Unsupervised Transformation Learning via Convex Relaxations
Authors: Tatsunori B. Hashimoto, Percy S. Liang, John C. Duchi
NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the effectiveness of our sampling-based convex relaxation for learning transformations in two ways. In section 4.1, we check whether we can recover a known set of rotation / translation transformations applied to a downsampled celebrity face image dataset. Next, in section 4.2 we perform a qualitative evaluation of learning transformations over raw celebrity faces (Celeb A) and MNIST digits, following recent evaluations of disentangling in adversarial networks [2]. |
| Researcher Affiliation | Academia | Tatsunori B. Hashimoto John C. Duchi Percy Liang Stanford University Stanford, CA 94305 {thashim,jduchi,pliang}@cs.stanford.edu |
| Pseudocode | No | The paper describes algorithms through mathematical formulations and text, but does not include any clearly labeled "Pseudocode" or "Algorithm" blocks. |
| Open Source Code | Yes | Code, data, and experiments can be found on Codalab Worksheets (http://bit.ly/2Aj5tti). |
| Open Datasets | Yes | On the handwritten digits (MNIST) and celebrity faces (Celeb A) datasets, our method finds interpretable and disentangled transformations for handwritten digits, the thickness of lines and the size of loops in digits such as 0 and 9; and for celebrity faces, the degree of a smile. |
| Dataset Splits | No | The paper uses specific datasets like MNIST and Celeb A and mentions a "20,000 example subset" for MNIST, but it does not specify explicit training, validation, or test split percentages, sample counts, or refer to standard predefined splits for reproducibility. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU models, or memory specifications used for running its experiments. |
| Software Dependencies | No | The paper mentions optimization methods like "minibatch proximal gradient descent with Adagrad [8]", but it does not provide specific software names with version numbers for dependencies (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | Optimization for our methods and gradient descent use minibatch proximal gradient descent with Adagrad [8], where the proximal step for trace norm penalties use subsampling down to five thousand points and randomized SVD. On MNIST digits we trained a five-dimensional linear transformation model over a 20,000 example subset of the data, which took 10 minutes. |