Engineering topics

Siegel, Charles M.

Publications and source records attributed to Siegel, Charles M..

9 records

DOE OSTI2019

Scaling studies for deep learning in Liquid Argon Time Projection Chamber event classification

Measurements in Liquid Argon Time Projection Chamber neutrino detectors feature large, high fidelity event images. Deep learning techniques have been extremely successful in classification tasks of photographs, but their application to these event images is challenging, due to the large size of the events, more two orders of magnitude larger than images found in classical chal- lenges like MNIST or ImageNet. This leads to extremely long training cycles, which slow down the exploration of new network architectures and hyperpa- rameter scans to improve the classification performance. We present studies of scaling an LArTPC classification problem on multiple architectures, spanning multiple nodes. The studies are carried out in simulated events in the MicroBooNE detector.

Strube, Jan F.↗

DOE OSTI2018

Deep Learning for Scientific Discovery

Deep learning has positively impacted fields were perceptive tasks are paramount, such as computer vision and natural language processing. Tremendous strides have also been made in deep reinforcement learning towards robotics and gaming. Given the success in these fields, it is apparent that deep learning may also accelerate scientific discovery. Herein, we discuss several topic areas where modern representation learning is driving innovation. This article presents a) learning non-linear models from data, b) learning scientific laws from models, c) deep learning on graphs, and d) approaches to scalable deep learning.

deep learning, scientific computing, koopman, mate↗

DOE OSTI2018

Using Rule-Based Models for Weak Supervised Learning: A ChemNet for Transferable Chemical Property Prediction

With access to large datasets, deep neural networks (DNN) have achieved human-level accuracy in image and speech recognition tasks. However, in chemistry, data is inherently small and fragmented. In this work, we develop an approach of integrating rule-based knowledge into the Chemception CNN model via transfer learning techniques. The resulting model, ChemNet, is a transferable and generalizable deep neural network for chemical property prediction that learns in a semi-supervised manner from large unlabeled chemical databases. When ChemNet is further fine-tuned on 3 smaller datasets to predict chemical properties that it was not originally trained on, we show that ChemNet exceeds the accuracy of existing Chemception CNN models and other contemporary DNN models that were trained using conventional supervised learning approaches. These results indicate that pre-training ChemNet on a large diverse chemical database while incorporating chemistry domain knowledge, enables the development of more generalizable deep neural networks for the prediction of novel chemical properties.

Artificial Intelligence, Cheminformatics, Computat↗

DOE OSTI2018

How Much Chemistry Does a Deep Neural Network Need to Know to Make Accurate Predictions?

In the last few years, we have seen the rise of deep learning applications in a broad range of computational chemistry research problems. Using human-engineered chemical features, such as molecular descriptors and fingerprints, deep learning models have shown similar, if not better performance that most traditional machine learning algorithms. Recently, we reported on the development of Chemception, a deep convolutional neural network (CNN) architecture for general-purpose small molecule property prediction. On average, Chemception matched the performance of expert-developed QSAR/QSPR models trained on chemical features (molecular fingerprints), despite that it was trained on just 2D images of molecular drawings with minimal chemical information. Here, we investigate the effects of systematically removing and adding basic chemical information to the image channels of the 2D images used to train Chemception. By augmenting our images with only 3 additional basic chemical information, we demonstrate the improvement of Chemception performance – that it is now more accurate than contemporary deep learning models trained on ECFP fingerprints for the prediction of toxicity, activity and solvation free energy, as well as physics-based free energy simulation methods for computing solvation properties. By altering the chemical information content in the image channels, and examining the resulting performance of Chemception, we also identify to two different “learning patterns” in toxicity/activity as compared to solvation free energy, and it parallels the fundamental differences in contemporary chemistry research for predicting toxicity/activity and solvation free energy.

Artificial Intelligence, Cheminformatics, Computat↗

DOE OSTI2018

A Learning Framework for Control-Oriented Modeling of Buildings

Buildings consume a significant amount of energy worldwide. Several building optimization and control use cases require models of energy consumption which are control oriented, have high predictive capability, imposes minimal data pre-processing requirements, and have the ability to be adapted continuously to account for changing conditions as new data becomes available. Data driven modeling techniques, that have been investigated so far, while promising in the context of buildings, have been unable to simultaneously satisfy all the requirements mentioned above. In this context, deep learning techniques such as Recurrent Neural Networks (RNNs) hold promise, empowered by advanced computational capabilities and big data opportunities. In this paper, we propose a deep learning based methodology for the development of control oriented models for building energy management and test in on data from a real building. Results show that the proposed methodology outperforms other data driven modeling techniques significantly. We perform a detailed analysis of the proposed methodology along dimensions such as topology, sensitivity, and downsampling. Lastly, we conclude by envisioning a building analytics suite empowered by the proposed deep framework, that can drive several use cases related to building energy management.

buildings↗

DOE OSTI2018

Evaluating On-Node GPU Interconnects for Deep Learning Workloads

Scaling deep learning workloads across multiple GPUs on a single node has become increasingly important in data analytics. A key question is how well a PCIe-based GPU interconnect can perform relative to a custom high-performance interconnect such as NVIDIA's NVLink. This paper evaluates two such on-node interconnects for eight NVIDIA Pascal P100 GPUs: (a) the NVIDIA DGX-1's NVLink 1.0 `hybrid cube mesh'; and (b) the Cirrascale GX8's two-level PCIe tree using dual SR3615 switch risers. To show the effects of a range of neural network workloads, we define a parameterized version of the popular ResNet. We define a workload intensity metric that characterizes the expected computation/communication ratio; we also locate AlexNet and GoogLeNet within that space. As expected, the DGX-1 typically has superior performance. However, when equalizing GPU SM frequencies, the GX8 is very competitive on all ResNet workloads. With 8 GPUs, the GX8 can outperform the DGX-1 on all-to-all reductions by 10% for medium-sized payloads; and in rare cases, the GX8 slightly outperforms on ResNet.

