Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Knowledge distillation”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 records
All sourcesNASA NTRSDOE OSTI
DOE OSTI2022

Distilling Knowledge from Ensembles of Cluster-Constrained-Attention Multiple-Instance Learners for Whole Slide Image Classification

The peculiar nature of whole slide imaging (WSI), digitizing conventional glass slides to obtain multiple high resolution images which capture microscopic details of a patient’s histopathological features, has garnered increased interest from the computer vision research community over the last two decades. Given the unique computational space and time complexity inherent to gigapixel-size whole slide image data, researchers have proposed novel machine learning algorithms to aid in the performance of diagnostic tasks in clinical pathology. One effective algorithm represents a Whole slide image as a bag of smaller image patches, which can be represented as low-dimension image patch embeddings. Weakly supervised deep-learning methods, such as cluster-constrained-attention multiple instance learning (CLAM), have shown promising results when combined with image patch embeddings. While traditional ensemble classifiers yield improved task performance, such methods come with a steep cost in model complexity. Through knowledge distillation, it is possible to retain some performance improvements from an ensemble, while minimizing costs to model complexity. In this work, we implement a weakly supervised ensemble using clustering-constrained-attention multiple-instance learners (CLAM), which uses attention and instance-level clustering to identify task salient regions and feature extraction in whole slides. By applying logit-based and attention-based knowledge distillation, we show it is possible to retain some performance improvements resulting from the ensemble at zero cost to model complexity.

Alamudun, Folami↗
DOE OSTI2023

Efficient and Robust Jet Tagging at the LHC with Knowledge Distillation

The challenging environment of real-time data processing systems at the Large Hadron Collider (LHC) strictly limits the computational complexity of algorithms that can be deployed. For deep learning models, this implies that only models with low computational complexity that have weak inductive bias are feasible. To address this issue, we utilize knowledge distillation to leverage both the performance of large models and the reduced computational complexity of small ones. In this paper, we present an implementation of knowledge distillation, demonstrating an overall boost in the student models' performance for the task of classifying jets at the LHC. Furthermore, by using a teacher model with a strong inductive bias of Lorentz symmetry, we show that we can induce the same inductive bias in the student model which leads to better robustness against arbitrary Lorentz boost.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2024

Robustness of topological persistence in knowledge distillation for wearable sensor data

Topological data analysis (TDA) has shown great success in various applications involving wearable sensor data. However, there are difficulties in leveraging topological features in machine learning and wearable sensors because of the large time consumption and computational resources required to extract the features. To address this problem, knowledge distillation (KD) is utilized to generate a small model and accommodate topological features with persistence image (PI) representations from the raw time series data. Deploying topological knowledge in KD enables the student to achieve better performance compared to the one trained solely on raw time series data. However, it is not yet known if there are coherent characteristics for topological features in PI, which can aid in improving the performance during KD. In this paper, we investigate the suitability and challenges of utilizing topological features in KD for wearable sensor data, thereby contributing to the advancement of the field. Our study explores the impact of transferred topological features by comparing the Teacher-to-Student framework with Multiple Teachers-to-Student where teachers utilize both time series data and persistence images obtained by TDA as inputs. Additionally, we conduct a rigorous examination of topological knowledge effects by testing under various corruptions, knowledge types, and learning strategies in the context of human activity recognition tasks. Our analysis of topological features in KD presents the optimal strategy for incorporating these features. This study includes datasets of varying scales, window lengths, and activity classes, providing a comprehensive evaluation. Our results demonstrate that leveraging topological features in KD to enhance performance across databases.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2023

Knowledge Distillation for Anomaly Detection

Unsupervised deep learning techniques are widely used to identify anomalous behaviour. The performance of such methods is a product of the amount of training data and the model size. However, the size is often a limiting factor for the deployment on resource-constrained devices. Here, we present a novel procedure based on knowledge distillation for compressing an unsupervised anomaly detection model into a supervised deployable one and we suggest a set of techniques to improve the detection sensitivity. Compressed models perform comparably to their larger counterparts while significantly reducing the size and memory footprint.

