluate BOLT on a number of information retrieval tasks including product recommendations, text classification, graph neural networks, and personalization. We find that our proposed system achieves competitive performance with state-of-the-art techniques at a fraction of the cost and energy consumption and an order-of-magnitude faster inference time. BOLT has also been successfully deployed by multiple businesses to address critical problems, and we highlight one customer case study in the field of e-commerce.

Aditya Desai, Anshumali Shrivastava

Embedding tables dominate industrial-scale recommendation model sizes, using up to terabytes of memory. A popular and the largest publicly available machine learning MLPerf benchmark on recommendation data is a Deep Learning Recommendation Model (DLRM) trained on a terabyte of click-through data. It contains 100GB of embedding memory (25+Billion parameters). DLRMs, due to their sheer size and the associated volume of data, face difficulty in training, deploying for inference, and memory bottlenecks due to large embedding tables. This paper analyzes and extensively evaluates a generic parameter sharing setup (PSS) for compressing DLRM models. We show theoretical upper bounds on the learnable memory requirements for achieving $(1 \pm ε)$ approximations to the embedding table. Our bounds indicate exponentially fewer parameters suffice for good accuracy. To this end, we demonstrate a PSS DLRM reaching 10000$\times$ compression on criteo-tb without losing quality. Such a compression, however, comes with a caveat. It requires 4.5 $\times$ more iterations to reach the same saturation quality. The paper argues that this tradeoff needs more investigations as it might be significantly favorable. Leveraging the small size of the compressed model, we show a 4.3$\times$ improvement in training latency leading to similar overall training times. Thus, in the tradeoff between system advantage of a small DLRM model vs. slower convergence, we show that scales are tipped towards having a smaller DLRM model, leading to faster inference, easier deployment, and similar training times.

Benjamin Coleman, David Torres Ramos, Vihan Lakshman et al.

Lookup tables are a fundamental structure in many data processing and systems applications. Examples include tokenized text in NLP, quantized embedding collections in recommendation systems, integer sketches for streaming data, and hash-based string representations in genomics. With the increasing size of web-scale data, such applications often require compression techniques that support fast random $O(1)$ lookup of individual parameters directly on the compressed data (i.e. without blockwise decompression in RAM). While the community has proposd a number of succinct data structures that support queries over compressed representations, these approaches do not fully leverage the low-entropy structure prevalent in real-world workloads to reduce space. Inspired by recent advances in static function construction techniques, we propose a space-efficient representation of immutable key-value data, called CARAMEL, specifically designed for the case where the values are multi-sets. By carefully combining multiple compressed static functions, CARAMEL occupies space proportional to the data entropy with low memory overheads and minimal lookup costs. We demonstrate 1.25-16x compression on practical lookup tasks drawn from real-world systems, improving upon established techniques, including a production-grade read-only database widely used for development within Amazon.com.

Sahil Joshi, Agniva Chowdhury, Wyatt Bellinger et al.

Exploiting sparsity during long-context inference is central to scaling large language models, as attention dominates the cost of autoregressive decoding. Sparse attention reduces this cost by restricting computation to a subset of tokens, but its effectiveness depends critically on efficient scoring and selection of relevant tokens at inference time. We revisit Locality-Sensitive Hashing (LSH) as a sparsification primitive and introduce SOCKET, a SOft Collision Kernel EsTimator that replaces hard bucket matches with probabilistic, similarity-aware aggregation. Our key insight is that hard LSH produces discrete collision signals and is therefore poorly suited for ranking. In contrast, soft LSH aggregates graded collision evidence across hash tables, preserving the stability of relative ordering among the true top-$k$ tokens. This transformation elevates LSH from a candidate-generation heuristic to a principled and mathematically grounded scoring kernel for sparse attention. Leveraging this property, SOCKET enables efficient token selection without ad-hoc voting mechanism, and matches or surpasses established sparse attention baselines across multiple long-context benchmarks using diverse set of models. With a custom CUDA kernel for scoring keys and a Flash Decode Triton backend for sparse attention, SOCKET achieves up to 1.5$\times$ higher throughput than FlashAttention, making it an effective tool for long-context inference. Code is open-sourced at https://github.com/amarka8/SOCKET.

Zhuang Wang, Zhaozhuo Xu, Jingyi Xi et al.

Distributed training is the de facto standard to scale up the training of deep learning models with multiple GPUs. Its performance bottleneck lies in communications for gradient synchronization. Although high tensor sparsity is widely observed, the optimal communication scheme to fully leverage sparsity is still missing. This paper aims to bridge this gap. We first analyze the characteristics of sparse tensors in popular models to understand the fundamentals of sparsity. We then systematically explore the design space of communication schemes for sparse tensors and find the optimal ones. These findings give a new understanding and inspire us to develop a holistic gradient synchronization system called Zen for sparse tensors. We demonstrate that Zen can achieve up to 5.09x speedup in communication time and up to $2.48\times$ speedup in training throughput compared to the state-of-the-art methods.

Aditya Desai, Keren Zhou, Anshumali Shrivastava

Advancements in deep learning are often associated with increasing model sizes. The model size dramatically affects the deployment cost and latency of deep models. For instance, models like BERT cannot be deployed on edge devices and mobiles due to their sheer size. As a result, most advances in Deep Learning are yet to reach the edge. Model compression has sought much-deserved attention in literature across natural language processing, vision, and recommendation domains. This paper proposes a model-agnostic, cache-friendly model compression approach: Random Operation Access Specific Tile (ROAST) hashing. ROAST collapses the parameters by clubbing them through a lightweight mapping. Notably, while clubbing these parameters, ROAST utilizes cache hierarchies by aligning the memory access pattern with the parameter access pattern. ROAST is up to $\sim 25 \times$ faster to train and $\sim 50 \times$ faster to infer than the popular parameter sharing method HashedNet. Additionally, ROAST introduces global weight sharing, which is empirically and theoretically superior to local weight sharing in HashedNet, and can be of independent interest in itself. With ROAST, we present the first compressed BERT, which is $100\times - 1000\times$ smaller but does not result in quality degradation. These compression levels on universal architecture like transformers are promising for the future of SOTA model deployment on resource-constrained devices like mobile and edge devices

Shabnam Daghaghi, Benjamin Coleman, Benito Geordie et al.

