SOOTT framework optimizes target tracking with robust and learning-augmented algorithms.
problem Optimizing target tracking in dynamic environments with adversarial perturbations.
method Integrates robust and learning-augmented algorithms for online decision-making.
result CoRT learning-augmented algorithm strictly improves over robust BEST when predictions are accurate.
Proposes a framework for energy-efficient AIGC workload scheduling in cloud data centers.
problem Challenges of scheduling AIGC workloads for energy efficiency and quality control.
method Joint energy management and coordinated AIGC workload scheduling framework with diffusion model-aided reward shaping.
result Effective learning of scheduling policies under sparse environmental feedback.
FPDeep accelerates CNN training on FPGA clusters with high parallelism and energy efficiency.
problem Scaling DNN training to large clusters with high utilization and balanced workload.
method Hybrid model and layer parallelism, fine-grained pipeline, balanced workload partitioning.
result FPDeep achieves high parallelism and utilization, reducing storage demand to on-chip memory.
Chicle tackles elastic machine learning training by avoiding micro-tasks.
problem Elasticity and load balancing in distributed machine learning training.
method Chicle is a new elastic distributed training framework that exploits machine learning algorithms to implement elasticity and load balancing without micro-tasks.
result Chicle achieves performance competitive with state-of-the-art rigid frameworks while enabling elastic execution and dynamic load balancing.
New principle reduces load imbalance in LLM serving systems, saving up to 52% energy.
problem Wasted computational power due to load imbalance in LLM serving systems.
method Developed a universal load-balancing principle for barrier-synchronized systems with non-migratable state.
result Proves worst-case theoretical guarantees for imbalance reduction and energy savings.
The paper predicts workload using process mining and neural networks.
problem Predicting workload in business processes.
method Process mining logs, combined with recurrent neural networks.
result Workload prediction achieved with an MAPE score of 19%.
Automated EEG analysis gauges mental workload in task evaluation.
problem Evaluating mental workload in user tasks, especially complex ones.
method Experimental setup, EEG data analysis, feature extraction, machine learning.
result Machine learning outperforms traditional methods in mental workload estimation.
Developing reliable workload predictive models can affect many aspects of clinical decision making procedure. The primary challenge in healthcare systems is handling the demand uncertainty over the time. This issue becomes more critical for the healthcare facilities that provide service for chronic disease treatment be…
This work introduces a new benchmark to compare neural network training algorithms.
problem Lack of reliable benchmarks to compare training algorithms effectively.
method Developed a new benchmark called AlgoPerf: Training Algorithms benchmark.
result Demonstrated the feasibility of the benchmark and set a provisional state-of-the-art.
Estimates statistical shifts across subjects for EEG-based mental workload assessment.
problem Variability in EEG correlates of mental workload across subjects makes model generalization difficult.
method Proposes a strategy to estimate marginal and conditional shifts between multiple data distributions.
result Estimates of statistical shifts can improve mental workload prediction accuracy.
We present ease.ml, a declarative machine learning service platform we built to support more than ten research groups outside the computer science departments at ETH Zurich for their machine learning needs. With ease.ml, a user defines the high-level schema of a machine learning application and submits the task via a W…
Paper proposes a human-algorithm approach to reduce medical device recall risk and workload.
problem High recall rate and regulatory workload in FDA's 510(k) pathway.
method Developed machine learning models to estimate recall risk and proposed a data-driven clearance policy.
result Conservative evaluation of policy shows a 32.9% improvement in recall rate and 40.5% reduction in workload.
Machine learning reduces workload in healthcare systematic reviews by 70%.
problem Efficiently screening abstracts for systematic reviews in healthcare.
method Training SVM classifiers on labeled abstracts to classify RCTs.
result SVM classifier achieved 90% accuracy and 0.84 F1 score.
ML4CO uses machine learning to improve combinatorial optimization solvers.
problem Solving combinatorial problems in practice often involves related data distributions.
method Replacing heuristic components with machine learning approaches.
result Improved state-of-the-art combinatorial optimization solvers.
