Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “Convolutional Neural Networks”

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 649 records · Page 36
All sourcesNASA NTRSDOE OSTI
DOE OSTI2021

Iterative self-organizing SCEne-LEvel sampling (ISOSCELES) for large-scale building extraction

Convolutional neural networks (CNN) provide state-of-the-art performance in many computer vision tasks, including those related to remote-sensing image analysis. Successfully training a CNN to generalize well to unseen data, however, requires training on samples that represent the full distribution of variation of both the target classes and their surrounding contexts. With remote sensing data, acquiring a sufficiently representative training set is a challenge due to both the inherent multi-modal variability of satellite or aerial imagery and the general high cost of labeling data. To address this challenge, we have developed ISOSCELES, an Iterative Self-Organizing SCEne LEvel Sampling method for hierarchical sampling of large image sets. Using affinity propagation, ISOSCELES automates the selection of highly representative training images. Compared to random sampling or using available reference data, the distribution of the training is principally data driven, reducing the chance of oversampling uninformative areas or undersampling informative ones. In comparison to manual sample selection by an analyst, ISOSCELES exploits descriptive features, spectral and/or textural, and eliminates human bias in sample selection. Using a hierarchical sampling approach, ISOSCELES can obtain a training set that reflects both between-scene variability, such as in viewing angle and time of day, and within-scene variability at the level of individual training samples. We verify the method by demonstrating its superiority to stratified random sampling in the challenging task of adapting a pre-trained model to a new image and spatial domain for country-scale building extraction. Using a pair of hand-labeled training sets comprising 1,987 sample image chips, a total of 496,000,000 individually labeled pixels, we show, across three distinct model architectures, an increase in accuracy, as measured by F1-score, of 2.2–4.2%.

42 ENGINEERING↗
DOE OSTI2018

Machine learning to classify animal species in camera trap images: Applications in ecology

Motion-activated cameras (“camera traps”) are increasingly used in ecological and management studies for remotely observing wildlife and are amongst the most powerful tools for wildlife research. However, studies involving camera traps result in millions of images that need to be analysed, typically by visually observing each image, in order to extract data that can be used in ecological analyses. We trained machine learning models using convolutional neural networks with the ResNet-18 architecture and 3,367,383 images to automatically classify wildlife species from camera trap images obtained from five states across the United States. We tested our model on an independent subset of images not seen during training from the United States and on an out-of-sample (or “out-of-distribution” in the machine learning literature) dataset of ungulate images from Canada. We also tested the ability of our model to distinguish empty images from those with animals in another out-of-sample dataset from Tanzania, containing a faunal community that was novel to the model. The trained model classified approximately 2,000 images per minute on a laptop computer with 16 gigabytes of RAM. The trained model achieved 98% accuracy at identifying species in the United States, the highest accuracy of such a model to date. Out-of-sample validation from Canada achieved 82% accuracy and correctly identified 94% of images containing an animal in the dataset from Tanzania. We provide an r package (Machine Learning for Wildlife Image Classification) that allows the users to (a) use the trained model presented here and (b) train their own model using classified images of wildlife from their studies. The use of machine learning to rapidly and accurately classify wildlife in camera trap images can facilitate non-invasive sampling designs in ecological studies by reducing the burden of manually analysing images. Our r package makes these methods accessible to ecologists.

54 ENVIRONMENTAL SCIENCES↗
DOE OSTI2025

Object detection with deep learning for rare event search in the GADGET II TPC

In the pursuit of identifying rare two-particle events within the GADGET II Time Projection Chamber (TPC), this paper presents a comprehensive approach for leveraging Convolutional Neural Networks (CNNs) and various data processing methods. To address the inherent complexities of 3D TPC track reconstructions, the data is expressed in 2D projections and 1D quantities. This approach capitalizes on the diverse data modalities of the TPC, allowing for the efficient representation of the distinct features of the 3D events, with no loss in topology uniqueness. Additionally, it leverages the computational efficiency of 2D CNNs and benefits from the extensive availability of pre-trained models. Given the scarcity of real training data for the rare events of interest, simulated events are used to train the models to detect real events. To account for potential distribution shifts when predominantly depending on simulations, significant perturbations are embedded within the simulations. This produces a broad parameter space that works to account for potential physics parameter and detector response variations and uncertainties. These parameter-varied simulations are used to train sensitive 2D CNN object detectors. When combined with 1D histogram peak detection algorithms, this multi-modal detection framework is highly adept at identifying rare, two-particle events in data taken during experiment 21072 at the Facility for Rare Isotope Beams (FRIB), demonstrating a 100% recall for events of interest. Here, we present the methods and outcomes of our investigation and discuss the potential future applications of these techniques.