Tallent, Nathan R.↗

DOE OSTI2017

ChemNet: A Transferable and Generalizable Deep Neural Network for Small-Molecule Property Prediction

With access to large datasets, deep neural networks through representation learning have been able to identify patterns from raw data, achieving human-level accuracy in image and speech recognition tasks. However, in chemistry, availability of large standardized and labelled datasets is scarce, and with a multitude of chemical properties of interest, chemical data is inherently small and fragmented. In this work, we explore transfer learning techniques in conjunction with the existing Chemception CNN model, to create a transferable and generalizable deep neural network for small-molecule property prediction. Our latest model, ChemNet learns in a semi-supervised manner from inexpensive labels computed from the ChEMBL database. When fine-tuned to the Tox21, HIV and FreeSolv dataset, which are 3 separate chemical tasks that ChemNet was not originally trained on, we demonstrate that ChemNet exceeds the performance of existing Chemception models, contemporary MLP models that trains on molecular fingerprints, and it matches the performance of the ConvGraph algorithm, the current state-of-the-art. Furthermore, as ChemNet has been pre-trained on a large diverse chemical database, it can be used as a universal “plug-and-play” deep neural network, which accelerates the deployment of deep neural networks for the prediction of novel small-molecule chemical properties.

Artificial Intelligence↗

DOE OSTI2017

What does fault tolerant Deep Learning need from MPI?

Deep Learning (DL) algorithms have become the {\em de facto} Machine Learning (ML) algorithm for large scale data analysis. DL algorithms are computationally expensive -- even distributed DL implementations which use MPI require days of training (model learning) time on commonly studied datasets. Long running DL applications become susceptible to faults -- requiring development of a fault tolerant system infrastructure, in addition to fault tolerant DL algorithms. This raises an important question: {\em What is needed from MPI for designing fault tolerant DL implementations?} In this paper, we address this problem for permanent faults. We motivate the need for a fault tolerant MPI specification by an in-depth consideration of recent innovations in DL algorithms and their properties, which drive the need for specific fault tolerance features. We present an in-depth discussion on the suitability of different parallelism types (model, data and hybrid); a need (or lack thereof) for check-pointing of any critical data structures; and most importantly, consideration for several fault tolerance proposals (user-level fault mitigation (ULFM), Reinit) in MPI and their applicability to fault tolerant DL implementations. We leverage a distributed memory implementation of Caffe, currently available under the Machine Learning Toolkit for Extreme Scale (MaTEx). We implement our approaches by extending MaTEx-Caffe for using ULFM-based implementation. Our evaluation using the ImageNet dataset and AlexNet neural network topology demonstrates the effectiveness of the proposed fault tolerant DL implementation using OpenMPI based ULFM.

Amatya, Vinay C.↗

DOE OSTI2017

Adaptive Neuron Apoptosis for Accelerating Deep Learning on Large Scale Systems

Machine Learning and Data Mining (MLDM) algorithms are becoming ubiquitous in {\em model learning} from the large volume of data generated using simulations, experiments and handheld devices. Deep Learning algorithms -- a class of MLDM algorithms -- are applied for automatic feature extraction, and learning non-linear models for unsupervised and supervised algorithms. Naturally, several libraries which support large scale Deep Learning -- such as TensorFlow and Caffe -- have become popular. In this paper, we present novel techniques to accelerate the convergence of Deep Learning algorithms by conducting low overhead removal of redundant neurons -- {\em apoptosis} of neurons -- which do not contribute to model learning, during the training phase itself. We provide in-depth theoretical underpinnings of our heuristics (bounding accuracy loss and handling apoptosis of several neuron types), and present the methods to conduct adaptive neuron apoptosis. We implement our proposed heuristics with the recently introduced TensorFlow and using its recently proposed extension with MPI. Our performance evaluation on two difference clusters -- one connected with Intel Haswell multi-core systems, and other with nVIDIA GPUs -- using InfiniBand, indicates the efficacy of the proposed heuristics and implementations. Specifically, we are able to improve the training time for several datasets by 2-3x, while reducing the number of parameters by 30x (4-5x on average) on datasets such as ImageNet classification. For the Higgs Boson dataset, our implementation improves the accuracy (measured by Area Under Curve (AUC)) for classification from 0.88/1 to 0.94/1, while reducing the number of parameters by 3x in comparison to existing literature, while achieving a 2.44x speedup in comparison to the default (no apoptosis) algorithm.

Siegel, Charles M.↗