We consider the machine teaching problem in a classroom-like setting wherein the teacher has to deliver the same examples to a diverse group of students. Their diversity stems from differences in their initial internal states as well as their learning rates. We prove that a teacher with full knowledge about the learnin…
Efficiently scheduling data processing jobs on distributed compute clusters requires complex algorithms. Current systems, however, use simple generalized heuristics and ignore workload characteristics, since developing and tuning a scheduling policy for each workload is infeasible. In this paper, we show that modern ma…
Paper develops a BERT-based classifier to reduce pathology report annotation workload.
problem Manual annotation of pathology reports is labor-intensive and time-consuming.
method Developed an automatic text classifier using BERT and introduced a human-centric metric to identify low-confidence cases.
result The model reduces manual annotation workload by 80% to 98%.
Paper proposes a method to predict optimal data partitioning based on query execution costs.
problem Finding optimal data partitioning for improved system performance and scalability.
method Formal model abstraction of workload queries, genetic algorithm for optimization, evaluation using PostgreSQL's query optimizer.
result The approach effectively reduces workload execution cost and improves system performance.
Preserving the privacy of individuals by protecting their sensitive attributes is an important consideration during microdata release. However, it is equally important to preserve the quality or utility of the data for at least some targeted workloads. We propose a novel framework for privacy preservation based on the …
Training large machine learning (ML) models with many variables or parameters can take a long time if one employs sequential procedures even with stochastic updates. A natural solution is to turn to distributed computing on a cluster; however, naive, unstructured parallelization of ML algorithms does not usually lead t…
Paper proposes semi-supervised learning for EEG analysis.
problem Reducing workload and delays in analyzing large unlabeled EEG datasets.
method Semi-supervised deep learning algorithm using minimal labeled data.
result Predictions can be made with as little as 5 labeled examples.
Simple policy outperforms complex ones in cloud auto-scaling.
problem Predicting resource scaling for large-scale cloud applications with limited deployment throughput.
method Probabilistic workload forecast for auto-scaling decisions based on risk aversion.
result The proposed policy outperforms sophisticated and simple benchmark policies in real-world and synthetic data.
Modern software systems provide many configuration options which significantly influence their non-functional properties. To understand and predict the effect of configuration options, several sampling and learning strategies have been proposed, albeit often with significant cost to cover the highly dimensional configu…
This paper optimizes how deep learning models are distributed across different devices.
problem Optimizing how large, complex neural networks are split across multiple devices.
method Identified and solved an optimization problem for device placement of DNN operators.
result Automated algorithms that solve the device placement problem for modern pipelined settings.
Classifying human cognitive states from behavioral and physiological signals is a challenging problem with important applications in robotics. The problem is challenging due to the data variability among individual users, and sensor artefacts. In this work, we propose an end-to-end framework for real-time cognitive wor…
AlgoPerf competition evaluates neural network training speed-ups.
problem Improving neural network training speed using better algorithms.
method Compared 18 diverse submissions from 10 teams on multiple workloads.
result Schedule Free AdamW algorithm achieved best results in self-tuning ruleset.
Benefitting from large-scale training datasets and the complex training network, Convolutional Neural Networks (CNNs) are widely applied in various fields with high accuracy. However, the training process of CNNs is very time-consuming, where large amounts of training samples and iterative operations are required to ob…
Optimizes resource allocation for virtualized network functions based on performance profiles.
problem Mapping SLA performance requirements to dynamic virtualized infrastructure resources.
method Profile-based resource allocation using VNF performance datasets and machine learning models.
result A method to predict and recommend optimal resource allocation for network services.
Sequence-to-sequence models predict resource usage for co-scheduled jobs in data centers.
problem Challenges in co-scheduling jobs due to resource interference and inefficiencies.
method Sequence-to-sequence models based on recurrent neural networks for workload interference prediction.
result Models accurately forecast resource usage trends from job profiles, improving scheduling decisions.
AutoPQ automates quantile forecasting for smart grids, reducing workload and environmental impact.
problem Accurate and unbiased uncertainty quantification in probabilistic forecasting for smart grid operations.
method AutoPQ uses a conditional Invertible Neural Network (cINN) to generate quantile forecasts from point forecasts, automating model selection and hyperparameter optimization.
result AutoPQ outperforms state-of-the-art methods while reducing computational effort and environmental impact.
This paper evaluates different data management methods for GBDT systems.
problem The impact of different data management methods on distributed GBDT performance.
method Categorization of data management policies, systematic analysis, and implementation of a novel system Vero.
result Vero, a novel distributed GBDT system, outperforms other systems in various datasets.