Data sampling is an effective method to improve the training speed of neural networks, with recent results demonstrating that it can even break the neural scaling laws. These results critically rely on high-quality scores to estimate the importance of an input to the network. We observe that there are two dominant strategies: static sampling, where the scores are determined before training, and dynamic sampling, where the scores can depend on the model weights. Static algorithms are computationally inexpensive but less effective than their dynamic counterparts, which can cause end-to-end slowdown due to their need to explicitly compute losses. To address this problem, we propose a novel sampling distribution based on nonparametric kernel regression that learns an effective importance score as the neural network trains. However, nonparametric regression models are too computationally expensive to accelerate end-to-end training. Therefore, we develop an efficient sketch-based approximation to the Nadaraya-Watson estimator. Using recent techniques from high-dimensional statistics and randomized algorithms, we prove that our Nadaraya-Watson sketch approximates the estimator with exponential convergence guarantees. Our sampling algorithm outperforms the baseline in terms of wall-clock time and accuracy on four datasets.

Sahil Joshi, Prithvi Dixit, Agniva Chowdhury et al.

Sparsity has long been a central theme in LLM efficiency, but its role in context processing remains unresolved. As LLM workloads shift toward longer contexts and agentic interactions, the compute and memory bottlenecks of attention become increasingly critical, raising the question of whether these constraints are fundamental. Our position is that these constraints are artificial and unnecessary, and that the future of LLM inference lies in extreme but principled sparsity along the context dimension. This position is supported by several strands of empirical and theoretical evidence. First, we find the insistence on dense attention unreasonable, since in a long context a query effectively projects O(N) attention information into a hidden space of dimension d << N, making the process inherently lossy. Second, we perform an extensive study of sparsity in LLMs spanning 20 models across five model families, varying context lengths, and different sparsity levels. We empirically demonstrate a strong trend: current LLMs, despite not being trained for context sparsity, are remarkably robust to inference-time decode sparsity across tasks of varying complexity, including retrieval, multi-hop QA, mathematical reasoning, and agentic coding. Importantly, we also show that current hardware is already sufficient to realize substantial gains from this sparsity. For example, our sparse decode kernels accelerate large-context processing by up to 10x over FlashInfer at 50x sparsity levels on hardware such as the H100. Overall, these results position extreme context sparsity not as a heuristic, but as a principled foundation for LLM inference, training, and architecture design: one that is both feasible and beneficial, and a compelling direction for future systems.

Tianyi Zhang, Jonah Wonkyu Yi, Bowen Yao et al.

Large language model inference on Central Processing Units (CPU) is challenging due to the vast quantities of expensive Multiply-Add (MAD) matrix operations in the attention computations. In this paper, we argue that there is a rare gem in modern CPUs, Single-Instruction-Multiple-Data (SIMD) registers, which allow for ultra-low-latency lookups in batch. We leverage this unique capability of CPUs to propose NoMAD-Attention, an efficient attention algorithm that replaces MAD operations with in-register lookups. Through hardware-aware algorithmic designs, NoMAD-Attention achieves the computation of attention scores using repeated fast accesses to SIMD registers despite their highly limited sizes. Moreover, NoMAD-Attention works with pre-trained attention-based LLMs without model finetuning. Empirical evaluations demonstrate that NoMAD-Attention maintains the quality of the original LLMs well, and speeds up the 4-bit quantized LLaMA-7B-based model by up to 2$\times$ at 16k context length. Our results are reproducible at https://github.com/tonyzhang617/nomad-dist.

Tianyi Zhang, Mohsen Hariri, Shaochen Zhong et al.

Large-scale AI models, such as Large Language Models (LLMs) and Diffusion Models (DMs), have grown rapidly in size, creating significant challenges for efficient deployment on resource-constrained hardware. In this paper, we introduce Dynamic-Length Float (DFloat11), a lossless compression framework that reduces LLM and DM size by 30% while preserving outputs that are bit-for-bit identical to the original model. DFloat11 is motivated by the low entropy in the BFloat16 weight representation of LLMs, which reveals significant inefficiency in the existing storage format. By applying entropy coding, DFloat11 assigns dynamic-length encodings to weights based on frequency, achieving near information-optimal compression without any loss of precision. To facilitate efficient inference with dynamic-length encodings, we develop a custom GPU kernel for fast online decompression. Our design incorporates the following: (i) compact, hierarchical lookup tables (LUTs) that fit within GPU SRAM for efficient decoding, (ii) a two-phase GPU kernel for coordinating thread read/write positions using lightweight auxiliary variables, and (iii) transformer-block-level decompression to minimize latency. Experiments on Llama 3.3, Qwen 3, Mistral 3, FLUX.1, and others validate our hypothesis that DFloat11 achieves around 30% model size reduction while preserving bit-for-bit identical outputs. Compared to a potential alternative of offloading parts of an uncompressed model to the CPU to meet memory constraints, DFloat11 achieves 2.3--46.2x higher throughput in token generation. With a fixed GPU memory budget, DFloat11 enables 5.7--14.9x longer generation lengths than uncompressed models. Notably, our method enables lossless inference of Llama 3.1 405B, an 810GB model, on a single node equipped with 8x80GB GPUs. Our code is available at https://github.com/LeanModels/DFloat11.

Haochen Zhang, Tianyi Zhang, Junze Yin et al.

Recommender systems play a pivotal role in providing relevant content to users. With the rapid development of large language models (LLMs), researchers have begun utilizing LLMs to build more powerful recommender systems. However, existing approaches that focus on aligning LLMs with recommendation tasks do not fully leverage their sequential information processing capabilities, leading to suboptimal performance. In this paper, we propose a novel system called compressed vocabulary expansion (CoVE). In CoVE, each item is assigned a unique ID within the expanded vocabulary. Our framework effectively capitalizes on sequence understanding abilities of LLMs, significantly enhancing their performance on recommendation tasks. Additionally, we compress the embedding layer, making CoVE practical for large-scale industrial applications. The effectiveness and performance of CoVE are demonstrated through comprehensive experiments on multiple recommendation datasets and comparisons with prior works. Our code can be found at https://github.com/HaochenZhang717/CoVE-official-Repo.

