Fine-grained Classes and How to Find Them
Authors: Matej Grcic, Artyom Gadetsky, Maria Brbic
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate FALCON on eight image classification tasks and a single-cell classification task. FALCON outperforms baselines by a large margin, achieving 22% improvement over the best baseline on the tiered Image Net dataset with over 600 fine-grained classes. |
| Researcher Affiliation | Academia | 1EPFL, Lausanne, Switzerland 2Faculty of Electrical Engineering and Computing, University of Zagreb, Croatia. Correspondence to: Maria Brbi c <mbrbic@epfl.ch>. |
| Pseudocode | Yes | Algorithm A1 shows the training procedure for simultaneous learning of a fine-grained classifier and class relationships. ... Algorithm E1 extends the Algorithm A1 to simultaneously train FALCON on multiple datasets. |
| Open Source Code | Yes | Our code is publicly available1. [Footnote 1: https://github.com/mlbio-epfl/falcon] |
| Open Datasets | Yes | We evaluate FALCON on eight image classification datasets including Living17, Nonliving26, Entity30, Entity13, tiered Image Net (Ren et al., 2018), CIFAR100 (Krizhevsky, 2009), CIFAR-SI, and CIFAR68 datasets. |
| Dataset Splits | Yes | For the tiered Image Net dataset(Ren et al., 2018), we joined training, validation and test taxonomies into a single dataset with 608 fine classes assigned across 34 coarse classes. ... The overview of all considered datasets is shown in Table 1. ... Table F1. Detailed overview of nine evaluation datasets. (includes VAL SAMPLES column) |
| Hardware Specification | Yes | We conducted all experiments using NVIDIA A100 GPUs. |
| Software Dependencies | Yes | The hyperparameter search was conducted using the TPE algorithm implemented with the Optuna framework (Akiba et al., 2019). We solve discrete optimization problem using Gurobi solver (Gurobi Optimization, LLC, 2023). |
| Experiment Setup | Yes | We train our method for 60 epochs with a batch size of 1024 across 2 GPUs for BREEDS datasets and for 90 epochs with batch size of 2048 across 4 GPUs for tiered Image Net. We use SGD with a momentum of 0.9 and no weight decay. The initial learning rate is set to 0.1 and annealed to 0.001 according to the cosine schedule with restarts every 30 epochs. ... Table G1. The selected hyperparameters. |