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 and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].

Advancing Out-of-Distribution Detection via Local Neuroplasticity

Authors: Alessandro Canevaro, Julian Schmidt, Sajad Marvi, Hang Yu, Georg Martius, Julian Jordan

ICLR 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We validate our approach on benchmarks from image and medical domains, demonstrating superior performance and robustness compared to state-of-the-art techniques.
Researcher Affiliation Collaboration 1Mercedes-Benz AG, Sindelfingen, Germany 2University of T ubingen, T ubingen, Germany
Pseudocode No The paper describes the step-by-step procedure of the method in Section 2.2 using numbered paragraphs and mathematical formulas, but it does not present it in a formal pseudocode block or algorithm environment.
Open Source Code Yes Our code is publicly available at the following link: https: //github.com/alessandro-canevaro/KAN-OOD.
Open Datasets Yes We tested our method on seven different benchmarks from two different domains: the Open OOD CIFAR10, CIFAR-100, Image Net-200 full-spectrum (FS), and Image Net-1K FS (Yang et al., 2022) for the image domain, and the Ethnicity, Age, and Synthetic OOD benchmarks for the tabular medical data domain (Azizmalayeri et al., 2023).
Dataset Splits Yes We evaluate the KAN detector on the CIFAR-10 benchmark, using CIFAR10 (Krizhevsky et al., b) as the In D dataset. The OOD datasets are categorized into near OOD datasets (CIFAR-100 (Krizhevsky et al., a) and Tiny Image Net (TIN) (Le & Yang, 2015)) and far OOD datasets (MNIST (Deng, 2012), SVHN (Netzer et al., 2011), Textures (Cimpoi et al., 2014), and Places365 (Zhou et al., 2018)).
Hardware Specification Yes All experiments are performed on a single NVIDIA Ge Force GTX 1080Ti GPU. For testing larger models and accelerating the hyperparameter optimization, we leveraged a cloud computing platform with an NVIDIA A100 GPU.
Software Dependencies Yes We used Python version 3.10 together with Py Torch 2.3.1 as the deep learning framework and leveraged CUDA 11.8 for GPU acceleration.
Experiment Setup Yes In our experiments, we use a learning rate of 0.1 and limit the training to a single epoch.