Shabnam Daghaghi, Nicholas Meisburger, Mengnan Zhao et al.

Deep learning implementations on CPUs (Central Processing Units) are gaining more traction. Enhanced AI capabilities on commodity x86 architectures are commercially appealing due to the reuse of existing hardware and virtualization ease. A notable work in this direction is the SLIDE system. SLIDE is a C++ implementation of a sparse hash table based back-propagation, which was shown to be significantly faster than GPUs in training hundreds of million parameter neural models. In this paper, we argue that SLIDE's current implementation is sub-optimal and does not exploit several opportunities available in modern CPUs. In particular, we show how SLIDE's computations allow for a unique possibility of vectorization via AVX (Advanced Vector Extensions)-512. Furthermore, we highlight opportunities for different kinds of memory optimization and quantizations. Combining all of them, we obtain up to 7x speedup in the computations on the same hardware. Our experiments are focused on large (hundreds of millions of parameters) recommendation and NLP models. Our work highlights several novel perspectives and opportunities for implementing randomized algorithms for deep learning on modern CPUs. We provide the code and benchmark scripts at https://github.com/RUSH-LAB/SLIDE

Nicholas Meisburger, Anshumali Shrivastava

We present SLASH (Sketched LocAlity Sensitive Hashing), an MPI (Message Passing Interface) based distributed system for approximate similarity search over terabyte scale datasets. SLASH provides a multi-node implementation of the popular LSH (locality sensitive hashing) algorithm, which is generally implemented on a single machine. We show how we can append the LSH algorithm with heavy hitters sketches to provably solve the (high) similarity search problem without a single distance computation. Overall, we mathematically show that, under realistic data assumptions, we can identify the near-neighbor of a given query $q$ in sub-linear ($ \ll O(n)$) number of simple sketch aggregation operations only. To make such a system practical, we offer a novel design and sketching solution to reduce the inter-machine communication overheads exponentially. In a direct comparison on comparable hardware, SLASH is more than 10000x faster than the popular LSH package in PySpark. PySpark is a widely-adopted distributed implementation of the LSH algorithm for large datasets and is deployed in commercial platforms. In the end, we show how our system scale to Tera-scale Criteo dataset with more than 4 billion samples. SLASH can index this 2.3 terabyte data over 20 nodes in under an hour, with query times in a fraction of milliseconds. To the best of our knowledge, there is no open-source system that can index and perform a similarity search on Criteo with a commodity cluster.

Zeyu Yang, Qi Ma, Jason Chen et al.

Retrieval-augmented agents are increasingly the interface to large organizational knowledge bases, yet most still treat retrieval as a black box: they issue exploratory queries, inspect returned snippets, and iteratively reformulate until useful evidence emerges. This approach resembles how a newcomer searches an unfamiliar database rather than how an expert navigates it with strong priors about terminology and likely evidence, and results in unnecessary retrieval rounds, increased latency, and poor recall. We introduce \textit{SuperIntelligent Retrieval Agent} (SIRA), which defines \emph{superintelligence} in retrieval as the ability to compress multi-round exploratory search into a single corpus-discriminative retrieval action. SIRA does not merely ask what terms are relevant to the query; it asks which terms are likely to separate the desired evidence from corpus-level confusers. On the corpus side, an LLM enriches each document offline with missing search vocabulary; on the query side, it predicts evidence vocabulary omitted by the query; and document-frequency statistics as a tool call to filter proposed terms that are absent, overly common, or unlikely to create retrieval margin. The final retrieval step is a single weighted BM25 call combining the original query with the validated expansion. Across ten BEIR benchmarks and downstream question-answering tasks, SIRA achieves the significantly superior performance outperforming dense retrievers and state-of-the-art multi-round agentic baselines, demonstrating that one well-formed lexical query, guided by LLM cognition and lightweight corpus statistics, can exceed substantially more expensive multi-round search while remaining interpretable, training-free, and efficient.

Tianyi Zhang, Jonah Yi, Zhaozhuo Xu et al.

Efficient deployment of Large Language Models (LLMs) requires batching multiple requests together to improve throughput. As the batch size, context length, or model size increases, the size of the key and value (KV) cache can quickly become the main contributor to GPU memory usage and the bottleneck of inference latency. Quantization has emerged as an effective technique for KV cache compression, but existing methods still fail at very low bit widths. We observe that distinct channels of a key/value activation embedding are highly inter-dependent, and the joint entropy of multiple channels grows at a slower rate than the sum of their marginal entropies. Based on this insight, we propose Coupled Quantization (CQ), which couples multiple key/value channels together to exploit their inter-dependency and encode the activations in a more information-efficient manner. Extensive experiments reveal that CQ outperforms or is competitive with existing baselines in preserving model quality. Furthermore, we demonstrate that CQ can preserve model quality with KV cache quantized down to 1-bit.

Zichang Liu, Qingyun Liu, Yuening Li et al.

Recent advancements in foundation models have yielded impressive performance across a wide range of tasks. Meanwhile, for specific applications, practitioners have been developing specialized application models. To enjoy the benefits of both kinds of models, one natural path is to transfer the knowledge in foundation models into specialized application models, which are generally more efficient for serving. Techniques from knowledge distillation may be applied here, where the application model learns to mimic the foundation model. However, specialized application models and foundation models have substantial gaps in capacity, employing distinct architectures, using different input features from different modalities, and being optimized on different distributions. These differences in model characteristics lead to significant challenges for distillation methods. In this work, we propose creating a teaching committee comprising both foundation model teachers and complementary teachers. Complementary teachers possess model characteristics akin to the student's, aiming to bridge the gap between the foundation model and specialized application models for a smoother knowledge transfer. Further, to accommodate the dissimilarity among the teachers in the committee, we introduce DiverseDistill, which allows the student to understand the expertise of each teacher and extract task knowledge. Our evaluations demonstrate that adding complementary teachers enhances student performance. Finally, DiverseDistill consistently outperforms baseline distillation methods, regardless of the teacher choices, resulting in significantly improved student performance.

Haochen Zhang, Junze Yin, Guanchu Wang et al.

