Deep and Wide Multiscale Recursive Networks for Robust Image Labeling
Authors: Gary B. Huang; Viren Jain
ICLR 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments in the challenging domain of connectomic reconstruction of neural circuity from 3d electron microscopy data show that these Deep And Wide Multiscale Recursive (DAWMR) networks lead to new levels of image labeling performance. We performed detailed experiments in the domain of boundary prediction in electron microscopy images of neural tissue. |
| Researcher Affiliation | Academia | Gary B. Huang and Viren Jain Janelia Farm Research Campus Howard Hughes Medical Institute 19700 Helix Drive, Ashburn, VA, USA {huangg, jainv}@janelia.hhmi.org |
| Pseudocode | No | The paper describes the methods textually and with diagrams, but does not include pseudocode or explicitly labeled algorithm blocks. |
| Open Source Code | Yes | Experiments were run using our parallel computing software package, available online1; for more details see Section C.1. 1http://sites.google.com/site/dawmrlib/ |
| Open Datasets | No | Two image volumes were obtained using the above acquisition process. The first volume was used for training and the second for both validation and testing. Initially, human annotators densely labeled subvolumes from both images. |
| Dataset Splits | Yes | These labels form a preliminary training set, referred to in the sequel as the small training set (5.2 million labels), and the validation set (16 million labels), respectively. Afterward, an interactive procedure was used wherein human annotators proofread machinegenerated segmentations by visually examining dense reconstructions within small subvolumes and correcting any mistakes. The proofread annotations were then added to the small training set and validation set to form the full training set (120 million labels) and test set (46 million labels), respectively. |
| Hardware Specification | No | To achieve a reasonable training time, in this paper we assume access to both graphics processing units (GPUs) and multi-core CPUs or cluster computing environments. A CPU cluster is used to pre-compute feature vectors prior to GPU-based training of a supervised classifier. Supervised learning is performed on a single GPU-equipped machine. The paper mentions types of hardware but lacks specific model numbers for CPUs or GPUs, or detailed specifications of the cluster. |
| Software Dependencies | No | An open source Matlab/C software package that implements DAWMR networks is available online: http://sites.google.com/site/dawmrlib/. The paper specifies the programming languages but does not provide version numbers for Matlab, C, or any specific libraries. |
| Experiment Setup | Yes | Unless explicitly stated otherwise, our experiments use the following set-up: the feature extraction modules produce a feature representation of dimension hu l = 8000, individual filters use 3d 53 patches, and classification is performed using an MLP with a single hidden layer with 200 hidden units and trained with a balanced sampling of positive and negative training examples. The multilayer perceptrons in DAWMR architectures were trained using minibatch sizes of 40 with a balanced sampling of positive and negative edges. Learning rates were set to 0.02. We used sigmoid output units and rectified linear units in the hidden layer. For networks trained with dropout regularization, the drop-out rate was set to 0.5 for the hidden layer and 0 for the input layer. We performed 5e5 updates. Optimization was performing using a cross-entropy loss function. To regularize and prevent overfitting, we used an inverse margin of 0.1, meaning that target labels were set to 0.1/0.9 rather than 0/1, penalizing over-confident predictions. |