Processing large amounts of data, especially in learning algorithms, poses a challenge for current embedded computing systems. Hyperdimensional (HD) computing (HDC) is a brain-inspired computing paradigm that works with high-dimensional vectors called hypervectors . HDC replaces several complex learning computations with bitwise and simpler arithmetic operations at the expense of an increased amount of data due to mapping the data into high-dimensional space. These hypervectors, more often than not, cannot be stored in memory, resulting in long data transfers from storage. In this article, we propose Store-n-Learn, an in-storage computing solution that performs HDC classification and clustering by implementing encoding, training, retraining, and inference across the flash hierarchy. To hide the latency of training and enable efficient computation, we introduce the concept of batching in HDC. We also present on-chip acceleration for HDC encoding in flash planes. This enables us to exploit the high parallelism provided by the flash hierarchy and encode multiple data points in parallel in both batched and non-batched fashion. Store-n-Learn also implements a single top-level FPGA accelerator with novel implementations for HDC classification training, retraining, inference, and clustering on the encoded data. Our evaluation over 10 popular datasets shows that Store-n-Learn is on average 222× (543×) faster than CPU and 10.6× (7.3×) faster than the state-of-the-art in-storage computing solution, INSIDER for HDC classification (clustering).
more »
« less
Multi-probe random projection clustering to secure very large distributed datasets
This paper presents a solution to the approximate k-means clustering problem for very large distributed datasets. Distributed data models have gained popularity in recent years following the efforts of commercial, academic and government organizations, to make data more widely accessible. Due to the sheer volume of available data, in-memory single-core computation quickly becomes infeasible, requiring distributed multi-processing. Our solution achieves comparable clustering performance to other popular clustering algorithms, with improved overall complexity growth while being amenable to distributed processing frameworks such as Map-Reduce. Our solution also maintains certain guarantees regarding data privacy deanonimization.
more »
« less
- Award ID(s):
- 1440420
- NSF-PAR ID:
- 10193710
- Date Published:
- Journal Name:
- 2015 IEEE International Conference on Big Data
- Page Range / eLocation ID:
- 1891 to 1900
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Computational science today depends on complex, data-intensive applications operating on datasets from a variety of scientific instruments. A major challenge is the integration of data into the scientist's workflow. Recent advances in dynamic, networked cloud resources provide the building blocks to construct reconfigurable, end-to-end infrastructure that can increase scientific productivity. However, applications have not adequately taken advantage of these advanced capabilities. In this work, we have developed a novel network-centric platform that enables high-performance, adaptive data flows and coordinated access to distributed cloud resources and data repositories for atmospheric scientists. We demonstrate the effectiveness of our approach by evaluating time-critical, adaptive weather sensing workflows, which utilize advanced networked infrastructure to ingest live weather data from radars and compute data products used for timely response to weather events. The workflows are orchestrated by the Pegasus workflow management system and were chosen because of their diverse resource requirements. We show that our approach results in timely processing of Nowcast workflows under different infrastructure configurations and network conditions. We also show how workflow task clustering choices affect throughput of an ensemble of Nowcast workflows with improved turnaround times. Additionally, we find that using our network-centric platform powered by advanced layer2 networking techniques results in faster, more reliable data throughput, makes cloud resources easier to provision, and the workflows easier to configure for operational use and automation.more » « less
-
In the era of big data and cloud computing, large amounts of data are generated from user applications and need to be processed in the datacenter. Data-parallel computing frameworks, such as Apache Spark, are widely used to perform such data processing at scale. Specifically, Spark leverages distributed memory to cache the intermediate results, represented as Resilient Distributed Datasets (RDDs). This gives Spark an advantage over other parallel frameworks for implementations of iterative machine learning and data mining algorithms, by avoiding repeated computation or hard disk accesses to retrieve RDDs. By default, caching decisions are left at the programmer’s discretion, and the LRU policy is used for evicting RDDs when the cache is full. However, when the objective is to minimize total work, LRU is woefully inadequate, leading to arbitrarily suboptimal caching decisions. In this paper, we design an algorithm for multi-stage big data processing platforms to adaptively determine and cache the most valuable intermediate datasets that can be reused in the future. Our solution automates the decision of which RDDs to cache: this amounts to identifying nodes in a direct acyclic graph (DAG) representing computations whose outputs should persist in the memory. Our experiment results show that our proposed cache optimization solution can improve the performance of machine learning applications on Spark decreasing the total work to recompute RDDs by 12%.more » « less
-
Clustering streaming data has gained importance in recent years due to an expanding opportunity to discover knowledge in widely available data streams. As streams are potentially evolving and unbounded sequence of data objects, clustering algorithms capable of performing fast and incremental processing of data points are necessary. This paper presents a method of clustering high-dimensional data streams using approximate methods called streamingRPHash. streamingRPHash combines random projections with locality-sensitivity hashing to construct a high-performance clustering method. streamingRPHash is amenable to distributed processing frameworks such as Map-Reduce, and also has the benefits of constrained overall complexity growth. This paper describes streamingRPHash algorithm and its various configurations. The clustering performance of streamingRPHash is compared to several alternatives. Experimental results show that streamingRPHash has comparable clustering accuracy and substantially lower runtime and memory usage.more » « less
-
Speaker diarization determines who spoke and when? in an audio stream. In this study, we propose a model-based approach for robust speaker clustering using i-vectors. The i-vectors extracted from different segments of same speaker are correlated. We model this correlation with a Markov Random Field (MRF) network. Leveraging the advancements in MRF modeling, we used Toeplitz Inverse Covariance (TIC) matrix to represent the MRF correlation network for each speaker. This approaches captures the sequential structure of i-vectors (or equivalent speaker turns) belonging to same speaker in an audio stream. A variant of standard Expectation Maximization (EM) algorithm is adopted for deriving closed-form solution using dynamic programming (DP) and the alternating direction method of multiplier (ADMM). Our diarization system has four steps: (1) ground-truth segmentation; (2) i-vector extraction; (3) post-processing (mean subtraction, principal component analysis, and length-normalization) ; and (4) proposed speaker clustering. We employ cosine K-means and movMF speaker clustering as baseline approaches. Our evaluation data is derived from: (i) CRSS-PLTL corpus, and (ii) two meetings subset of the AMI corpus. Relative reduction in diarization error rate (DER) for CRSS-PLTL corpus is 43.22% using the proposed advancements as compared to baseline. For AMI meetings IS1000a and IS1003b, relative DER reduction is 29.37% and 9.21%, respectively.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1109/BigData.2015.7363964
