Stable and Efficient Representation Learning with Nonnegativity Constraints

Authors: Tsung-Han Lin, H. T. Kung

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

Reproducibility Variable Result LLM Response
Research Type Experimental In our experiments, we use a multi-layer deep architecture for representation learning, where we use K-means for feature learning and NOMP for representation encoding.
Researcher Affiliation Academia Tsung-Han Lin THLIN@EECS.HARVARD.EDU H. T. Kung KUNG@HARVARD.EDU School of Engineering and Applied Sciences, Harvard University, Cambridge MA, USA
Pseudocode Yes NOMP iterates the following steps for up to k rounds: 1. Initialize the residual vector r(0) = x and round number l = 1. Select the atom dil that has the highest positive correlation with the residual, il = arg maxi di, r(l 1) . Terminate if the largest correlation is less than or equal to zero. 2. Approximate the coefficients of the selected atoms by nonnegative least squares. z(l) = arg minz x Pl h=1 dihzih 2 s.t. zih 0 3. Compute the new residual r(l) = x Dz(l). Increment l by 1.
Open Source Code No The paper does not provide concrete access to source code for the methodology described, nor does it contain explicit statements about code release or links to repositories.
Open Datasets Yes We use a dictionary learned from 6 6 images patches from CIFAR-10. We first evaluate NOMP using the full CIFAR-10 dataset... The CIFAR-100 dataset is one of such datasets... Finally, we evaluate NOMP on the STL-10 dataset...
Dataset Splits Yes We choose k as 20 and 5 for NOMP and OMP, respectively, in the experiments.9 This choice is cross-validated over k = {1, 3, 5, 10, 20}. We report accuracy from the 5-fold cross validation on the training set.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory, or cluster specifications) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers required to replicate the experiment.
Experiment Setup Yes Standard preprocessing steps are applied on image data to generate data vectors for layer-1. These include mean subtraction, contrast normalization , and ZCA-whitening, followed by sign-splitting as described in Section 3.2. We encode 6 6 patches, pool the sparse codes over the four quadrants of an image, and concatenate the four representations. We choose k as 20 and 5 for NOMP and OMP, respectively, in the experiments. The patch sizes and features sizes are 6 6, 9 9, 15 15, and 3200, 6400, 6400 for the three layers, respectively. we employ a linear classifier (L2-SVM).