A Unified Semantic Embedding: Relating Taxonomies and Attributes
Authors: Sung Ju Hwang, Leonid Sigal
NeurIPS 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our method for multiclass categorization performance on two different datasets generated from a public image collection, and also test for knowledge transfer on few-shot learning. |
| Researcher Affiliation | Industry | Sung Ju Hwang Disney Research Pittsburgh, PA sungju.hwang@disneyresearch.com Leonid Sigal Disney Research Pittsburgh, PA lsigal@disneyresearch.com |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide concrete access to its own source code. It only mentions using code provided by authors of a baseline method (NCM [11]) in a footnote. |
| Open Datasets | Yes | We use Animals with Attributes dataset [1], which consists of 30, 475 images of 50 animal classes, with 85 class-level attributes. |
| Dataset Splits | Yes | Since there is no fixed training/test split, we use {30,30,30} random split for training/validation/test. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running its experiments, only general mentions of features derived from deep convolutional networks. |
| Software Dependencies | No | The paper mentions the use of 'De CAF features [18]' but does not provide specific version numbers for any software dependencies used in their own experimental setup. |
| Experiment Setup | Yes | For parameters, the projection dimension de = 50 for all our models. For other parameters, we find the optimal value by cross-validation on the validation set. We set µ1 = 1 that balances the main and auxiliary task equally, and search for µ2 for discriminative/generative tradeoff, in the range of {0.01, 0.1, 0.2 . . . , 1, 10}, and set ℓ-2 norm regularization parameter λ = 1. For sparsity parameter γ1, we set it to select on average several (3 or 4) attributes per class, and for disjoint parameter γ2, we use 10γ1, without tuning for performance. |