Low-rank optimization has emerged as a promising approach to enabling memory-efficient training of large language models (LLMs). Existing low-rank optimization methods typically project gradients onto a low-rank subspace, reducing the memory cost of storing optimizer states. A key challenge in these methods is selecting suitable subspaces to ensure an effective optimization trajectory. Most existing approaches select the dominant subspace to preserve gradient information, as this intuitively provides the best approximation. However, we find that in practice, the dominant subspace stops changing during pretraining, thereby constraining weight updates to similar subspaces. In this paper, we propose importance sampling for low-rank optimization in LLM pretraining with a provable convergence guarantee, which the dominant subspace approach does not have. Empirically, we demonstrate that our method significantly outperforms previous methods in LLM pretraining tasks.

Tianyi Zhang, Haoteng Yin, Rongzhe Wei et al.

Graph neural networks (GNNs) have shown great potential in learning on graphs, but they are known to perform sub-optimally on link prediction tasks. Existing GNNs are primarily designed to learn node-wise representations and usually fail to capture pairwise relations between target nodes, which proves to be crucial for link prediction. Recent works resort to learning more expressive edge-wise representations by enhancing vanilla GNNs with structural features such as labeling tricks and link prediction heuristics, but they suffer from high computational overhead and limited scalability. To tackle this issue, we propose to learn structural link representations by augmenting the message-passing framework of GNNs with Bloom signatures. Bloom signatures are hashing-based compact encodings of node neighborhoods, which can be efficiently merged to recover various types of edge-wise structural features. We further show that any type of neighborhood overlap-based heuristic can be estimated by a neural network that takes Bloom signatures as input. GNNs with Bloom signatures are provably more expressive than vanilla GNNs and also more scalable than existing edge-wise models. Experimental results on five standard link prediction benchmarks show that our proposed model achieves comparable or better performance than existing edge-wise GNN models while being 3-200 $\times$ faster and more memory-efficient for online inference.

Xiaoqiang Lin, Aritra Ghosh, Bryan Kian Hsiang Low et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in leveraging extensive external knowledge to enhance responses in multi-turn and agentic applications, such as retrieval-augmented generation (RAG). However, processing long-context inputs introduces significant system latency and demands substantial memory for the key-value cache, resulting in reduced throughput and a fundamental trade-off between knowledge enrichment and system efficiency. While minimizing latency for long-context inputs is a primary objective for LLMs, we contend that RAG require specialized consideration. In RAG, much of the LLM context consists of concatenated passages from retrieval, with only a small subset directly relevant to the query. These passages often exhibit low semantic similarity due to diversity or deduplication during re-ranking, leading to block-diagonal attention patterns that differ from those in standard LLM generation tasks. Based on this observation, we argue that most computations over the RAG context during decoding are unnecessary and can be eliminated with minimal impact on performance. To this end, we propose REFRAG, an efficient decoding framework that compresses, senses, and expands to improve latency in RAG applications. By exploiting the sparsity structure, we demonstrate a 30.85 the time-to-first-token acceleration (3.75 improvement to previous work) without loss in perplexity. In addition, our optimization framework for large context enables REFRAG to extend the context size of LLMs by 16. We provide rigorous validation of REFRAG across diverse long-context tasks, including RAG, multi-turn conversations, and long document summarization, spanning a wide range of datasets. Experimental results confirm that REFRAG delivers substantial speedup with no loss in accuracy compared to LLaMA models and other state-of-the-art baselines across various context sizes.

Sahil Joshi, Agniva Chowdhury, Amar Kanakamedala et al.

Softmax Attention has a quadratic time complexity, which becomes prohibitive to run at long contexts, even with highly optimized GPU kernels. For example, FlashAttention (an exact, GPU-optimized implementation of Softmax Attention) cannot complete a single forward-backward pass of a multi-head attention layer once the context exceeds ~4 million tokens on an NVIDIA GH200 (96 GB). We introduce RACE Attention, a kernel-inspired alternative to Softmax Attention that is linear in sequence length and embedding dimension. RACE Attention replaces the exponential kernel with a sharpened angular (cosine) similarity, and approximates attention outputs via randomized projections and soft Locality-Sensitive Hashing (LSH). Across language modeling, masked language modeling, and text classification, RACE Attention matches the accuracy of strong baselines while reducing runtime and memory. In a controlled scale test, it processes up to 12 million tokens during a single forward-backward pass on an NVIDIA GH200 GPU and 75 million tokens on an Intel Xeon Gold 5220R CPU, well beyond the practical limits of the current state-of-the-art attention implementations. RACE Attention thus offers a practical, theoretically grounded mechanism for outrageously long context windows on today's hardware. We hope that it gets adopted in practice.

Zeyu Yang, Tianyi Zhang, Jianwen Xie et al.

The scaling of Generative AI (GenAI) models into the hundreds of billions of parameters makes low-precision computation indispensable for efficient deployment. We argue that the fundamental solution lies in developing low-precision floating-point formats, which inherently provide numerical stability, memory savings, and hardware efficiency without dequantization overhead. In this paper, we present a theoretical and empirical study of an exponent concentration phenomenon in GenAI weights: exponents consistently exhibit low entropy across architectures and modalities. We show that this arises naturally from $α$-stable distributions induced by stochastic gradient descent, and we prove tight bounds on the entropy of exponents. Our analysis establishes a theoretical compression limit near FP4.67, which motivates the design of a practical FP8 format. Building on these insights, we propose Exponent-Concentrated FP8 (ECF8), a lossless compression framework with entropy-aware encoding and GPU-optimized decoding. Experiments on LLMs and DiTs up to 671B parameters demonstrate up to 26.9% memory savings and 177.1% throughput acceleration, with perfectly lossless computations, i.e., no deviation in model outputs. Our results establish exponent concentration as a statistical law of trained models and open a principled path for lossless low-precision floating-point design in the FP8 era.

Bingcong Li, Shuai Zheng, Parameswaran Raman et al.