Pol, Adrian Alan↗
DOE OSTI2024

Distilling particle knowledge for fast reconstruction at high-energy physics experiments

Knowledge distillation is a form of model compression that allows artificial neural networks of different sizes to learn from one another. Its main application is the compactification of large deep neural networks to free up computational resources, in particular on edge devices. In this article, we consider proton-proton collisions at the High-Luminosity Large Hadron Collider (HL-LHC) and demonstrate a successful knowledge transfer from an event-level graph neural network (GNN) to a particle-level small deep neural network (DNN). Our algorithm, DistillNet, is a DNN that is trained to learn about the provenance of particles, as provided by the soft labels that are the GNN outputs, to predict whether or not a particle originates from the primary interaction vertex. The results indicate that for this problem, which is one of the main challenges at the HL-LHC, there is minimal loss during the transfer of knowledge to the small student network, while improving significantly the computational resource needs compared to the teacher. This is demonstrated for the distilled student network on a CPU, as well as for a quantized and pruned student network deployed on a field programmable gate array. Our study proves that knowledge transfer between networks of different complexity can be used for fast artificial intelligence (AI) in high-energy physics that improves the expressiveness of observables over non-AI-based reconstruction algorithms. Such an approach can become essential at the HL-LHC experiments, e.g. to comply with the resource budget of their trigger stages.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗
DOE OSTI2023

A survey of techniques for optimizing transformer inference

Recent years have seen a phenomenal rise in the performance and applications of transformer neural networks. The family of transformer networks, including Bidirectional Encoder Representations from Transformer (BERT), Generative Pretrained Transformer (GPT) and Vision Transformer (ViT), have shown their effectiveness across Natural Language Processing (NLP) and Computer Vision (CV) domains. Transformer-based networks such as ChatGPT have impacted the lives of common men. However, the quest for high predictive performance has led to an exponential increase in transformers' memory and compute footprint. Researchers have proposed techniques to optimize transformer inference at all levels of abstraction. Further, this paper presents a comprehensive survey of techniques for optimizing the inference phase of transformer networks. We survey techniques such as knowledge distillation, pruning, quantization, neural architecture search and lightweight network design at the algorithmic level. We further review hardware-level optimization techniques and the design of novel hardware accelerators for transformers. We summarize the quantitative results on the number of parameters/FLOPs and the accuracy of several models/techniques to showcase the tradeoff exercised by them. We also outline future directions in this rapidly evolving field of research. We believe that this survey will educate both novice and seasoned researchers and also spark a plethora of research efforts in this field.

97 MATHEMATICS AND COMPUTING↗
DOE OSTI2025

An ecological framework for microbial metabolites in the ocean ecosystem

The ocean microbe‐metabolite network involves thousands of individual metabolites that encompass a breadth of chemical diversity and biological functions. These microbial metabolites mediate biogeochemical cycles, facilitate ecological relationships, and impact ecosystem health. While analytical advancements have begun to illuminate such roles, a challenge in navigating the deluge of marine metabolomics information is to identify a subset of metabolites that have the greatest ecosystem impact. Here, we present an ecological framework to distill knowledge of fundamental metabolites that underpin marine ecosystems. We borrow terms from macroecology that describe important species, namely “dominant,” “keystone,” and “indicator” species, and apply these designations to metabolites within the ocean microbial metabolome. These selected metabolites may shape marine community structure, function, and health and provide focal points for enhanced study of microbe‐metabolite networks. Applying ecological concepts to marine metabolites provides a path to leverage metabolomics data to better describe and predict marine microbial ecosystems.