Convolutional neural network↗
DOE OSTI2026

ML-based Micro-CT SOFC Microstructure Models (from Kent 2026 Microstructural Augmentation paper)

Overview -------------------------- This repository contains datasets from the manuscript **"Enhanced Generalizability to Deep-Learning Quantification of 3D Microstructural Characteristics through Microstructurally Aware Augmentation of Scarce Data"** (*William F. Kent, Rochan Bajpai, Rachel C. Kurchin, William K. Epting, Harry W. Abernathy, Paul A. Salvador. Submitted 2026*). The methods are also described in the dissertation **Data Intensive Analysis of Solid Oxide Cell Microstructures** (*Doctoral dissertation, Carnegie Mellon University, 2025*). The datasets here are trained convolutional neural network (CNN) models for predicting key microstructural properties of solid oxide cell (SOC) electrodes from low-res, 2-channel 3D images, as well as some helpful code. The parameters for input images are provided in the paper. Sample data is provided in the file `Combined_anode_aug_dual_1k_examples` - that particular data was used to train `anode_all_aug.pth` and will work most accurately with that model. Please familiarize yourself with all caveats on accuracy and applicability, as detailed in the associated paper. Usage -------------------------- The basic usage is as follows, assuming `model_fn` is the path to the .pth file, and `X` is 2-channel input image(s) of the proper dimensions (either one image of shape `[2,12,24,24]`, or a batch of N input images of shape `[N,2,12,24,24]`): from CNN_inferencer import load_model_for_inference model = load_model_for_inference(model_fn) y_predicted = model(X) The model object automatically handles input scaling and output de-scaling based on the way the models were trained - in other words, pass in a 2-channel micro-CT image, and it will output microstructural property values in real units. ## Other model object attributes Note that model has useful attributes other than its forward pass model(X). * `model.output_descaler` - returns the output descaler object. Model does the de-scaling when generating inferences, but you may want to re-use this de-scaler on other values to e.g. compare predictions to ground truth from already-scaled training data. * `model.prop_names` - Gives the property names of the predicted y values, in order. Only exists if there's an output scaler as part of the model object, which there will be in the models provided here. ## Usage with sample data Here is a short script to use with the included sample data. from CNN_inferencer import display_predictions, load_model_for_inference, calculate_mape, parity_plot import h5py import numpy as np model_fn = 'anode_all_aug.pth' data_fn = 'Combined_anode_aug_dual_1k_examples.h5' N_samples = 200 figure_outdir = '.' model = load_model_for_inference(model_fn) with h5py.File(data_fn,'r') as f: XX = f['X'] #These are the 2-channel 3D images yy = f['y'] #These are the ground-truth microstructural properties, but they have been scaled for training - need to de-scale below N = XX.shape[0] #How many images total in the input data file #Run inferences on N_samples random samples from XX. #Run in a batch, much more efficient than one at a time. ii = np.random.choice(N,N_samples,replace=False) ii.sort() y_pred = model(XX[ii]) #Get the original/true (but normalized/scaled) values from the training dataset... #Because they were normalized, they are not in real units yet. So let's also de-scale them using model.output_scaler. y_true = model.output_scaler.transform(yy[ii]) #Let's display actual values for just 5 random ones for i in np.random.choice(N_samples,5,replace=False): display_predictions(y_true[i], y_pred[i], model.prop_names) #Make parity plots for each property (ground truth vs predicted values) #Also label each plot with the mean abs. percent error (MAPE) of the predicted values for i,key in enumerate(model.prop_names): mape = calculate_mape(y_true[:,i], y_pred[:,i]) parity_plot(y_true[:,i], y_pred[:,i], figure_outdir, key, extra_title=f' ({mape:.2f}% MAPE)')

3D microstructure↗
DOE OSTI2022

Photometric redshifts from SDSS images with an interpretable deep capsule network

