A Multicore Path to Connectomics-on-Demand

Abstract

The current design trend in large scale machine learning is to use distributed clusters of CPUs and GPUs with MapReduce-style programming. Some have been led to believe that this type of horizontal scaling can reduce or even eliminate the need for traditional algorithm development, careful parallelization, and performance engineering. This paper is a case study showing the contrary: that the benefits of algorithms, parallelization, and performance engineering, can sometimes be so vast that it is possible to solve "clusterscale" problems on a single commodity multicore machine. Connectomics is an emerging area of neurobiology that uses cutting edge machine learning and image processing to extract brain connectivity graphs from electron microscopy images. It has long been assumed that the processing of connectomics data will require mass storage, farms of CPU/GPUs, and will take months (if not years) of processing time. We present a high-throughput connectomics-ondemand system that runs on a multicore machine with less than 100 cores and extracts connectomes at the terabyte per hour pace of modern electron microscopes. CCS Concepts • Computing methodologies → Concurrent algorithms; • Computer systems organization → Multicore architectures; • Computing methodologies → Neural networks; • Applied computing → Computational biology

Cross-Classification Clustering: An Efficient Multi-Object Tracking Technique for 3-D Instance Segmentation in Connectomics

“…Many of these techniques use the hierarchical approach of Andres et al [2] that employs the well-known hierarchical image segmentation framework [3,15,39,47]. This is still the most common approach in connectomics segmentation pipelines: first detecting object borders in 2-D/3-D and then gradually agglomerating information to form the final objects [6,9,20,27,31,34,35,52]. The elevation maps obtained from the border detectors are treated as estimators of the true border probabilities [9], which are used to define an over-segmentation of the image, foreground connected components on top of a background canvas.…”

“…Therefore it may need to be agglomerated with other connected components (heuristically [17,18,26] or based on learned weights of handcrafted features [2,27,40,41]), but it should not be broken down into smaller segments. Numerous 3-D reconstruction systems follow this bottom-up design [6,7,27,31,35,40,41,44]. A heavily engineered implementation of hierarchical segmentation [31] still occupies the leading entry in the (still active) classical SNEMI3D connectomics contest of 2013 [5], evaluated in terms of the uniform instance segmentation correctness metrics (normalized Rand-Error [53] and Variation of Information [36]).…”

“…In order to illustrate the accuracy of 3C, we submitted our result to the public SNEMI3D challenge website alongside two common baseline models, the 3-D watershed transform (a region-growing technique) and Neuroproof agglomeration [41]. Our watershed code was adopted from [35]. Similar to other traditional agglomerating-techniques, Neuroproof trains a random forest on merge decisions of neighboring objects [40,41,42].…”

“…For the first test, we used a lightweight yet successful FCN model [35] (Maxout) (for border and 3C computations), reconstructing the test set of [24] (S1) (3C with FOV of 109 pixels). Maxout is currently the fastest border detector in connectomics, which was previously successfully used for single-object tracking [37].…”

FlashR: R-Programmed Parallel and Scalable Machine Learning using SSDs

“…While conventional wisdom addresses large-scale data analysis and machine learning with clusters [6,33,7,1,32,14], recent works [35,37,16,17] demonstrate a single-machine solution can deal with large-scale data analysis efficiently in a multicore machine. The advance of solid-state drives (SSDs) allows us to tackle data analysis in a single machine efficiently at a larger scale with a cheaper price.…”