microbial metabolites↗
DOE OSTI2024

Efficient distributed continual learning for steering experiments in real-time

Deep learning has emerged as a powerful method for extracting valuable information from large volumes of data. However, when new training data arrives continuously (i.e., is not fully available from the beginning), incremental training suffers from catastrophic forgetting (i.e., new patterns are reinforced at the expense of previously acquired knowledge). Training from scratch each time new training data becomes available would result in extremely long training times and massive data accumulation. Rehearsal-based continual learning has shown promise for addressing the catastrophic forgetting challenge, but research to date has not addressed performance and scalability. To fill this gap, we propose an approach based on a distributed rehearsal buffer that efficiently complements data-parallel training on multiple GPUs to achieve high accuracy, short runtime, and scalability. It leverages a set of buffers (local to each GPU) and uses several asynchronous techniques for updating these local buffers in an embarrassingly parallel fashion, all while handling the communication overheads necessary to augment input minibatches using unbiased, global sampling. We further propose a generalization of rehearsal buffers to support both classification and generative learning tasks, as well as more advanced rehearsal strategies (notably Dark Experience Replay, leveraging knowledge distillation). We illustrate this approach with a real-life HPC streaming application from the domain of ptychographic image reconstruction. Furthermore, we run extensive experiments on up to 128 GPUs of the ThetaGPU supercomputer to compare our approach with baselines representative of training-from-scratch (the upper bound in terms of accuracy) and incremental training (the lower bound). Results show that rehearsal-based continual learning achieves a top-5 validation accuracy close to the upper bound, while simultaneously exhibiting a runtime close to the lower bound.

Asynchronous data management↗
DOE OSTI2023

Let’s Unleash the Network Judgment: A Self-Supervised Approach for Cloud Image Analysis

Accurate cloud type identification and coverage analysis are crucial in understanding the Earth’s radiative budget. Traditional computer vision methods rely on low-level visual features of clouds for estimating cloud coverage or sky conditions. Several handcrafted approaches have been proposed; however, scope for improvement still exists. Newer deep neural networks (DNNs) have demonstrated superior performance for cloud segmentation and categorization. These methods, however, need expert engineering intervention in the preprocessing steps—in the traditional methods—or human assistance in assigning cloud or clear sky labels to a pixel for training DNNs. Such human mediation imposes considerable time and labor costs. We present the application of a new self-supervised learning approach to autonomously extract relevant features from sky images captured by ground-based cameras, for the classification and segmentation of clouds. We evaluate a joint embedding architecture that uses self-knowledge distillation plus regularization. We use two datasets to demonstrate the network’s ability to classify and segment sky images—one with ~85,000 images collected from our ground-based camera and another with 400 labeled images from the WSISEG database. We find that this approach can discriminate full-sky images based on cloud coverage, diurnal variation, and cloud base height. Additionally, it semantically segments the cloud areas without labels. The approach shows competitive performance in all tested tasks, suggesting a new alternative for cloud characterization.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Real-time Anomaly Detection for Liquid Argon Time Projection Chambers

We present a real-time anomaly detection framework for liquid argon time projection chambers (LArTPCs), targeting applications in particle physics experiments such as the Short Baseline Near Detector (SBND) or the future Deep Underground Neutrino Experiment (DUNE). These experiments employ detectors that generate and stream high-resolution but sparse images of neutrino and other particle interactions. Our approach utilizes anomaly detection with autoencoders, compressed through knowledge distillation (KD), to enable the detection of anomalous signals in the data through efficient inference on resource-constrained hardware. The framework is targeted for deployment on computing platforms equipped with field-programmable gate arrays (FPGAs), GPUs, or CPUs, allowing low-latency selection of relevant activity directly from the raw detector data stream. We demonstrate that our approach is suitable for the detection and localization of anomalously "high-multiplicity" activity, and outline promising applications for LArTPC online data filtering and triggering.

