Freezing intermediate layers reduces DNN inference latency.
problem High latency in deep learning inference due to computationally expensive models.
method Introduce approximate caching at each intermediate layer to avoid running all layers for many inference requests.
result Reduces the number of effective layers by half for 91.58% of CIFAR-10 requests.
We developed a caching method to speed up concept learning in complex knowledge bases.
problem Complex concept learning requires many instance retrieval calls, increasing runtime.
method Semantics-aware caching that links concepts to instances via crisp set operations.
result Our cache reduces concept retrieval and learning runtime by an order of magnitude.
KVCOMM optimizes multi-agent LLM systems by reusing KV-caches, reducing redundant processing.
problem Substantial overhead from reprocessing overlapping contexts across multi-agent systems.
method KVCOMM reuses KV-caches and adjusts offsets for shared content using a pool of cached examples (anchors).
result Achieves over 70% reuse rate across diverse multi-agent tasks, up to 7.8x speedup.
RLCache uses reinforcement learning to optimize cache management decisions.
problem Optimizing cache hit rate and storage size in computer systems.
method Designing three reinforcement learning agents for cache manager tasks and two advanced architectures.
result Reinforcement learning agents achieve higher cache hit rates and minimize storage space compared to heuristics.
New method speeds up diffusion models without sacrificing quality.
problem Slow inference in diffusion models.
method Adams-Bashforth method for caching and acceleration.
result Achieved nearly 3x speedup with maintained quality.
Training large-scale image recognition models is computationally expensive. This raises the question of whether there might be simple ways to improve the test performance of an already trained model without having to re-train or fine-tune it with new data. Here, we show that, surprisingly, this is indeed possible. The …
A deep learning model predicts future cache accesses with high accuracy.
problem Improving cache management and performance through better prediction of future data accesses.
method Proposed a LSTM-based recurrent neural network model to predict future cache accesses using only a cache trace as input.
result The proposed model achieves high prediction accuracy and outperforms state-of-the-art practical policies.
Parrot learns optimal cache replacement policies using imitation learning.
problem Improving cache hit rates in complex access patterns.
method Imitation learning approach using Belady's oracle policy.
result Parrot increases cache hit rates by 61% on a web search benchmark.
A caching mechanism improves sequence to logical form generation accuracy.
problem Generating logical forms from natural language sequences.
method Proposes a caching mechanism to increase output probability of source input tokens and weigh them based on context.
result Improves sequence/token-level accuracy on sequence to logical form tasks.
Bandana stores deep learning models using NVM with DRAM caching.
problem Reducing memory footprint of deep learning models stored on DRAM.
method Uses NVM as primary storage, DRAM as cache, hypergraph partitioning, and cache simulation.
result Significantly reduces the cost of ownership for storing deep learning models.
Adaptive caching strategy improves SVM training efficiency.
problem Expensive SVM training for large datasets.
method Proposed EFU and HCST caching strategies for kernel value reuse.
result HCST achieves 20% more reduction in training time.
Proposes a method to predict content preferences for mobile users in decentralized caching networks.
problem Determining caching schemes for decentralized caching networks with mobile traffic.
method Formulates content preference learning as a DRMTL problem, integrates mobility prediction, and uses ADMM for optimization.
result Mobility-aware content preference learning provides more accurate predictions and improved hit ratios.
DEAP Cache learns prefetching, eviction, and admission using machine learning.
problem Improving cache performance through better prefetching, eviction, and admission strategies.
method End-to-end pipeline using machine learning, inspired by pretraining on large corpora and online reinforcement learning.
result Optimal policy distribution between two orthogonal eviction strategies based on frequency and recency.
Graph-Structured Cache improves code completion and variable naming tasks.
problem Learning open vocabulary in source code.
method Graph-Structured Cache for handling new words in code.
result Improves code completion and variable naming tasks by over 100%.
New MAB model for online caching costs.
problem Learning costs of cached items online.
method Synchronization bandits, MirrorSync algorithm.
result Adversarial regret of O(T2/3) for MirrorSync. Paper tackles caching in fog radio access networks using RL.
problem Distributed edge caching in fog radio access networks with fluctuating traffic demands.
method Q-learning framework for optimal caching policy, value function approximation.
result Proposed method outperforms traditional methods in simulations.
