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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
ConceptScope: Characterizing Dataset Bias via Disentangled Visual Concepts
Authors: Jinho Choi, Hyesu Lim, Steffen Schneider, Jaegul Choo
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we show that our method captures a broad range of visual concepts, including objects, textures, backgrounds, facial attributes, emotions, and actions. We further demonstrate that the spatial attributions derived from SAEs aligned closely with ground-truth segmentation masks. Moreover, Concept Scope reliably identifies known dataset biases (e.g., background bias in Waterbirds [59]) and uncovers previously unannotated ones (e.g., cultural biases or co-occurring objects in Image Net [12]). Finally, we highlight its practical utility for evaluating model robustness by identifying out-of-distribution samples and quantifying the resulting degradation in model accuracy. |
| Researcher Affiliation | Academia | Jinho Choi KAIST AI Hyesu Lim KAIST AI Steffen Schneider1,2 Helmholtz Munich Jaegul Choo Correspondence: EMAIL; 1Institute of Computational Biology, Computational Health Center, Helmholtz Munich; 2Munich Center for Machine Learning (MCML) |
| Pseudocode | No | The paper describes the methodology using prose, mathematical formulas, and figures (e.g., Figure 2 illustrates the Concept Scope framework). However, it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at https://github.com/jjho-choi/Concept Scope. |
| Open Datasets | Yes | We conduct experiments on three widely used benchmarks: Celeb A [43], Waterbirds [59], and NICO++ [76]. The Waterbirds dataset contains two classes: waterbirds and landbirds. ... We train the SAE using Image Net-1K [12]. |
| Dataset Splits | Yes | Caltech-101 [18]. ...includes 3,066 training images and 6,080 test images. Waterbirds [59]. ...4,056 images in the training split and 5,794 images in the test split. Of these, 95% of the waterbird images have ocean or lake backgrounds, while the remaining 5% feature forest or bamboo forest backgrounds. Conversely, 95% of landbird images have forest or bamboo forest backgrounds, with the remaining 5% featuring ocean or lake backgrounds. The test set consists of 5,794 images, balanced equally across backgrounds (50% water backgrounds, 50% land backgrounds) for both classes. NICO++ [76]. ...The training sets contain 9,359 samples for NICO++75, 8,124 for NICO++90, and 7,844 for NICO++95. In contrast, the test set is designed to be balanced: for each class, images are evenly distributed across the six background types. This results in 300 samples per class (50 per background), for a total of 1,800 test images. |
| Hardware Specification | Yes | Computation resources. We ran all experiments on a single NVIDIA RTX A6000 GPU. With precomputed CLIP image embeddings, each iteration (batch size 64) took about 0.79 seconds, used 24 GB of GPU memory, and required 2,208 GFLOPs. |
| Software Dependencies | Yes | We train the SAE using Open AI CLIP Vi T-L/143 as the vision encoder... We query the GPT-4o API5... We employ the BLIP-2 [35] model architecture with a 2.7B parameter model15... For LLa VA-Ne XT [34], we utilize the 8B parameter model16... Specifically, we use gemini-2.0-flash-lite14 as the LLM... |
| Experiment Setup | Yes | Hyperparameter settings. We train the SAE using Open AI CLIP Vi T-L/143 as the vision encoder and extract embeddings from its penultimate (23rd) layer with embedding dimension d = 1024. The vision encoder resizes the input images to 224 × 224, producing 257 tokens consisting of one class token ([CLS]) and 16 × 16 spatial tokens. We apply an expansion factor of 32, resulting in a SAE latent dimensionality d = 32, 768. We set the L1 sparsity loss weight λ = 8 × 10−5, and the learning rate to 4 × 10−4 with 500 warmup steps using constant scheduling. We initialized decoder bias with the geometric median [53] of the first training batch and applied ghost gradients [5] to resample dead neurons caused by sparsity regularization. We use a batch size of 64 and train on Image Net-1K 4 training split, comprising approximately 1.28 million images, for 5 epochs. |