Eric Goodman

Spatiotemporal Pattern Recognition via Liquid State Machines

The applicability of complex networks of spiking neurons as a general purpose machine learning technique remains open. Building on previous work using macroscopic exploration of the parameter space of an (artificial) neural microcircuit, we investigate the possibility of using a liquid state machine to solve two real-world problems: stockpile surveillance signal alignment and spoken phoneme recognition.

High-performance computing applied to semantic databases

To-date, the application of high-performance computing resources to Semantic Web data has largely focused on commodity hardware and distributed memory platforms. In this paper we make the case that more specialized hardware can offer superior scaling and close to an order of magnitude improvement in performance. In particular we examine the Cray XMT. Its key characteristics, a large, global sharedmemory, and processors with a memory-latency tolerant design, offer an environment conducive to programming for the Semantic Web and have engendered results that far surpass current state of the art. We examine three fundamental pieces requisite for a fully functioning semantic database: dictionary encoding, RDFS inference, and query processing. We show scaling up to 512 processors (the largest configuration we had available), and the ability to process 20 billion triples completely in-memory.

Hashing strategies for the Cray XMT

Two of the most commonly used hashing strategies-linear probing and hashing with chaining-are adapted for efficient execution on a Cray XMT. These strategies are designed to minimize memory contention. Datasets that follow a power law distribution cause significant performance challenges to shared memory parallel hashing implementations. Experimental results show good scalability up to 128 processors on two power law datasets with different data types: integer and string. These implementations can be used in a wide range of applications.

Effectively using recurrently-connected spiking neural networks

Recurrently connected spiking neural networks are difficult to use and understand because of the complex nonlinear dynamics of the system. Through empirical studies of spiking networks, we deduce several principles which are critical to success. Network parameters such as synaptic time delays and time constants and the connection probabilities can be adjusted to have a significant impact on accuracy. We show how to adjust these parameters to fit the type of problem.

Scalable hashing for shared memory supercomputers

Hashing is a fundamental technique in computer science to allow O(l) insert and lookups of items in an associative array. Here we present several thread coordination and hashing strategies and compare and contrast their performance on large, shared memory symmetric multiprocessor machines, each possessing between a half to a full terabyte of memory. We show how our approach can be used as a key kernel for fundamental paradigms such as dynamic programming and MapReduce. We further show that a set of approaches yields close to linear speedup for both uniform random and more difficult power law distributions. This scalable performance is in spite of the fact that our set of approaches is not completely lock-free. Our experimental results utilize and compare an SGI Altix UV with 4 Xeon processors (32 cores) and a Cray XMT with 128 processors. On the scale of data we addressed, on the order of 5 billion integers, we show that the Altix UV far exceeds the performance of the Cray XMT for power law distributions. However, the Cray XMT exhibits greater scalability.

Using Bipartite Anomaly Features for Cyber Security Applications

In this paper we use anomaly scores derived from a technique for bipartite graphs as features for a supervised machine learning algorithm for two cyber security problems: classifying Short Message Service (SMS) text messages as either spam or non-spam and detecting malicious lateral movement within a network. While disparate problems, both spam and lateral movement detection can be viewed as bipartite graphs and we can compute bipartite anomaly scores for each situation. The bipartite anomaly scores by themselves are not very predictive, but used as auxiliary features can boost the receiver operating characteristic (ROC) curve of a supervised classifier. We examine the UCI SMS Spam Collection Data Set for the SPAM problem and use an authentication graph from Los Alamos National Laboratory. We create features by dimensionality reduction through principal component analysis (PCA) on the message-term or user-computer matrix, and then augment those features with anomaly scores. By using the anomaly scores we are able to improve the area under the curve (AUC) for the receiver operating characteristic (ROC) up to 27.5% for the spam data and 21.4% for the authentication data.

Irregular large-scale computed tomography on multiple graphics processors improves energy-efficiency metrics for industrial applications

This paper will investigate energy-efficiency for various real-world industrial computed-tomography reconstruction algorithms, both CPU- and GPU-based implementations. This work shows that the energy required for a given reconstruction is based on performance and problem size. There are many ways to describe performance and energy efficiency, thus this work will investigate multiple metrics including performance-per-watt, energy-delay product, and energy consumption. This work found that irregular GPU-based approaches1 realized tremendous savings in energy consumption when compared to CPU implementations while also significantly improving the performance-per- watt and energy-delay product metrics. Additional energy savings and other metric improvement was realized on the GPU-based reconstructions by improving storage I/O by implementing a parallel MIMD-like modularization of the compute and I/O tasks.

Optimizing graph queries with graph joins and Sprinkle SPARQL

Big data problems are often more akin to sparse graphs rather than relational tables. As such we argue that graph-based physical representations provide advantages in terms of both size and speed for executing queries. Drawing from research in sparse matrices, we use a compressed sparse row (CSR) format to model graph-oriented data. We also present two novel mechanisms for exploiting the CSR format that both find optimal join strategies and also prune variable bindings before expensive join operations occur. The first tactic we call Sprinkle SPARQL, which takes triple patterns of SPARQL queries and performs low-cost, linear-time set intersections to produce a constrained list of variable bindings for each variable in a query. Besides constrained lists of variable bindings, Sprinkle SPARQL also produces metrics that are consumed by the join algorithm to select an optimal execution path. The second tactic, graph joins, utilizes the CSR data structure as an index to efficiently join two variables expressed in a triple pattern together. We evaluate our approach on two data sets with over a billion edges: LUBM(8000) and an R-MAT graph generated with Graph5001 parameters and extended to have edge labels.