EnsLoss: Stochastic Calibrated Loss Ensembles for Preventing Overfitting in Classification
Authors: Ben Dai
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The numerical effectiveness of ENSLOSS compared to fixed loss methods is demonstrated through experiments on a broad range of 45 pairs of CIFAR10 datasets, the PCam image dataset, and 14 Open ML tabular datasets and with various deep learning architectures. Python repository and source code are available on GITHUB. |
| Researcher Affiliation | Academia | 1Department of Statistics, The Chinese University of Hong Kong. Correspondence to: Ben Dai <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 (Minibatch) Calibrated ensemble SGD. and Algorithm 2 Inverse Box-Cox transform of loss-derivatives. |
| Open Source Code | Yes | Python repository and source code are available on GITHUB. All Python codes is openly accessible at our GITHUB. |
| Open Datasets | Yes | Image datasets. We present the empirical results for image benchmark datasets: the CIFAR10 (Krizhevsky et al., 2009) and the Patch Camelyon (PCam; (Veeling et al., 2018))... Tabular datasets. We applied a filtering (n ≥ 1000, d ≤ 1000) across all Open ML (Vanschoren et al., 2014) binary classification dense datasets... |
| Dataset Splits | No | The paper frequently refers to 'training' and 'testing' datasets, such as 'significant gap often persisting between the training (close to zero) and testing errors', but does not explicitly provide specific percentages, counts, or methodology for dataset splits (e.g., 80/10/10 split, or specific sample counts for train/validation/test sets) for the datasets used. |
| Hardware Specification | No | The paper does not explicitly mention specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It refers to 'deep learning models' and 'neural network architectures' in a general sense. |
| Software Dependencies | No | The paper states 'All Python codes is openly accessible at our GITHUB.', indicating Python is used. However, it does not provide specific version numbers for Python or any other libraries, frameworks, or solvers utilized in the implementation. |
| Experiment Setup | No | The paper mentions that 'The implementation settings for each method are identical', and refers to a 'learning rate γ' and a 'minibatch size B' in Algorithm 1, as well as a hyperparameter 'λ = 0' for the Box-Cox transformation. Table 8 shows 'minimum epochs required for training accuracy to stabilize'. However, it does not provide specific numerical values for critical hyperparameters such as the learning rate, the exact batch size used for the experiments, or the total number of epochs for the main results presented in Section 4. |