FOS: Physical sciences↗
DOE OSTI2026

Dispersion Loss Counteracts Embedding Condensation and Improves Generalization in Small Language Models

Large language models (LLMs) achieve remarkable performance through ever-increasing parameter counts, but scaling incurs steep computational costs. To better understand LLM scaling, we study representational differences between LLMs and their smaller counterparts, with the goal of replicating the representational qualities of larger models in smaller models. We observe a geometric phenomenon which we term embedding condensation, where token embeddings collapse into a narrow cone-like subspace in some language models. Through systematic analyses across multiple Transformer families, we show that small models such as GPT2 and Qwen3-0.6B exhibit severe condensation, whereas larger models such as GPT2-x1 and Qwen3-32B are more resistant to this phenomenon. Additional observations show that embedding condensation is not reliably mitigated by knowledge distillation from larger models. To fight against it, we formulate a dispersion loss that explicitly encourages embedding dispersion during training. Experiments demonstrate that it mitigates condensation, recovers dispersion patterns seen in larger models, and yields performance gains across 10 benchmarks. We believe this work offers a principled path toward improving smaller Transformers without additional parameters.

Xiao, Xi [ORNL] (ORCID:0009000009316982)↗
DOE OSTI2022

Advancing river corridor science beyond disciplinary boundaries with an inductive approach to catalyse hypothesis generation

Abstract A unified conceptual framework for river corridors requires synthesis of diverse site‐, method‐ and discipline‐specific findings. The river research community has developed a substantial body of observations and process‐specific interpretations, but we are still lacking a comprehensive model to distill this knowledge into fundamental transferable concepts. We confront the challenge of how a discipline classically organized around the deductive model of systematically collecting of site‐, scale‐, and mechanism‐specific observations begins the process of synthesis. Machine learning is particularly well‐suited to inductive generation of hypotheses. In this study, we prototype an inductive approach to holistic synthesis of river corridor observations, using support vector machine regression to identify potential couplings or feedbacks that would not necessarily arise from classical approaches. This approach generated 672 relationships linking a suite of 157 variables each measured at 62 locations in a fifth order river network. Eighty four percent of these relationships have not been previously investigated, and representing potential (hypothetical) process connections. We document relationships consistent with current understanding including hydrologic exchange processes, microbial ecology, and the River Continuum Concept, supporting that the approach can identify meaningful relationships in the data. Moreover, we highlight examples of two novel research questions that stem from interpretation of inductively‐generated relationships. This study demonstrates the implementation of machine learning to sieve complex data sets and identify a small set of candidate relationships that warrant further study, including data types not commonly measured together. This structured approach complements traditional modes of inquiry, which are often limited by disciplinary perspectives and favour the careful pursuit of parsimony. Finally, we emphasize that this approach should be viewed as a complement to, rather than in place of, more traditional, deductive approaches to scientific discovery.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2022

Enhanced Oblique Decision Tree Enabled Policy Extraction for Deep Reinforcement Learning in Power System Emergency Control

Deep reinforcement learning (DRL) algorithms have successfully solved many challenging problems in various power system control scenarios. However, their decision-making process is usually regarded as black-boxes. Furthermore, how DRL models interact with human intelligence remains an open problem. Thus, this paper proposes a policy extraction framework to extract a complex DRL model into an explainable policy. This framework includes three parts: 1) DRL training and data generation. We train an agent for a specific control task and generate data, which contains the control policy of the agent. 2) Policy extraction. We propose an information gain rate based weighted oblique decision tree (IGR-WODT) for DRL policy extraction. 3) Policy evaluation. We define three metrics to evaluate the performance of the proposed approach. A case study for the under-voltage load shedding problem shows that the IGR-WODT presents a performance enhancement compared with DRL, weighted oblique decision tree, and univariate decision tree. The proposed policy extraction method could provide an intuitive explanation of the neural network decision-making process to the dispatchers when making final decisions on power grid operation. Also, the resulted rule-based controller could replace the deep neural network-based controller in many field edge devices with limited computing resources, providing comparable performance.