ABSTRACT Studies of cosmology, galaxy evolution, and astronomical transients with current and next-generation wide-field imaging surveys like the Rubin Observatory Legacy Survey of Space and Time are all critically dependent on estimates of photometric redshifts. Capsule networks are a new type of neural network architecture that is better suited for identifying morphological features of the input images than traditional convolutional neural networks. We use a deep capsule network trained on ugriz images, spectroscopic redshifts, and Galaxy Zoo spiral/elliptical classifications of ∼400 000 Sloan Digital Sky Survey galaxies to do photometric redshift estimation. We achieve a photometric redshift prediction accuracy and a fraction of catastrophic outliers that are comparable to or better than current methods for SDSS main galaxy sample-like data sets (r ≤ 17.8 and zspec ≤ 0.4) while requiring less data and fewer trainable parameters. Furthermore, the decision-making of our capsule network is much more easily interpretable as capsules act as a low-dimensional encoding of the image. When the capsules are projected on a two-dimensional manifold, they form a single redshift sequence with the fraction of spirals in a region exhibiting a gradient roughly perpendicular to the redshift sequence. We perturb encodings of real galaxy images in this low-dimensional space to create synthetic galaxy images that demonstrate the image properties (e.g. size, orientation, and surface brightness) encoded by each dimension. We also measure correlations between galaxy properties (e.g. magnitudes, colours, and stellar mass) and each capsule dimension. We publicly release our code, estimated redshifts, and additional catalogues at https://biprateep.github.io/encapZulate-1.

79 ASTRONOMY AND ASTROPHYSICS↗
DOE OSTI2019

Analyzing machine learning models to accelerate generation of fundamental materials insights

Abstract Machine learning for materials science envisions the acceleration of basic science research through automated identification of key data relationships to augment human interpretation and gain scientific understanding. A primary role of scientists is extraction of fundamental knowledge from data, and we demonstrate that this extraction can be accelerated using neural networks via analysis of the trained data model itself rather than its application as a prediction tool. Convolutional neural networks excel at modeling complex data relationships in multi-dimensional parameter spaces, such as that mapped by a combinatorial materials science experiment. Measuring a performance metric in a given materials space provides direct information about (locally) optimal materials but not the underlying materials science that gives rise to the variation in performance. By building a model that predicts performance (in this case photoelectrochemical power generation of a solar fuels photoanode) from materials parameters (in this case composition and Raman signal), subsequent analysis of gradients in the trained model reveals key data relationships that are not readily identified by human inspection or traditional statistical analyses. Human interpretation of these key relationships produces the desired fundamental understanding, demonstrating a framework in which machine learning accelerates data interpretation by leveraging the expertize of the human scientist. We also demonstrate the use of neural network gradient analysis to automate prediction of the directions in parameter space, such as the addition of specific alloying elements, that may increase performance by moving beyond the confines of existing data.

36 MATERIALS SCIENCE↗
DOE OSTI2017

Uncertainties in Parameters Estimated with Neural Networks: Application to Strong Gravitational Lensing

In Hezaveh et al. (2017) we showed that deep learning can be used for model parameter estimation and trained convolutional neural networks to determine the parameters of strong gravitational lensing systems. Here we demonstrate a method for obtaining the uncertainties of these parameters. We review the framework of variational inference to obtain approximate posteriors of Bayesian neural networks and apply it to a network trained to estimate the parameters of the Singular Isothermal Ellipsoid plus external shear and total flux magnification. We show that the method can capture the uncertainties due to different levels of noise in the input data, as well as training and architecture-related errors made by the network. To evaluate the accuracy of the resulting uncertainties, we calculate the coverage probabilities of marginalized distributions for each lensing parameter. By tuning a single hyperparameter, the dropout rate, we obtain coverage probabilities approximately equal to the confidence levels for which they were calculated, resulting in accurate and precise uncertainty estimates. Our results suggest that neural networks can be a fast alternative to Monte Carlo Markov Chains for parameter uncertainty estimation in many practical applications, allowing more than seven orders of magnitude improvement in speed.

Perreault Levasseur, Laurence↗
NASA NTRS2018

NeMO-Net & Fluid Lensing: The Neural Multi-Modal Observation & Training Network for Global Coral Reef Assessment Using Fluid Lensing Augmentation of NASA EOS Data