New algorithm improves cache management with delayed feedback and decaying costs.
problem Improving cache replacement policies with delayed and decaying costs feedback.
method Proposed EXP4-DFDC adaptive reinforcement learning algorithm.
result Expected regret is a vanishing quantity as a function of time.
This paper optimizes caching and model multiplexing for large model inference.
problem Resource consumption and latency challenges in large model deployment.
method Jointly optimizing a caching algorithm (GDSF or LEC) and a model multiplexer for large model inference.
result Achieves optimal rates in offline and online settings with up to 50x improvement over baseline.
A new method for energy-efficient file delivery in small cell networks.
problem Efficient resource management in femto-caching with time-variant statistical properties.
method Formulates a resource allocation problem as a stochastic knapsack problem and a multi-armed bandit problem, developing solutions for each.
result The proposed method maximizes the accumulated utility over the horizon, especially suitable for networks with time-variant statistical properties.
This paper is about metric data structures in high-dimensional or non-Euclidean space that permit cached sufficient statistics accelerations of learning algorithms. It has recently been shown that for less than about 10 dimensions, decorating kd-trees with additional "cached sufficient statistics" such as first and sec…
LATM framework uses LLMs to create and reuse tools for efficient problem-solving.
problem Improving problem-solving capabilities of LLMs with external tools.
method Closed-loop framework with two phases: tool making and tool using. LLMs act as both tool makers and users. Resource-intensive model for tool making, lightweight model for tool using.
result LATM framework achieves performance equivalent to using a powerful LLM for both roles but with significantly reduced costs.
FibQuant improves KV-cache compression for long-context inference.
problem Memory traffic bottleneck in long-context inference due to KV cache growth.
method Introduces FibQuant, a universal vector quantizer that combines Beta-quantile radii, Fibonacci/Roberts-Kronecker directions, and Lloyd-Max refinement.
result FibQuant achieves high compression rates with minimal loss in attention cosine similarity.
Optimizes parallel training of linear models, improving convergence.
problem Improving convergence of parallel training of linear models.
method Data partitioning scheme across threads to improve convergence.
result Achieved up to 42x speedup in convergence compared to state of the art implementations.
New algorithm SAGA++ improves on SAGA for faster convergence in stochastic batch size methods.
problem Optimizing convergence rate in stochastic variance reduction methods with batch size.
method Proposes SAGA++ algorithm with optimal average batch size considering cache/disk IO effects.
result SAGA++ outperforms SAGA and other solvers on real datasets.
We address challenges of active learning under scarce informational resources in non-stationary environments. In real-world settings, data labeled and integrated into a predictive model may become invalid over time. However, the data can become informative again with switches in context and such changes may indicate un…
Research optimizes C++ patterns for HFT, reducing latency and improving profitability.
problem Optimizing latency-critical code for high-frequency trading systems.
method Creation of a Low-Latency Programming Repository, optimisation of trading strategy, implementation of Disruptor pattern.
result Significant performance improvements in speed and profitability.
IVF k-means algorithm improves performance on large sparse data sets.
problem Efficiently clustering large-scale sparse data sets with numerous classes.
method Sparse data representation and inverted-file structure for high-speed and low-memory clustering.
result IVF achieves better performance than other algorithms on real document data sets.
Unified architecture for multi-modal multi-task learning using transformer.
problem Training multiple tasks concurrently with varying modalities.
method Spatio-temporal cache mechanism for multi-modal learning.
result Training multiple tasks together reduces model size by about three times.
DeltaGrad rapidly retrain models with minimal data changes.
problem Rapid retraining of machine learning models with minimal data changes.
method DeltaGrad algorithm based on cached training information.
result DeltaGrad compares favorably to state-of-the-art methods.
Efficient DSP features reduce speech recognition memory usage.
problem Limited memory on DSPs for speech recognition.
method Developed efficient bottleneck features (BNFs) for DSPs.
result Reduced speech recognition memory usage by 10x with minimal accuracy loss.
In this paper, we propose several improvements on the block-coordinate Frank-Wolfe (BCFW) algorithm from Lacoste-Julien et al. (2013) recently used to optimize the structured support vector machine (SSVM) objective in the context of structured prediction, though it has wider applications. The key intuition behind our i…
Accelerates data loading in deep neural network training by 30x.
problem Data loading is a bottleneck in deep neural network training.
method Locality-aware data loading method using software caches.
result More than 30x speedup in data loading.
