LASS: A Simple Assignment Model with Laplacian Smoothing
Authors: Miguel Carreira-Perpinan, Weiran Wang
AAAI 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We test LASS in: (1) a single-label task (digit recognition), where using assignment vectors is not strictly necessary, but LASS still achieves the best classification performance. (2) Document categorization and image tagging, where learning assignment vectors and using partial labels is necessary. ... We randomly sample 10 000 MNIST digit images ... Fig. 1(left) shows the classification error over 20 different labeled/unlabeled partitions of the dataset as a function of Nl (errorbars not shown to avoid clutter). |
| Researcher Affiliation | Academia | Miguel A. Carreira-Perpi n an and Weiran Wang Electrical Engineering and Computer Science, School of Engineering, University of California, Merced |
| Pseudocode | Yes | The final algorithm is as follows. The input are the affinity matrices GN K and WN N, from which we construct the graph Laplacian LN N. We then choose ρ > 0 and set K 1N RRT = 2λL + ρI. The Cholesky factor R is used to solve linear system (3b). We then iterate, in order, eqs. (3a) (3d) until convergence: K (Y U)1K h (3a) Z (2λL + ρI) 1(ρ(Y U) + G ν1T K) (3b) Y (Z + U)+ (3c) U U + Z Y (3d) |
| Open Source Code | No | No statement regarding the release of open-source code or a link to a code repository was found. |
| Open Datasets | Yes | We randomly sample 10 000 MNIST digit images (http://yann.lecun.com/ exdb/mnist) and compute the 10-nearest-neighbor graph with similarities based on the Euclidean distance between images. |
| Dataset Splits | No | The paper describes using labeled/unlabeled partitions and a test set but does not explicitly specify exact train/validation/test splits (e.g., percentages or sample counts) for reproducibility or use the term 'validation' for a distinct split. |
| Hardware Specification | No | As a sample runtime, for a problem with N = 10 000 items and K = 10 categories (i.e., Z has 105 parameters) and using a 100-nearest-neighbor graph, the Cholesky factorization takes 0.5 s and each iteration takes 0.15 s in a PC. This is too vague for specific hardware. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software dependencies (e.g., libraries, frameworks, or specific solvers). |
| Experiment Setup | No | We let all algorithms use their respective optimal parameters (e.g. λ in LASS is determined by a grid search). ... hinge loss penalty parameter C selected from {10 3, 10 2, 10 1, 1, 101, 102, 103} and SSL (Zhu, Ghahramani, and Lafferty 2003) using 1-out-of-10 coding for the labeled points. While parameters are mentioned, specific values for the optimal settings or full training setup details are not provided in the main text. |