On-device memory concerns in distributed deep learning have become severe due to (i) the growth of model size in multi-GPU training, and (ii) the wide adoption of deep neural networks for federated learning on IoT devices which have limited storage. In such settings, communication efficient optimization methods are attractive alternatives, however they still struggle with memory issues. To tackle these challenges, we propose an communication efficient method called contractive error feedback (ConEF). As opposed to SGD with error-feedback (EFSGD) that inefficiently manages memory, ConEF obtains the sweet spot of convergence and memory usage, and achieves communication efficiency by leveraging biased and all-reducable gradient compression. We empirically validate ConEF on various learning tasks that include image classification, language modeling, and machine translation and observe that ConEF saves 80\% - 90\% of the extra memory in EFSGD with almost no loss on test performance, while also achieving 1.3x - 5x speedup of SGD. Through our work, we also demonstrate the feasibility and convergence of ConEF to clear up the theoretical barrier of integrating ConEF to popular memory efficient frameworks such as ZeRO-3.

Zichang Liu, Aditya Desai, Fangshuo Liao et al.

Large language models(LLMs) have sparked a new wave of exciting AI applications. Hosting these models at scale requires significant memory resources. One crucial memory bottleneck for the deployment stems from the context window. It is commonly recognized that model weights are memory hungry; however, the size of key-value embedding stored during the generation process (KV cache) can easily surpass the model size. The enormous size of the KV cache puts constraints on the inference batch size, which is crucial for high throughput inference workload. Inspired by an interesting observation of the attention scores, we hypothesize the persistence of importance: only pivotal tokens, which had a substantial influence at one step, will significantly influence future generations. Based on our empirical verification and theoretical analysis around this hypothesis, we propose Scissorhands, a system that maintains the memory usage of the KV cache at a fixed budget without finetuning the model. In essence, Scissorhands manages the KV cache by storing the pivotal tokens with a higher probability. We validate that Scissorhands reduces the inference memory usage of the KV cache by up to 5X without compromising model quality. We further demonstrate that Scissorhands can be combined with 4-bit quantization, traditionally used to compress model weights, to achieve up to 20X compression.

Zhaozhuo Xu, Zirui Liu, Beidi Chen et al.

While the numerous parameters in Large Language Models (LLMs) contribute to their superior performance, this massive scale makes them inefficient and memory-hungry. Thus, they are hard to deploy on commodity hardware, such as one single GPU. Given the memory and power constraints of such devices, model compression methods are widely employed to reduce both the model size and inference latency, which essentially trades off model quality in return for improved efficiency. Thus, optimizing this accuracy-efficiency trade-off is crucial for the LLM deployment on commodity hardware. In this paper, we introduce a new perspective to optimize this trade-off by prompting compressed models. Specifically, we first observe that for certain questions, the generation quality of a compressed LLM can be significantly improved by adding carefully designed hard prompts, though this isn't the case for all questions. Based on this observation, we propose a soft prompt learning method where we expose the compressed model to the prompt learning process, aiming to enhance the performance of prompts. Our experimental analysis suggests our soft prompt strategy greatly improves the performance of the 8x compressed LLaMA-7B model (with a joint 4-bit quantization and 50% weight pruning compression), allowing them to match their uncompressed counterparts on popular benchmarks. Also, we demonstrate that these learned prompts can be transferred across various datasets, tasks, and compression levels. Hence with this transferability, we can stitch the soft prompt to a newly compressed model to improve the test-time accuracy in an ``in-situ'' way.

Minghao Yan, Nicholas Meisburger, Tharun Medini et al.

More than 70% of cloud computing is paid for but sits idle. A large fraction of these idle compute are cheap CPUs with few cores that are not utilized during the less busy hours. This paper aims to enable those CPU cycles to train heavyweight AI models. Our goal is against mainstream frameworks, which focus on leveraging expensive specialized ultra-high bandwidth interconnect to address the communication bottleneck in distributed neural network training. This paper presents a distributed model-parallel training framework that enables training large neural networks on small CPU clusters with low Internet bandwidth. We build upon the adaptive sparse training framework introduced by the SLIDE algorithm. By carefully deploying sparsity over distributed nodes, we demonstrate several orders of magnitude faster model parallel training than Horovod, the main engine behind most commercial software. We show that with reduced communication, due to sparsity, we can train close to a billion parameter model on simple 4-16 core CPU nodes connected by basic low bandwidth interconnect. Moreover, the training time is at par with some of the best hardware accelerators.

Anshumali Shrivastava, Zhao Song, Zhaozhuo Xu

Conditional gradient methods (CGM) are widely used in modern machine learning. CGM's overall running time usually consists of two parts: the number of iterations and the cost of each iteration. Most efforts focus on reducing the number of iterations as a means to reduce the overall running time. In this work, we focus on improving the per iteration cost of CGM. The bottleneck step in most CGM is maximum inner product search (MaxIP), which requires a linear scan over the parameters. In practice, approximate MaxIP data-structures are found to be helpful heuristics. However, theoretically, nothing is known about the combination of approximate MaxIP data-structures and CGM. In this work, we answer this question positively by providing a formal framework to combine the locality sensitive hashing type approximate MaxIP data-structures with CGM algorithms. As a result, we show the first algorithm, where the cost per iteration is sublinear in the number of parameters, for many fundamental optimization algorithms, e.g., Frank-Wolfe, Herding algorithm, and policy gradient.

Zhenwei Dai, Chen Dun, Yuxin Tang et al.

Federated learning enables many local devices to train a deep learning model jointly without sharing the local data. Currently, most of federated training schemes learns a global model by averaging the parameters of local models. However, most of these training schemes suffer from high communication cost resulted from transmitting full local model parameters. Moreover, directly averaging model parameters leads to a significant performance degradation, due to the class-imbalanced non-iid data on different devices. Especially for the real life federated learning tasks involving extreme classification, (1) communication becomes the main bottleneck since the model size increases proportionally to the number of output classes; (2) extreme classification (such as user recommendation) normally have extremely imbalanced classes and heterogeneous data on different devices. To overcome this problem, we propose federated multiple label hashing (FedMLH), which leverages label hashing to simultaneously reduce the model size (up to 3.40X decrease) with communication cost (up to 18.75X decrease) and achieves significant better accuracy (up to 35.5%} relative accuracy improvement) and faster convergence rate (up to 5.5X increase) for free on the federated extreme classification tasks compared to federated average algorithm.