We present preliminary results from NASA NeMO-Net, the first neural multi-modal observation and training network for global coral reef assessment. NeMO-Net is an open-source deep convolutional neural network (CNN) and interactive active learning training software in development which will assess the present and past dynamics of coral reef ecosystems. NeMO-Net exploits active learning and data fusion of mm-scale remotely sensed 3D images of coral reefs captured using fluid lensing with the NASA FluidCam instrument, presently the highest-resolution remote sensing benthic imaging technology capable of removing ocean wave distortion, as well as hyperspectral airborne remote sensing data from the ongoing NASA CORAL mission and lower-resolution satellite data to determine coral reef ecosystem makeup globally at unprecedented spatial and temporal scales. Aquatic ecosystems, particularly coral reefs, remain quantitatively misrepresented by low-resolution remote sensing as a result of refractive distortion from ocean waves, optical attenuation, and remoteness. Machine learning classification of coral reefs using FluidCam mm-scale 3D data show that present satellite and airborne remote sensing techniques poorly characterize coral reef percent living cover, morphology type, and species breakdown at the mm, cm, and meter scales. Indeed, current global assessments of coral reef cover and morphology classification based on km-scale satellite data alone can suffer from segmentation errors greater than 40%, capable of change detection only on yearly temporal scales and decameter spatial scales, significantly hindering our understanding of patterns and processes in marine biodiversity at a time when these ecosystems are experiencing unprecedented anthropogenic pressures, ocean acidification, and sea surface temperature rise. NeMO-Net leverages our augmented machine learning algorithm that demonstrates data fusion of regional FluidCam (mm, cm-scale) airborne remote sensing with global low-resolution (m, km-scale) airborne and spaceborne imagery to reduce classification errors up to 80% over regional scales. Such technologies can substantially enhance our ability to assess coral reef ecosystems dynamics.

satellite data↗
NASA NTRS2018

NASA NeMO-Net

We present preliminary results from NASA NeMO-Net, the first neural multi-modal observation and training network for global coral reef assessment. NeMO-Net is an open-source deep convolutional neural network (CNN) and interactive active learning training software in development which will assess the present and past dynamics of coral reef ecosystems. NeMO-Net exploits active learning and data fusion of mm-scale remotely sensed 3D images of coral reefs captured using fluid lensing with the NASA FluidCam instrument, presently the highest-resolution remote sensing benthic imaging technology capable of removing ocean wave distortion, as well as hyperspectral airborne remote sensing data from the ongoing NASA CORAL mission and lower-resolution satellite data to determine coral reef ecosystem makeup globally at unprecedented spatial and temporal scales. Aquatic ecosystems, particularly coral reefs, remain quantitatively misrepresented by low- resolution remote sensing as a result of refractive distortion from ocean waves, optical attenuation, and remoteness. Machine learning classification of coral reefs using FluidCam mm-scale 3D data show that present satellite and airborne remote sensing techniques poorly characterize coral reef percent living cover, morphology type, and species breakdown at the mm, cm, and meter scales. Indeed, current global assessments of coral reef cover and morphology classification based on km-scale satellite data alone can suffer from segmentation errors greater than 40%, capable of change detection only on yearly temporal scales and decameter spatial scales, significantly hindering our understanding of patterns and processes in marine biodiversity at a time when these ecosystems are experiencing unprecedented anthropogenic pressures, ocean acidification, and sea surface temperature rise. NeMO-Net leverages our augmented machine learning algorithm that demonstrates data fusion of regional FluidCam (mm, cm-scale) airborne remote sensing with global low-resolution (m, km-scale) airborne and spaceborne imagery to reduce classification errors up to 80% over regional scales. Such technologies can substantially enhance our ability to assess coral reef ecosystems dynamics.

NASA↗
DOE OSTI2020

A Multiple Filter Based Neural Network Approach to the Extrapolation of Adsorption Energies on Metal Surfaces for Catalysis Applications

Computational catalyst discovery involves the development of microkinetic reactor models based on estimated parameters determined from density functional theory (DFT). For complex surface chemistries, the number of reaction intermediates can be very large, and the cost of calculating the adsorption energies by DFT for all surface intermediates even for one active site model can become prohibitive. In this paper, we have identified appropriate descriptors and machine learning models that can be used to predict a significant part of these adsorption energies given data on the rest of them. Furthermore, our investigations also included the case when the species data used to train the predictive model are of different size relative to the species the model tries to predict—this is an extrapolation in the data space which is typically difficult with regular machine learning models. Due to the relative size of the available data sets, we have attempted to extrapolate from the larger species to the smaller ones in the current work. Here, we have developed a neural network based predictive model that combines an established additive atomic contribution based model with the concepts of a convolutional neural network that, when extrapolating, achieves a statistically significant improvement over the previous models.

