The Balanced-Pairwise-Affinities Feature Transform
Authors: Daniel Shalam, Simon Korman
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, the transform is highly effective and flexible in its use and consistently improves networks it is inserted into, in a variety of tasks and training schemes. We demonstrate state-of-the-art results in few-shot classification, unsupervised image clustering and person re-identification. |
| Researcher Affiliation | Academia | 1 Department of Computer Science, University of Haifa, Israel. |
| Pseudocode | Yes | Algorithm 1 BPA transform on a set of n features. |
| Open Source Code | Yes | Code is available at github.com/Daniel Shalam/BPA . |
| Open Datasets | Yes | Mini Imagenet (Vinyals et al., 2016) and CIFAR-FS (Bertinetto et al., 2019), with detailed results in Tables 2 and 3 respectively. We evaluate the performance of the proposed BPA, applying it to a variety of FSC methods including the recent state-of-the-art (PTMap (Hu et al., 2020), Sill Net (Zhang et al., 2021), PTMap-SF (Chen & Wang, 2021) and PMF (Hu et al., 2022)) as well as to conventional methods like the popular Proto Net (Snell et al., 2017). ... We experiment on 3 standard datasets, STL-10 (Coates et al., 2011), CIFAR-10 and CIFAR-100-20 (Krizhevsky & Hinton, 2009), ... and tested on the large-scale Re ID benchmarks CUHK03 (Li et al., 2014) (both detected and labeled ) as well as the Market-1501 (Zheng et al., 2015) set |
| Dataset Splits | Yes | We measured accuracy on the validation set of Mini Imagenet (Vinyals et al., 2016), using Proto Net-BPAp (which is the non-fine-tuned drop-in version of BPA within Proto Net (Snell et al., 2017)). |
| Hardware Specification | No | The paper does not specify any particular hardware components such as GPU models (e.g., 'NVIDIA A100'), CPU models, or cloud computing instance types used for running the experiments. |
| Software Dependencies | No | The paper mentions 'Py Torch-style pseudocode' (Algorithm 1) which implies the use of PyTorch, but it does not specify a version number for PyTorch or any other software dependencies. |
| Experiment Setup | Yes | BPA has two hyper-parameters that were chosen through cross-validation and kept fixed for each application over all datasets. The number of Sinkhorn iterations for computing the optimal transport plan was fixed to 5 and entropy regularization parameter λ (Eq. (3.1)) was set to 0.1 for UIC and FSC and to 0.25 for Re ID. |