Learning from Small Sample Sets by Combining Unsupervised Meta-Training with CNNs
Authors: Yu-Xiong Wang, Martial Hebert
NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The resulting CNNs significantly improve the performance in scene classification, fine-grained recognition, and action recognition with small training samples. 4 Experimental evaluation In this section, we explore the use of low-density separator networks (LDS+CNNs) on a number of supervised learning tasks with limited data, including scene classification, fine-grained recognition, and action recognition. |
| Researcher Affiliation | Academia | Yu-Xiong Wang Martial Hebert Robotics Institute, Carnegie Mellon University {yuxiongw, hebert}@cs.cmu.edu |
| Pseudocode | No | The paper describes various optimization procedures and steps (e.g., in sections 2.2 and 2.3) but does not present them in a formally labeled "Pseudocode" or "Algorithm" block. |
| Open Source Code | No | The paper mentions using third-party tools like Caffe [32] and Liblinear [37] but does not provide any statement or link indicating that the source code for the proposed methodology is openly available. |
| Open Datasets | Yes | We implement the unsupervised meta-training on Yahoo! Flickr Creative Commons100M dataset (YFCC100M) [26], which is the largest single publicly available image and video database. We evaluate on standard benchmark datasets for scene classification: SUN397 [33] and MIT-67 [34], fine-grained recognition: Oxford 102 Flowers [35], and action recognition (compositional semantic recognition): Stanford-40 actions [36]. |
| Dataset Splits | Yes | We follow the standard experimental setup (e.g., the train/test splits) for these datasets. |
| Hardware Specification | No | The paper mentions "NVIDIA for donating GPUs" and "AWS Cloud Credits for Research program", but does not specify exact GPU models, CPU models, or detailed cloud instance specifications used for experiments. |
| Software Dependencies | No | The paper mentions using "Caffe [32]" and "Liblinear [37]" but does not provide specific version numbers for these software components. |
| Experiment Setup | Yes | Implementation Details. During unsupervised meta-training, we use 99.2 million unlabeled images on YFCC100M [26]. After resizing the smallest side of each image to be 256, we generate the standard 10 crops (4 corners plus one center and their flips) of size 224 224 as implemented in Caffe [32]. For single-scale structures, we learn LDS in the fc7 activation space of dimension 4,096. [...] For learning LDS in Eqn. (2), η and λ1 are set to 1 and λ2 is set to normalize for the size of quasi-classes, which is the same setup and default parameters as in [23]. For generating high-density quasi-classes in Eqn. (3), following [31, 24], we set the minimum and maximum number of selected samples per quasi-classes to be τ0 =6 and τ =56, and produce C =30 quasi-classes in total. We use the same setup and parameters as in [24], where α=1, β =1. |