37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
DOE OSTI2019

Evolving Larger Convolutional Layer Kernel Sizes for a Settlement Detection Deep-Learner on Summit

Deep-learner hyper-parameters, such as kernel sizes, batch sizes, and learning rates, can significantly influence the quality of trained models. The state of the art for finding optimal hyper-parameters generally uses a brute force, grid search approach, random search, or Bayesian-based optimization among other techniques. We applied an evolutionary algorithm to optimize kernel sizes for a convolutional neural network used to detect settlements in satellite imagery. Usually convolutional layer kernel sizes are small - typically one, three, or five - but we found that the system converged at, or near, kernel sizes of nine for the last convolutional layer, and that this occurred for multiple runs using two different datasets. Moreover, the larger kernel sizes had fewer false positives than the 3x3 kernel sizes found as optimal via a brute force uniform grid search. This suggests that this large kernel size may be leveraging patterns found in larger areal features in the source imagery, and that this may be generalized as possible guidance for similar remote sensing deep-learning tasks.

Coletti, Mark↗
DOE OSTI2022

Effectiveness of Deep Learning Trained on SynthCity Data for Urban Point-Cloud Classification

3D object recognition is one of the most popular areas of study in computer vision. Many of the more recent algorithms focus on indoor point clouds, classifying 3D geometric objects, and segmenting outdoor 3D scenes. One of the challenges of the classification pipeline is finding adequate and accurate training data. Hence, this article seeks to evaluate the accuracy of a synthetically generated data set called SynthCity, tested on two mobile laser-scan data sets. Varying levels of noise were applied to the training data to reflect varying levels of noise in different scanners. The chosen deep-learning algorithm was Kernel Point Convolution, a convolutional neural network that uses kernel points in Euclidean space for convolution weights.

Geology↗
DOE OSTI2020

AI Denoising to Accelerate Detector Simulation

Detector simulation is critical to experimental HEP; however this simulation (commonly done through toolkits such as Geant4) is computationally intensive. Performance can be improved somewhat through technical optimization, but more is needed. Using machine learning (ML) to accelerate simulation is a promising field, however efforts to use generative adversarial networks (GANs) or optimized autoencoders have faced issues. Using convolutional neural networks (CNNs) for denoising has been successful in non-HEP applications such as image processing. This poster investigates the efficacy of using CNNs to denoise Geant4 simulations. This could increase the accuracy of simulations performed under settings designed to increase computational efficiency.

Franklin, Lena↗
DOE OSTI2018

Evaluation and Comparison of Machine Learning Techniques for Rapid QSTS Simulations

Rapid and accurate quasi-static time series (QSTS) analysis is becoming increasingly important for distribution system analysis as the complexity of the distribution system intensifies with the addition of new types, and quantities, of distributed energy resources (DER). The expanding need for hosting capacity analysis, control systems analysis, photovoltaic (PV) and DER impact analysis, and maintenance cost estimations are just a few reasons that QSTS is necessary. Historically, QSTS analysis has been prohibitively slow due to the number of computations required for a full-year analysis. Therefore, new techniques are required that allow QSTS analysis to rapidly be performed for many different use cases. This research demonstrates a novel approach to doing rapid QSTS analysis for analyzing the number of voltage regulator tap changes in a distribution system with PV components. A representative portion of a yearlong dataset is selected and QSTS analysis is performed to determine the number of tap changes, and this is used as training data for a machine learning algorithm. The machine learning algorithm is then used to predict the number of tap changes in the remaining portion of the year not analyzed directly with QSTS. The predictions from the machine learning algorithms are combined with the results of the partial year simulation for a final prediction for the entire year, with the goal of maintaining an error <10% on the full-year prediction. Five different machine learning techniques were evaluated and compared with each other; a neural network ensemble, a random forest decision tree ensemble, a boosted decision tree ensemble, support vector machines, and a convolutional neural network deep learning technique. A combination of the neural network ensemble together with the random forest produced the best results. Using 20% of the year as training data, analyzed with QSTS, the average performance of the technique resulted in ~2.5% error in the yearly tap changes, while maintaining a <10% 99.9 th percentile error bound on the results. This is a 5x speedup compared to a standard, full-length QSTS simulation. These results demonstrate the potential for applying machine learning techniques to facilitate modern distribution system analysis and further integration of distributed energy resources into the power grid.

