Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]

Zero-Shot Natural Language Explanations

Authors: Fawaz Sammani, Nikos Deligiannis

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on 38 vision models, including both CNNs and Transformers. Our method outperforms supervised baselines on many metrics, while remaining comparable on others.
Researcher Affiliation Collaboration Fawaz Sammani & Nikos Deligiannis ETRO Department, Vrije Universiteit Brussel, Pleinlaan 2, B-1050 Brussels, Belgium imec, Kapeldreef 75, B-3001 Leuven, Belgium
Pseudocode No The paper describes the methods in prose and with diagrams (Figure 3, Figure 4), but does not contain explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement about releasing its own source code for the described methodology, nor does it provide a direct link to a code repository. It mentions using third-party libraries and models.
Open Datasets Yes We use the challenging Image Net-1K dataset as our benchmark... Image Net-X (Sammani & Deligiannis, 2023)... COCO dataset (Lin et al., 2014)... Places365 dataset (Zhou et al., 2017)... DTD dataset (Cimpoi et al., 2014)
Dataset Splits Yes We use the challenging Image Net-1K dataset as our benchmark, splitting its 1,000 classes into 900 for training and 100 for testing. For validation and hyperparameter tuning, we use 100 non-overlapping classes from the Image Net-21K dataset. ... Image Net-X (Sammani & Deligiannis, 2023) dataset ... It consists of 141K training samples, 2K for validation and 1K for testing. ... We report results on the common Karpathy test split benchmark using various vision classifiers.
Hardware Specification Yes On a single RTX3090 GPU, it takes roughly 10 seconds to train.
Software Dependencies No The paper mentions software components like 'Stable Diffusion v1.5 model', 'Hugging Face Diffusers library', 'smallest GPT-2 (Radford et al., 2019) model', 'Adam Optimizer (Kingma & Ba, 2015)', 'Sentence Transformers library', 'torchvision library', 'timm library', and 'huggingface library'. However, it does not provide specific version numbers for general ancillary software such as Python, PyTorch, or CUDA libraries.
Experiment Setup Yes The MLP is trained with full batch gradient descent for 2500 epochs using the Adam Optimizer (Kingma & Ba, 2015) with a learning rate of 5e-3 and a cosine annealing schedule (Loshchilov & Hutter, 2017). ... The number of learnable prefixes is set to 5 in each attention block of the GPT-2 model. ... We train the prefixes with the Adam optimizer ... with a learning rate of 0.01 and a weight decay of 0.3 with a cosine annealing schedule for I = 20 iterations. ... The number of K tokens sampled at each timestep is set to 512. We use a maximum NLE length of 20. ... We add the fluency loss from Tewel et al. (2021) with a weight of 0.8. ... We also prevent the generation of repeated n-grams of order 3 to avoid repetitive phrases, by setting their score to negative infinity... We also enforce a minimum sequence length of 10 tokens by setting the <.> token score to 0, in order to prevent premature termination.