Attentive Recurrent Comparators
Authors: Pranav Shyam, Shubham Gupta, Ambedkar Dukkipati
ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We first test ARCs across many tasks that require assessment of visual similarity. We find that ARCs that do not use any convolutions show comparable performance to Deep Convolutional Neural Networks on challenging datasets like CASIA Web Face and Omniglot. Though dense ARCs are as capable as Conv Nets, a combination of both ARCs and convolutions (Conv ARCs) produces much more superior models. In the task of estimating the similarity of two characters from the Omniglot dataset, ARCs and Deep Conv Nets both achieve about 93.4% accuracy, whereas Conv ARCs achieve 96.10% accuracy. In the task of face verification on the CASIA Webface dataset, Conv ARCs achieved 81.73% accuracy surpassing the 79.48% accuracy achieved by a CNN baseline considered. |
| Researcher Affiliation | Academia | 1Department of Computer Science and Engineering, Rashtreeya Vidyalaya College of Engineering, Bengaluru, India 2Department of Computer Science and Automation, Indian Institute of Science, Bengaluru, India. |
| Pseudocode | No | No pseudocode or clearly labeled algorithm blocks were found in the paper. |
| Open Source Code | No | The paper does not provide any explicit statements about releasing source code for their methodology, nor does it include a link to a code repository. |
| Open Datasets | Yes | Omniglot is a dataset by (Lake et al., 2015) that is specially designed to compare and contrast the learning abilities of humans and machines. |
| Dataset Splits | Yes | The data split up of the Omniglot dataset used for this comparison is different from the above: 30 alphabets were used for training, 10 for validation and 10 for testing (this was in order to be comparable to the Conv Nets in (Koch et al.)). ... We split the data as follows: Training set: 70% (7402 people), validation set: 15% (1586 people) and Test set: 15% (1587 people). |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments. |
| Software Dependencies | No | The paper describes the model architecture and components (e.g., LSTM, Wide Resnets) but does not list specific software dependencies with version numbers (e.g., 'PyTorch 1.9', 'TensorFlow 2.x'). |
| Experiment Setup | Yes | The number of glimpses per image was fixed to 8, thus making the total number of recurrent steps 16. 32x32 greyscale images of characters were used and the attention glimpse resolution of 4x4 was used. We used moderate data augmentation consisting of translation, flipping, rotation and shearing. |