The Role of Minimal Complexity Functions in Unsupervised Learning of Semantic Mappings
Authors: Tomer Galanti, Lior Wolf, Sagie Benaim
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Various predictions are made based on the hypothesis that semantic alignment can be captured by the minimal mapping. These are verified extensively. In addition, a new mapping algorithm is proposed and shown to lead to better mapping results. (Abstract); 5 EXPERIMENTS The first group of experiments is dedicated to test the validity of the three predictions made, in order to give further support to the simplicity hypothesis. Next, we evaluate the success of the proposed algorithm in comparison to the Disco GAN method of Kim et al. (2017). |
| Researcher Affiliation | Collaboration | Tomer Galanti The Blavatnik School of Computer Science Tel Aviv University Tel Aviv, Israel tomerga2@post.tau.ac.il Lior Wolf Facebook AI Research & The Blavatnik School of Computer Science Tel Aviv University Tel Aviv, Israel wolf@fb.com wolf@cs.tau.ac.il Sagie Benaim The Blavatnik School of Computer Science Tel Aviv University Tel Aviv, Israel sagieb@mail.tau.ac.il |
| Pseudocode | Yes | Algorithm 1 Complexity Based Regularization Alignment Require: Unlabeled training sets SA i.i.d Dm A and SB i.i.d Dn B, a desired complexity k2, and a trade-off parameter λ 1: Identify a complexity k1, which leads to a small discrepancy min g s.t: C(g)=k1disc(g DA, DB). 2: Train g of complexity k1 to minimize disc(g DA, DB). 3: Train h of complexity k2 to minimize disc(h DA, DB) + λRDA[h, g]. |
| Open Source Code | No | The paper mentions using the 'public implementation' of Disco GAN (Kim et al., 2017) in Section 5.1, but does not state that the code for *their* proposed new method or experiments is publicly available or provide a link. |
| Open Datasets | Yes | The first set of experiments considers the Celeb A face dataset. (Section 5.1); In the case of mapping handbags to shoes... (Section 5.1); in the case of edges to shoes... (Section 5.1); We chose to experiment with the Disco GAN architecture since it focuses on semantic tasks that contain a lesser component of texture or style transfer. (Section 5) |
| Dataset Splits | No | The paper mentions 'training images' and 'training samples' and 'test images' but does not specify a clear training/validation/test split or percentages used for each, nor does it explicitly mention a validation set. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments. |
| Software Dependencies | No | The paper mentions various architectures (Disco GAN, Cycle GAN, U-net) and activation functions (Leaky ReLU, ReLU, Sigmoid), and mentions VGG face descriptors, but it does not specify any software names with version numbers (e.g., PyTorch, TensorFlow, or specific library versions) that were used for the experiments. |
| Experiment Setup | Yes | In (Kim et al., 2017), 8 or 10 layers are employed (counting both convolution and deconvolution) depending on the experiment. In our experiment, we vary the number of layers and inspect the influence on the results. (Section 5.1); In Disco GAN, the generator is built of: (i) an encoder consisting of convolutional layers with 4x4 filters followed by Leaky Re LU activation units and (ii) a decoder consisting of deconvolutional layers with 4x4 filters followed by a Re LU activation units. Sigmoid is used for the output layer. Between four to five convolutional/deconvolutional layers are used, depending on the domains used in A and B (we match the published code architecture per dataset). The discriminator is similar to the encoder, but has an additional convolutional layer as the first layer and a sigmoid output unit. (Section 5.1); Specifically, we use k1 = 6, which, as discussed above, typically leads to a low (but not very low) discrepancy, while the alignment is still unambiguous. (Section 5.2) |