Aditya Desai, Li Chou, Anshumali Shrivastava

Deep learning for recommendation data is one of the most pervasive and challenging AI workload in recent times. State-of-the-art recommendation models are one of the largest models matching the likes of GPT-3 and Switch Transformer. Challenges in deep learning recommendation models (DLRM) stem from learning dense embeddings for each of the categorical tokens. These embedding tables in industrial scale models can be as large as hundreds of terabytes. Such large models lead to a plethora of engineering challenges, not to mention prohibitive communication overheads, and slower training and inference times. Of these, slower inference time directly impacts user experience. Model compression for DLRM is gaining traction and the community has recently shown impressive compression results. In this paper, we present Random Offset Block Embedding Array (ROBE) as a low memory alternative to embedding tables which provide orders of magnitude reduction in memory usage while maintaining accuracy and boosting execution speed. ROBE is a simple fundamental approach in improving both cache performance and the variance of randomized hashing, which could be of independent interest in itself. We demonstrate that we can successfully train DLRM models with same accuracy while using $1000 \times$ less memory. A $1000\times$ compressed model directly results in faster inference without any engineering effort. In particular, we show that we can train DLRM model using ROBE array of size 100MB on a single GPU to achieve AUC of 0.8025 or higher as required by official MLPerf CriteoTB benchmark DLRM model of 100GB while achieving about $3.1\times$ (209\%) improvement in inference throughput.

Zichang Liu, Benjamin Coleman, Anshumali Shrivastava

Large machine learning models achieve unprecedented performance on various tasks and have evolved as the go-to technique. However, deploying these compute and memory hungry models on resource constraint environments poses new challenges. In this work, we propose mathematically provable Representer Sketch, a concise set of count arrays that can approximate the inference procedure with simple hashing computations and aggregations. Representer Sketch builds upon the popular Representer Theorem from kernel literature, hence the name, providing a generic fundamental alternative to the problem of efficient inference that goes beyond the popular approach such as quantization, iterative pruning and knowledge distillation. A neural network function is transformed to its weighted kernel density representation, which can be very efficiently estimated with our sketching algorithm. Empirically, we show that Representer Sketch achieves up to 114x reduction in storage requirement and 59x reduction in computation complexity without any drop in accuracy.

Zhaozhuo Xu, Minghao Yan, Junyan Zhang et al.

Transformer models have demonstrated superior performance in natural language processing. The dot product self-attention in Transformer allows us to model interactions between words. However, this modeling comes with significant computational overhead. In this work, we revisit the memory-compute trade-off associated with Transformer, particularly multi-head attention, and show a memory-heavy but significantly more compute-efficient alternative to Transformer. Our proposal, denoted as PairConnect, a multilayer perceptron (MLP), models the pairwise interaction between words by explicit pairwise word embeddings. As a result, PairConnect substitutes self dot product with a simple embedding lookup. We show mathematically that despite being an MLP, our compute-efficient PairConnect is strictly more expressive than Transformer. Our experiment on language modeling tasks suggests that PairConnect could achieve comparable results with Transformer while reducing the computational cost associated with inference significantly.

Anshumali Shrivastava, Zhao Song, Zhaozhuo Xu

We present the first provable Least-Squares Value Iteration (LSVI) algorithms that have runtime complexity sublinear in the number of actions. We formulate the value function estimation procedure in value iteration as an approximate maximum inner product search problem and propose a locality sensitive hashing (LSH) [Indyk and Motwani STOC'98, Andoni and Razenshteyn STOC'15, Andoni, Laarhoven, Razenshteyn and Waingarten SODA'17] type data structure to solve this problem with sublinear time complexity. Moreover, we build the connections between the theory of approximate maximum inner product search and the regret analysis of reinforcement learning. We prove that, with our choice of approximation factor, our Sublinear LSVI algorithms maintain the same regret as the original LSVI algorithms while reducing the runtime complexity to sublinear in the number of actions. To the best of our knowledge, this is the first work that combines LSH with reinforcement learning resulting in provable improvements. We hope that our novel way of combining data-structures and iterative algorithm will open the door for further study into cost reduction in optimization.

Gaurav Gupta, Tharun Medini, Anshumali Shrivastava et al.

Neural models have transformed the fundamental information retrieval problem of mapping a query to a giant set of items. However, the need for efficient and low latency inference forces the community to reconsider efficient approximate near-neighbor search in the item space. To this end, learning to index is gaining much interest in recent times. Methods have to trade between obtaining high accuracy while maintaining load balance and scalability in distributed settings. We propose a novel approach called IRLI (pronounced `early'), which iteratively partitions the items by learning the relevant buckets directly from the query-item relevance data. Furthermore, IRLI employs a superior power-of-$k$-choices based load balancing strategy. We mathematically show that IRLI retrieves the correct item with high probability under very natural assumptions and provides superior load balancing. IRLI surpasses the best baseline's precision on multi-label classification while being $5x$ faster on inference. For near-neighbor search tasks, the same method outperforms the state-of-the-art Learned Hashing approach NeuralLSH by requiring only ~ {1/6}^th of the candidates for the same recall. IRLI is both data and model parallel, making it ideal for distributed GPU implementation. We demonstrate this advantage by indexing 100 million dense vectors and surpassing the popular FAISS library by >10% on recall.

Zhaozhuo Xu, Aditya Desai, Menal Gupta et al.

Traditional seismic processing workflows (SPW) are expensive, requiring over a year of human and computational effort. Deep learning (DL) based data-driven seismic workflows (DSPW) hold the potential to reduce these timelines to a few minutes. Raw seismic data (terabytes) and required subsurface prediction (gigabytes) are enormous. This large-scale, spatially irregular time-series data poses seismic data ingestion (SDI) as an unconventional yet fundamental problem in DSPW. Current DL research is limited to small-scale simplified synthetic datasets as they treat seismic data like images and process them with convolution networks. Real seismic data, however, is at least 5D. Applying 5D convolutions to this scale is computationally prohibitive. Moreover, raw seismic data is highly unstructured and hence inherently non-image like. We propose a fundamental shift to move away from convolutions and introduce SESDI: Set Embedding based SDI approach. SESDI first breaks down the mammoth task of large-scale prediction into an efficient compact auxiliary task. SESDI gracefully incorporates irregularities in data with its novel model architecture. We believe SESDI is the first successful demonstration of end-to-end learning on real seismic data. SESDI achieves SSIM of over 0.8 on velocity inversion task on real proprietary data from the Gulf of Mexico and outperforms the state-of-the-art U-Net model on synthetic datasets.

