Learning by Stretching Deep Networks

Authors: Gaurav Pandey, Ambedkar Dukkipati

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experimental results suggest that the proposed stretched deep convolutional networks are capable of achieving good performance for many object recognition tasks. More importantly, for a fixed network architecture, one can achieve much better accuracy using stretching rather than learning the weights using backpropagation. 6. Experimental results
Researcher Affiliation Academia Gaurav Pandey GP88@CSA.IISC.ERNET.IN Ambedkar Dukkipati AD@CSA.IISC.ERNET.IN Department of Computer Science and Automation Indian Institute of Science, Bangalore-560012, India
Pseudocode Yes Algorithm 1 Iterative computation of the convolved kernel matrix
Open Source Code No The paper does not provide an explicit statement about releasing source code or a direct link to a code repository for the described methodology.
Open Datasets Yes MNIST (Le Cun et al., 1998) is a standard dataset for character recognition with 50000 training and 10000 test samples of digits ranging from 0 to 9. The Caltech-101 dataset (Fei-Fei et al., 2007) consists of pictures of objects belonging to 101 categories with about 40 to 800 objects per categories. STL-10 dataset (Coates et al., 2011) is an image recognition dataset with 10 classes and 500 training and 800 test images per class.
Dataset Splits No For MNIST, the paper mentions '50000 training and 10000 test samples' but no explicit validation split. For Caltech-101 and STL-10, it similarly only specifies training and testing samples without mentioning a validation set.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers).
Experiment Setup Yes The weight matrix used for the stretched and the unstretched model is the same and is obtained after training the model for 5 epochs ( 150s). We randomly extract 9 9 patches from the images and learn a weight matrix with 64 weight vectors from these patches using an Re Lu RBM. The weight matrix is then stretched by multiplication with a random matrix of size 64 64. We use average pooling with a 10 10 boxcar filter and 5 5 down-sampling.