A representation-learning game for classes of prediction tasks
Authors: Neria Uzan, Nir Weinberger
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Example 6. We validated Algorithm 1 in the linear MSE setting (Section 3), for which a closed-form solution exists. We ran Algorithm 1 on randomly drawn diagonal Σx, and computed the ratio between the regret obtained by the algorithm to the theoretical value. The left panel of Figure 2 shows that the ratio is between 1.15 1.2 in a wide range of d values. [...] Example 9 (Comparison with PCA for multi-label classification). We constructed a dataset of images, each containing 4 shapes randomly selected from a dictionary of 6 shapes. [...] We ran the simplified version of Algorithm 1 on a dataset of 1000 images, and compared the cross-entropy loss and the accuracy of optimized representation to that of PCA on a fresh dataset of 1000 images. The results in Figure 3 show that the regret of PCA is much larger, not only for the worse-case function but also for the average-case function. |
| Researcher Affiliation | Academia | Neria Uzan Technion Israel Institute of Technology neriauzan@gmail.com Nir Weinberger Technion Israel Institute of Technology nirwein@technion.ac.il |
| Pseudocode | Yes | Algorithm 1 Solver of (22): An iterative algorithm for learning mixed representations. [...] Algorithm 2 A procedure for finding a new function via the solution of (23) [...] Algorithm 3 A procedure for finding a new representation R(k+1) via the solution of (24) |
| Open Source Code | Yes | The code for the experiments is available at this link. |
| Open Datasets | No | The paper constructs its own datasets for examples (e.g., 'We constructed a dataset of images', 'We drawn empirical distributions of features from an isotropic normal distribution', 'randomly drawn diagonal Σx') but does not provide concrete access information (link, DOI, or citation) for them to be considered publicly available. |
| Dataset Splits | No | The paper uses datasets for experiments (e.g., 'a dataset of 1000 images' and 'a fresh dataset of 1000 images' in Example 9) but does not provide specific details on training, validation, and test splits (percentages, counts, or methodology for creating them). |
| Hardware Specification | Yes | Table 1: Hardware details CPU RAM GPU Intel i9 13900k 64GB RTX 3090 Ti |
| Software Dependencies | No | The paper states 'The code for the experiments was written in Python 3.6 and is available at this link. The optimization of hyperparameters was done using the Optuna library.' While Python version is given, no version is provided for the Optuna library or other potentially critical software dependencies like deep learning frameworks, making full reproducibility of the software environment difficult. |
| Experiment Setup | Yes | The algorithm parameters used for Example 6 are shown in Table 2. The parameters were optimally tuned for σ0 = 1. [...] Table 2: Parameters for linear MSE setting example Parameter βr βf ηr ηf Value 0.94 0.653 0.713 0.944 Parameter TR Tf Tavg Tstop Value 100 until convergence 10 80 |