Aditya Desai, Yanzhou Pan, Kuangyuan Sun et al.

Deep learning-based models are utilized to achieve state-of-the-art performance for recommendation systems. A key challenge for these models is to work with millions of categorical classes or tokens. The standard approach is to learn end-to-end, dense latent representations or embeddings for each token. The resulting embeddings require large amounts of memory that blow up with the number of tokens. Training and inference with these models create storage, and memory bandwidth bottlenecks leading to significant computing and energy consumption when deployed in practice. To this end, we present the problem of \textit{Memory Allocation} under budget for embeddings and propose a novel formulation of memory shared embedding, where memory is shared in proportion to the overlap in semantic information. Our formulation admits a practical and efficient randomized solution with Locality sensitive hashing based Memory Allocation (LMA). We demonstrate a significant reduction in the memory footprint while maintaining performance. In particular, our LMA embeddings achieve the same performance compared to standard embeddings with a 16$\times$ reduction in memory footprint. Moreover, LMA achieves an average improvement of over 0.003 AUC across different memory regimes than standard DLRM models on Criteo and Avazu datasets

Aditya Desai, Benjamin Coleman, Anshumali Shrivastava

We introduce Density sketches (DS): a succinct online summary of the data distribution. DS can accurately estimate point wise probability density. Interestingly, DS also provides a capability to sample unseen novel data from the underlying data distribution. Thus, analogous to popular generative models, DS allows us to succinctly replace the real-data in almost all machine learning pipelines with synthetic examples drawn from the same distribution as the original data. However, unlike generative models, which do not have any statistical guarantees, DS leads to theoretically sound asymptotically converging consistent estimators of the underlying density function. Density sketches also have many appealing properties making them ideal for large-scale distributed applications. DS construction is an online algorithm. The sketches are additive, i.e., the sum of two sketches is the sketch of the combined data. These properties allow data to be collected from distributed sources, compressed into a density sketch, efficiently transmitted in the sketch form to a central server, merged, and re-sampled into a synthetic database for modeling applications. Thus, density sketches can potentially revolutionize how we store, communicate, and distribute data.

Shabnam Daghaghi, Tharun Medini, Nicholas Meisburger et al.

Softmax classifiers with a very large number of classes naturally occur in many applications such as natural language processing and information retrieval. The calculation of full softmax is costly from the computational and energy perspective. There have been various sampling approaches to overcome this challenge, popularly known as negative sampling (NS). Ideally, NS should sample negative classes from a distribution that is dependent on the input data, the current parameters, and the correct positive class. Unfortunately, due to the dynamically updated parameters and data samples, there is no sampling scheme that is provably adaptive and samples the negative classes efficiently. Therefore, alternative heuristics like random sampling, static frequency-based sampling, or learning-based biased sampling, which primarily trade either the sampling cost or the adaptivity of samples per iteration are adopted. In this paper, we show two classes of distributions where the sampling scheme is truly adaptive and provably generates negative samples in near-constant time. Our implementation in C++ on CPU is significantly superior, both in terms of wall-clock time and accuracy, compared to the most optimized TensorFlow implementations of other popular negative sampling approaches on powerful NVIDIA V100 GPU.

Zichang Liu, Li Chou, Anshumali Shrivastava

On-device localization and tracking are increasingly crucial for various applications. Along with a rapidly growing amount of location data, machine learning (ML) techniques are becoming widely adopted. A key reason is that ML inference is significantly more energy-efficient than GPS query at comparable accuracy, and GPS signals can become extremely unreliable for specific scenarios. To this end, several techniques such as deep neural networks have been proposed. However, during training, almost none of them incorporate the known structural information such as floor plan, which can be especially useful in indoor or other structured environments. In this paper, we argue that the state-of-the-art-systems are significantly worse in terms of accuracy because they are incapable of utilizing these essential structural information. The problem is incredibly hard because the structural properties are not explicitly available, making most structural learning approaches inapplicable. Given that both input and output space potentially contain rich structures, we study our method through the intuitions from manifold-projection. Whereas existing manifold based learning methods actively utilized neighborhood information, such as Euclidean distances, our approach performs Neighbor Oblivious Learning (NObLe). We demonstrate our approach's effectiveness on two orthogonal applications, including WiFi-based fingerprint localization and inertial measurement unit(IMU) based device tracking, and show that it gives significant improvement over state-of-art prediction accuracy.

Constantinos Chamzas, Zachary Kingston, Carlos Quintero-Peña et al.

Earlier work has shown that reusing experience from prior motion planning problems can improve the efficiency of similar, future motion planning queries. However, for robots with many degrees-of-freedom, these methods exhibit poor generalization across different environments and often require large datasets that are impractical to gather. We present SPARK and FLAME , two experience-based frameworks for sampling-based planning applicable to complex manipulators in 3 D environments. Both combine samplers associated with features from a workspace decomposition into a global biased sampling distribution. SPARK decomposes the environment based on exact geometry while FLAME is more general, and uses an octree-based decomposition obtained from sensor data. We demonstrate the effectiveness of SPARK and FLAME on a Fetch robot tasked with challenging pick-and-place manipulation problems. Our approaches can be trained incrementally and significantly improve performance with only a handful of examples, generalizing better over diverse tasks and environments as compared to prior approaches.