deep reinforcement learning↗
NASA NTRSDate not supplied

An Overview of NASA’s Newest Engineering Model, ORDEM 4.0

Since the mid-1990s, one of the most important products produced by the NASA Orbital Debris Program Office (ODPO) has been the Orbital Debris Engineering Model (ORDEM). This series of models distills down our knowledge of the orbital debris environment to compute debris fluxes on satellites in a given orbit. This information can be used by spacecraft and upper stage designers and operators to design missions for better protection against the debris environment. The current version of the model is ORDEM 3.2, but the ODPO is working on the next generation of ORDEM, to be designated ORDEM 4.0. ORDEM 4.0 will include many known features from previous models, such as the ability to input a spacecraft orbit and time and to compute the flux as a function of debris size, impact speed, impact direction, and debris material densities, as well as uncertainty information on the flux. ORDEM 4.0 will update debris populations using the most recent measurements, including radar observations by the Haystack Ultrawideband Satellite Imaging Radar (HUSIR), NASA’s Goldstone radar, data from the new Space Surveillance Network Space Fence, and observations of Geosynchronous Earth Orbits (GEO) using the Eugene Stansbery-Meter Class Autonomous Telescope (ES-MCAT). The latest in situ impact data from returned hardware surfaces will be used. In addition, ORDEM 4.0 will introduce a parameterized debris shape model based on laboratory hypervelocity impact tests, such as DebriSat. This will allow analysts to implement shape characteristics in their damage equations and more accurately predict impact damage risk by debris of different shapes and orientations. This paper provides an overview of some of the new features forthcoming in ORDEM 4.0 and a status report on its development.

Mark Matney↗
NASA NTRSDate not supplied

An Overview of NASA’s Newest Engineering Model, ORDEM 4.0

Since the mid-1990s, one of the most important products produced by the NASA Orbital Debris Program Office (ODPO) has been the Orbital Debris Engineering Model (ORDEM). This series of models distills down our knowledge of the orbital debris environment to compute debris fluxes on satellites in a given orbit. This information can be used by spacecraft and upper stage designers and operators to design missions for better protection against the debris environment. The current version of the model is ORDEM 3.2, but the ODPO is working on the next generation of ORDEM, to be designated ORDEM 4.0. ORDEM 4.0 will include many known features from previous models, such as the ability to input a spacecraft orbit and time and to compute the flux as a function of debris size, impact speed, impact direction, and debris material densities, as well as uncertainty information on the flux. ORDEM 4.0 will update debris populations using the most recent measurements, including radar observations by the Haystack Ultrawideband Satellite Imaging Radar (HUSIR), NASA’s Goldstone radar, data from the new Space Surveillance Network Space Fence, and observations of Geosynchronous Earth Orbits (GEO) using the Eugene Stansbery-Meter Class Autonomous Telescope (ES-MCAT). The latest in situ impact data from returned hardware surfaces will be used. In addition, ORDEM 4.0 will introduce a parameterized debris shape model based on laboratory hypervelocity impact tests, such as DebriSat. This will allow analysts to implement shape characteristics in their damage equations and more accurately predict impact damage risk by debris of different shapes and orientations. This paper provides an overview of some of the new features forthcoming in ORDEM 4.0 and a status report on its development.

Mark Matney↗
DOE OSTI2026

A transfer learning approach to energy-efficient control of small and medium-sized commercial buildings

Model-free reinforcement learning (RL) provides a data-driven and adaptive approach to optimize building energy use while satisfying occupant comfort. This powerful tool does not need any prior knowledge about the environment and system it is optimizing and can adapt its policy based on the changes in captures. Like any other data-driven tool, it faces high training costs due to the extensive agent-environment interactions required to capture long-term building dynamics and user comfort. Transfer learning, particularly policy distillation, offers a promising way to accelerate training by leveraging pretrained RL agents in different building and system types. Here, this study investigates online student distillation, in which the student model updates its neural network weights using outputs from teacher models. The work introduces a student distillation strategy designed for efficient knowledge transfer, along with a teacher selection method that ensures high-quality guidance. The approach is validated using a highly calibrated whole building energy model for a small/medium commercial building test facility. Results show substantial reductions in training time and data requirements while surpassing the performance of ASHRAE Guideline 36, an advanced rule-based control strategy. The distilled RL model required 45% less data and achieved 20% higher cumulative rewards than a state-of-the-art RL model, with faster convergence and lower energy consumption. These outcomes demonstrate that effective transfer learning enables a scalable and data-efficient energy management solution for commercial buildings.

ASHRAE guideline 36↗