Power and thermal management are critical components of High-Performance-Computing (HPC) systems, due to their high power density and large total power consumption. The assessment of thermal dissipation by means of compact models directly from the thermal response of the final device enables more robust and precise the…
FLAME auto-labels mobile data efficiently on diverse processors.
problem Accurately and efficiently labeling mobile data with unknown labels on heterogeneous processors.
method Self-adaptive auto-labeling system Flame that schedules and executes workloads on mobile processors.
result Flame achieves high labeling accuracy and performance on heterogeneous mobile processors.
We introduce a learning-based framework to optimize tensor programs for deep learning workloads. Efficient implementations of tensor operators, such as matrix multiplication and high dimensional convolution, are key enablers of effective deep learning systems. However, existing systems rely on manually optimized librar…
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%.
We improve private training accuracy with learning rate schedules and matrix factorizations.
problem Private training with learning rate schedules and correlated noise.
method General upper and lower bounds for learning rate schedules, memory-efficient constructions, and schedule-aware factorizations.
result Schedule-aware factorizations improve accuracy in private training.
SliceOut speeds up deep learning training without sacrificing accuracy.
problem Frequent model re-training and large model training workloads in deep learning.
method SliceOut uses dropout-inspired scheme to drop contiguous sets of units at random, leveraging GPU memory layout.
result 10-40% speedups and memory reduction with minimal accuracy loss.
Deep learning (DL) training-as-a-service (TaaS) is an important emerging industrial workload. The unique challenge of TaaS is that it must satisfy a wide range of customers who have no experience and resources to tune DL hyper-parameters, and meticulous tuning for each user's dataset is prohibitively expensive. Therefo…
Stochastic gradient descent (SGD) is a popular stochastic optimization method in machine learning. Traditional parallel SGD algorithms, e.g., SimuParallel SGD, often require all nodes to have the same performance or to consume equal quantities of data. However, these requirements are difficult to satisfy when the paral…
A new method for private query release using Johnson-Lindenstrauss projection.
problem Private release of query answers with minimal privacy loss.
method Random projection of query answers to a lower dimension, followed by noise addition.
result Optimal worst-case sample complexity for answering a workload of k queries.
Data compression is a popular technique for improving the efficiency of data processing workloads such as SQL queries and more recently, machine learning (ML) with classical batch gradient methods. But the efficacy of such ideas for mini-batch stochastic gradient descent (MGD), arguably the workhorse algorithm of moder…
Recent hardware developments have dramatically increased the scale of data parallelism available for neural network training. Among the simplest ways to harness next-generation hardware is to increase the batch size in standard mini-batch neural network training algorithms. In this work, we aim to experimentally charac…
Stage-based hyper-parameter optimization reduces GPU-hours and training time.
problem Efficiently executing hyper-parameter optimization for deep learning models.
method Stage-based execution strategy to remove redundant computations.
result Stage-based execution outperforms trial-based method by up to 6.60 times in GPU-hours and 4.13 times in training time.
AutoYara generates effective Yara rules faster than humans.
problem Developing high-quality Yara rules for malware families is labor-intensive.
method Leverages biclustering on large n-grams to automate Yara rule generation.
result AutoYara reduces analyst workload by 44-86% and matches human performance.
A new kernel simplifies deep learning code and boosts performance.
problem Complex and hard-to-maintain DL codebases.
method Introducing a batch-reduce GEMM kernel for deep learning.
result Deep learning algorithms can be implemented with just 3K lines of code.
XLabel tool reduces medical experts' workload by 40% and explains its decisions.
problem Efficiently labeling large electronic health records.
method Visual-interactive tool using Explainable Boosting Machine (EBM) for classification and explanation.
result EBM achieves high accuracy and explainability, even with mislabeled data.
A human-in-the-loop ML framework for precision dosing reduces expert workload and removes bias.
problem High cost of data annotation and lack of appropriate data for ML models.
method Incorporates human experts into the model learning loop to improve interpretability and reduce bias.
result The approach learns interpretable rules from data and potentially lowers expert workload.
Sage platform protects ML models trained on sensitive data from leakage.
problem Protecting sensitive data in machine learning models exposed to untrusted domains.
method Develops block composition for privacy accounting and privacy-adaptive training to manage privacy budget and utility tradeoff.
result Enables continuous training of models on sensitive data streams while maintaining global DP guarantees.