Paper accelerates K-means clustering for large sparse document data.
problem Efficiently clustering large-scale sparse document data.
method Designs an AFM algorithm leveraging UCs and inverted-index structure.
result Significantly improves clustering speed for large-scale documents.
Pipeline-aware hyperparameter tuning speeds up machine learning pipelines by reusing intermediate computations.
problem High computational burden in hyperparameter tuning of multi-stage pipelines.
method Proposes a hybrid hyperparameter tuning method and a caching problem formulated as an ILP to maximize reuse.
result Pipeline-aware approach offers over an order-of-magnitude speedup over independent evaluations.
Developing efficient and scalable algorithms for Latent Dirichlet Allocation (LDA) is of wide interest for many applications. Previous work has developed an O(1) Metropolis-Hastings sampling method for each token. However, the performance is far from being optimal due to random accesses to the parameter matrices and fr…
Efficiently combines autoregressive and set-based models for joint distributions.
problem Joint distributions over multiple predictions from set-based models.
method Causal autoregressive buffer that caches context and captures dependencies.
result Up to 20x faster joint sampling and density evaluation, up to 7x lower memory usage.
Paper details Hilbert-curve for high-performance data mining.
problem Efficiently mapping multi-dimensional data to one dimension.
method Defines Hilbert-curve using finite automaton and context-free grammar.
result Cache-oblivious algorithms for matrix operations and clustering.
Automated process discovery is a class of process mining methods that allow analysts to extract business process models from event logs. Traditional process discovery methods extract process models from a snapshot of an event log stored in its entirety. In some scenarios, however, events keep coming with a high arrival…
Paper proposes EEIPU, a memoization-aware BO algorithm to reduce hyperparameter tuning costs.
problem High costs in GPU-days for training and fine-tuning language models.
method Memoization-aware Bayesian Optimization (EEIPU) algorithm in tandem with pipeline caching.
result EEIPU produces 103% more hyperparameter candidates and 108% more validation metric improvement.
Efficient index structures for fast approximate nearest neighbor queries are required in many applications such as recommendation systems. In high-dimensional spaces, many conventional methods suffer from excessive usage of memory and slow response times. We propose a method where multiple random projection trees are c…
Small-GAN speeds up GAN training by using coresets.
problem Slowness and high memory usage in GAN training with large batches.
method Draw a large batch of samples from the prior, compress using Coreset-selection, and use cached Inception activations for random projection.
result Significantly reduces training time and memory usage for modern GAN variants.
Study uses three sources to evaluate language models fairly.
problem Bias in offline model evaluation due to confounded model choice.
method Combines observational logs, randomized experiments, and simulators.
result Randomized experiment and simulator together recover causal model values.
We propose a general algorithmic framework for constrained matrix and tensor factorization, which is widely used in signal processing and machine learning. The new framework is a hybrid between alternating optimization (AO) and the alternating direction method of multipliers (ADMM): each matrix factor is updated in tur…
New algorithm reduces costs and latency for large language model inference.
problem Optimizing inference costs and latency for large language models with GPU constraints.
method Formulated as an online scheduling problem with endogenous memory growth, introduced fluid model and WAIT algorithms.
result Reduced costs and latency, especially in near-overloaded and overloaded regimes.
Generative profiling improves real-time task timing for varied resource contexts.
problem Inaccurate task timing analysis for complex hardware architectures.
method Nonparametric, conditional multi-marginal Schrödinger Bridge (MSB) formulation for synthesizing context-dependent timing profiles.
result Maximum likelihood accurate execution profiles for unseen resource contexts.
Online structure learning approaches, such as those stemming from Statistical Relational Learning, enable the discovery of complex relations in noisy data streams. However, these methods assume the existence of fully-labelled training data, which is unrealistic for most real-world applications. We present a novel appro…
A new approach uses deep learning to manage vehicular content efficiently.
problem Managing content replication and caching in vehicular networks efficiently.
method Data-driven, centralized approach using a Convolutional Neural Network (CNN).
result Effective strategies derived to modulate FC operation in space and adapt to mobility changes.
Service-induced congestion in memory-constrained LLM serving
problem Service-induced congestion in memory-constrained large language model (LLM) serving
method Developing a discrete-time dynamical model of memory-constrained LLM inference
result The system converges to a unique worst-case limit cycle that is asymptotically stable outside a Lebesgue-measure-zero exact-capture set, with throughput losses as large as 50%.