24 POWER TRANSMISSION AND DISTRIBUTION↗
DOE OSTI2023

Graph neural networks predict energetic and mechanical properties for models of solid solution metal alloy phases

Here, we developed a PyTorch-based architecture called HydraGNN that implements graph convolutional neural networks (GCNNs) to predict the formation energy and the bulk modulus for models of solid solution alloys for various atomic crystal structures and relaxed volumes. We trained the GCNN surrogate model on a dataset for nickel–niobium (NiNb) generated by the embedded atom model (EAM) empirical interatomic potential for demonstration purposes. The dataset was generated by calculating the formation energy and the bulk modulus as a prototypical elastic property for optimized geometries starting from initial body-centered cubic (BCC), face-centered cubic (FCC), and hexagonal compact packed (HCP) crystal structures, with configurations spanning the possible compositional range for each of the three types of initial crystal structures. Numerical results show that the GCNN model effectively predicts both the formation energy and the bulk modulus as function of the optimized crystal structure, relaxed volume, and configurational entropy of the model structures for solid solution alloys.

36 MATERIALS SCIENCE↗
DOE OSTI2023

Laser Wakefield Accelerator modelling with Variational Neural Networks

A machine learning model was created to predict the electron spectrum generated by a GeV-class laser wakefield accelerator. The model was constructed from variational convolutional neural networks, which mapped the results of secondary laser and plasma diagnostics to the generated electron spectrum. An ensemble of trained networks was used to predict the electron spectrum and to provide an estimation of the uncertainty of that prediction. It is anticipated that this approach will be useful for inferring the electron spectrum prior to undergoing any process that can alter or destroy the beam. In addition, the model provides insight into the scaling of electron beam properties due to stochastic fluctuations in the laser energy and plasma electron density.

43 PARTICLE ACCELERATORS↗
DOE OSTI2018

Comparison of Shallow and Deep Learning Methods on Classifying the Regional Pattern of Diffuse Lung Disease

This study aimed to compare shallow and deep learning of classifying the patterns of interstitial lung diseases (ILDs). Using high-resolution computed tomography images, two experienced radiologists marked 1200 regions of interest (ROIs), in which 600 ROIs were each acquired using a GE or Siemens scanner and each group of 600 ROIs consisted of 100 ROIs for subregions that included normal and five regional pulmonary disease patterns (ground-glass opacity, consolidation, reticular opacity, emphysema, and honeycombing). We employed the convolution neural network (CNN) with six learnable layers that consisted of four convolution layers and two fully connected layers. The classification results were compared with the results classified by a shallow learning of a support vector machine (SVM). The CNN classifier showed significantly better performance for accuracy compared with that of the SVM classifier by 6–9%. As the convolution layer increases, the classification accuracy of the CNN showed better performance from 81.27 to 95.12%. Especially in the cases showing pathological ambiguity such as between normal and emphysema cases or between honeycombing and reticular opacity cases, the increment of the convolution layer greatly drops the misclassification rate between each case. Conclusively, the CNN classifier showed significantly greater accuracy than the SVM classifier, and the results implied structural characteristics that are inherent to the specific ILD patterns.

62 RADIOLOGY AND NUCLEAR MEDICINE↗
DOE OSTI2021

Mitigating Catastrophic Forgetting in Deep Learning in a Streaming Setting Using Historical Summary

Recent advancements in scientific equipment and the adaptation of electronics and the Internet of Things (IoT) in our everyday lives resulted in large and complex data production at a high rate. Making meaningful and timely knowledge discovery at a modest cost from this big data is difficult for computing power and storage limitations. Training deep learning models incrementally in a streaming setting can help us with overcoming these limitations. However, in a well-known phenomenon named catastrophic forgetting, incrementally trained models increasingly perform poorly on the past data. To mitigate catastrophic forgetting in training in a streaming setting, we propose constructing a historical summary over time and use the summary with newly arrived data during incremental training. We propose various data summarization techniques such as random sampling, micro clustering, coreset computation, and Auto Encoders to counteract catastrophic forgetting. We built a pipeline for incremental training with a historical summary for training deep learning models for streaming data. We demonstrate the effectiveness of historical summary in mitigating catastrophic forgetting using three case studies involving three different deep learning applications: an Artificial Neural Network (ANN) for classification task on MNIST dataset, a language model (RNN-LM) on the WikiText2 dataset, and a Convolutional Neural Network (CNN), ResNet50 to classify the ImageNet dataset. Through the training of the models, we observe that catastrophic forgetting is evident in ANN and CNN but not in an RNN. For the first task, our method recovers up to 47.9% lost accuracy due to catastrophic forgetting. For the third task, the historical summary recovers classification accuracy by up to 25%. For the second task, though there is not proof of catastrophic forgetting, the training performance (PPL) improves by up to 26% with historical summary.

Dash, Sajal↗