Tharun Medini, Beidi Chen, Anshumali Shrivastava

Dense embedding models are commonly deployed in commercial search engines, wherein all the document vectors are pre-computed, and near-neighbor search (NNS) is performed with the query vector to find relevant documents. However, the bottleneck of indexing a large number of dense vectors and performing an NNS hurts the query time and accuracy of these models. In this paper, we argue that high-dimensional and ultra-sparse embedding is a significantly superior alternative to dense low-dimensional embedding for both query efficiency and accuracy. Extreme sparsity eliminates the need for NNS by replacing them with simple lookups, while its high dimensionality ensures that the embeddings are informative even when sparse. However, learning extremely high dimensional embeddings leads to blow up in the model size. To make the training feasible, we propose a partitioning algorithm that learns such high dimensional embeddings across multiple GPUs without any communication. This is facilitated by our novel asymmetric mixture of Sparse, Orthogonal, Learned and Random (SOLAR) Embeddings. The label vectors are random, sparse, and near-orthogonal by design, while the query vectors are learned and sparse. We theoretically prove that our way of one-sided learning is equivalent to learning both query and label embeddings. With these unique properties, we can successfully train 500K dimensional SOLAR embeddings for the tasks of searching through 1.6M books and multi-label classification on the three largest public datasets. We achieve superior precision and recall compared to the respective state-of-the-art baselines for each of the tasks with up to 10 times faster speed.

Louis Abraham, Gary Becigneul, Benjamin Coleman et al.

We study the problem usually referred to as group testing in the context of COVID-19. Given n samples collected from patients, how should we select and test mixtures of samples to maximize information and minimize the number of tests? Group testing is a well-studied problem with several appealing solutions, but recent biological studies impose practical constraints for COVID-19 that are incompatible with traditional methods. Furthermore, existing methods use unnecessarily restrictive solutions, which were devised for settings with more memory and compute constraints than the problem at hand. This results in poor utility. In the new setting, we obtain strong solutions for small values of n using evolutionary strategies. We then develop a new method combining Bloom filters with belief propagation to scale to larger values of n (more than 100) with good empirical results. We also present a more accurate decoding algorithm that is tailored for specific COVID-19 settings. This work demonstrates the practical gap between dedicated algorithms and well-known generic solutions. Our efforts results in a new and practical multiplex method yielding strong empirical performance without mixing more than a chosen number of patients into the same probe. Finally, we briefly discuss adaptive methods, casting them into the framework of adaptive sub-modularity.

Zichang Liu, Zhaozhuo Xu, Alan Ji et al.

Efficient inference for wide output layers (WOLs) is an essential yet challenging task in large scale machine learning. Most approaches reduce this problem to approximate maximum inner product search (MIPS), which relies heavily on the observation that for a given model, ground truth labels correspond to logits of highest value during full model inference. However, such an assumption is restrictive in practice. In this paper, we argue that approximate MIPS subroutines, despite having sub-linear computation time, are sub-optimal because they are tailored for retrieving large inner products with high recall instead of retrieving the correct labels. With WOL, the labels often have moderate inner products, which makes approximate MIPS more challenging. We propose an alternative problem formulation, called Label Superior Sampling (LSS), where the objective is to tailor the system to ensure retrieval of the correct label. Accordingly, we propose a novel learned hash approach, which is significantly more efficient and sufficient for high inference accuracy than MIPS baselines. Our extensive evaluation indicates that LSS can match or even outperform full inference accuracy with around 5x speed up and 87% energy reduction.

Benjamin Coleman, Gaurav Gupta, John Chen et al.

Empirical risk minimization is perhaps the most influential idea in statistical learning, with applications to nearly all scientific and technical domains in the form of regression and classification models. To analyze massive streaming datasets in distributed computing environments, practitioners increasingly prefer to deploy regression models on edge rather than in the cloud. By keeping data on edge devices, we minimize the energy, communication, and data security risk associated with the model. Although it is equally advantageous to train models at the edge, a common assumption is that the model was originally trained in the cloud, since training typically requires substantial computation and memory. To this end, we propose STORM, an online sketch for empirical risk minimization. STORM compresses a data stream into a tiny array of integer counters. This sketch is sufficient to estimate a variety of surrogate losses over the original dataset. We provide rigorous theoretical analysis and show that STORM can estimate a carefully chosen surrogate loss for the least-squares objective. In an exhaustive experimental comparison for linear regression models on real-world datasets, we find that STORM allows accurate regression models to be trained.

Benjamin Coleman, Anshumali Shrivastava

Differential privacy (DP) is a compelling privacy definition that explains the privacy-utility tradeoff via formal, provable guarantees. Inspired by recent progress toward general-purpose data release algorithms, we propose a private sketch, or small summary of the dataset, that supports a multitude of machine learning tasks including regression, classification, density estimation, near-neighbor search, and more. Our sketch consists of randomized contingency tables that are indexed with locality-sensitive hashing and constructed with an efficient one-pass algorithm. We prove competitive error bounds for DP kernel density estimation. Existing methods for DP kernel density estimation scale poorly, often exponentially slower with an increase in dimensions. In contrast, our sketch can quickly run on large, high-dimensional datasets in a single pass. Exhaustive experiments show that our generic sketch delivers a similar privacy-utility tradeoff when compared to existing DP methods at a fraction of the computation cost. We expect that our sketch will enable differential privacy in distributed, large-scale machine learning settings.

Sicong Liu, Junzhao Du, Anshumali Shrivastava et al.

The remarkable success of machine learning has fostered a growing number of cloud-based intelligent services for mobile users. Such a service requires a user to send data, e.g. image, voice and video, to the provider, which presents a serious challenge to user privacy. To address this, prior works either obfuscate the data, e.g. add noise and remove identity information, or send representations extracted from the data, e.g. anonymized features. They struggle to balance between the service utility and data privacy because obfuscated data reduces utility and extracted representation may still reveal sensitive information. This work departs from prior works in methodology: we leverage adversarial learning to a better balance between privacy and utility. We design a \textit{representation encoder} that generates the feature representations to optimize against the privacy disclosure risk of sensitive information (a measure of privacy) by the \textit{privacy adversaries}, and concurrently optimize with the task inference accuracy (a measure of utility) by the \textit{utility discriminator}. The result is the privacy adversarial network (\systemname), a novel deep model with the new training algorithm, that can automatically learn representations from the raw data. Intuitively, PAN adversarially forces the extracted representations to only convey the information required by the target task. Surprisingly, this constitutes an implicit regularization that actually improves task accuracy. As a result, PAN achieves better utility and better privacy at the same time! We report extensive experiments on six popular datasets and demonstrate the superiority of \systemname compared with alternative methods reported in prior work.