Transductive Decoupled Variational Inference for Few-Shot Classification

Authors: Anuj Rajeeva Singh, Hadi Jamali-Rad

TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experimental results corroborate the efficacy of TRIDENT and demonstrate that, using the simplest of backbones and a meta-learning strategy, it sets a new state-of-the-art in the most commonly adopted datasets mini Image Net and tiered Image Net (offering up to 4% and 5% improvements, respectively), as well as for the recent challenging cross-domain mini Imagenet CUB scenario offering a significant margin (up to 20% improvement) beyond the best existing baselines.
Researcher Affiliation Collaboration 1Delft University of Technology, The Netherlands 2Shell Global Solutions International B.V., Amsterdam, The Netherlands
Pseudocode Yes Algorithm 1: TRIDENT Algorithm 2: End to End Meta-Training of TRIDENT
Open Source Code Yes 1Codebase available at https://github.com/anujinho/trident.
Open Datasets Yes We evaluate TRIDENT on the three most commonly adopted datasets: mini Imagenet (Ravi & Larochelle, 2017a), tiered Imagenet (Ren et al., 2018) and CUB (Welinder et al., 2010).
Dataset Splits Yes We follow the predominantly adopted settings of Ravi & Larochelle (2017a); Chen et al. (2019) where we split the entire dataset into 64 classes for training, 16 for validation and 20 for testing. tiered Imagenet is a larger subset of Image Net with 608 classes and 779, 165 total images, which are grouped into 34 higher-level nodes in the Image Net human-curated hierarchy. This set of nodes is partitioned into 20, 6, and 8 disjoint sets of training, validation, and testing nodes, and the corresponding classes form the respective meta-sets. CUB (Welinder et al., 2010) dataset has a total of 200 classes, split into training, validation and test sets following Chen et al. (2019).
Hardware Specification Yes This translates to an average training time of 110 hours on an 11GB NVIDIA 1080Ti GPU.
Software Dependencies No We use Py Torch (Paszke et al., 2019) and learn2learn (Arnold et al., 2020) for all our implementations. Specific version numbers for these software components are not provided.
Experiment Setup Yes The hyperparameter values (H.P.) used for training TRIDENT on mini Imagenet and tiered Imagenet are shown in Table 1. We apply the same hyperparameters for the cross-domain testing scenario of mini Imagenet CUB used for training TRIDENT on mini Imagenet, for the given (N-way, K-shot) configuration. Hyperparameters are kept fixed throughout training, validation and testing for a given configuration. Adam (Kingma & Ba, 2015) optimizer is used for inner and meta-updates.