Research map / Label-efficient and robust learning / Recognition and long-tail

Restyling Unsupervised Concept Based Interpretable Networks with Generative Models

ICLR 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Self-explaining and visual explanation networks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VisCoIN, an unsupervised concept-based interpretable network that maps concept activations into the latent space of a pretrained StyleGAN2-ADA via a concept translator, enabling high-quality concept visualization by activation imputation.

Problem

Visualizing and interpreting unsupervised concept dictionaries in concept-based interpretable networks is hard for large-scale images, since activation maximization yields unnatural patterns and maximally activating samples lack granularity.

Approach

Use a frozen pretrained generative model (StyleGAN2-ADA) as the decoder: a learned concept translator maps concept activations to the generator's W+ latent space, with LPIPS and classifier-consistency reconstruction losses, so that concept activations can be scaled and visualized through generated images.

Key results

• VisCoIN reaches 79.44% on CUB-200 and 79.89% on Stanford Cars, versus 77.2% (FLINT) and 50.02% (FLAEM) on those datasets, close to the pretrained ResNet50 (80.56%, 82.28%).
• Reconstruction FID improves to 8.55 on CelebA-HQ, 15.85 on CUB-200 and 6.77 on Stanford Cars, versus FLINT's 30.45, 53.16 and 64.16.
• Consistency of visualization is 85.5 on CelebA-HQ, 85.0 on CUB-200 and 82.7 on Stanford Cars (lambda=2), above FLINT and FLAEM; faithfulness at tau=0.1 on CUB-200 is 0.251 vs 0.004 for FLINT.

Measured effects reported in the paper

Surprising findings

• FLINT achieves lower pixel MSE than VisCoIN despite VisCoIN's much better perceptual and distributional reconstruction quality.
• FLAEM's accuracy collapses to about 50% on CUB-200 and Stanford Cars, far below its CelebA-HQ result.

Limitations

• Learned concepts can modify more than one high-level feature at once (e.g., eye-squint and beard).
• Reconstruction quality and visualization depend on the quality of the pretrained generator; ProgressiveGAN reconstruction was worse.
• Extreme activation scaling (lambda above about 3-4) pushes latents away from the average and makes generated outputs less reliable.

Experimental setup

Datasets: CelebA-HQ CUB-200 Stanford Cars FashionMNIST

Benchmarks: CelebA-HQ CUB-200 Stanford Cars FashionMNIST

Metrics: accuracy MSE LPIPS FID faithfulness FF_x median consistency CC_k accuracy

Baselines: FLINT FLAEM SENN ResNet50 pretrained classifier Random concept selection

Builds on: FLINT StyleGAN2-ADA LPIPS SENN

Closest papers by meaning

• Interpretable Generative Models through Post-hoc Concept Bottlenecks (CVPR 2025 · similarity 0.89)
• Enhancing Performance of Explainable AI Models with Constrained Concept Refinement (ICML 2025 · similarity 0.88)
• From Causal to Concept-Based Representation Learning (NeurIPS 2024 · similarity 0.88)
• Explainable Concept Generation through Vision-Language Preference Learning for Understanding Neural Networks' Internal Representations (ICML 2025 · similarity 0.88)
• ConceptBed: Evaluating Concept Learning Abilities of Text-to-Image Diffusion Models (AAAI 2024 · similarity 0.87)
• Visual Concept Connectome (VCC): Open World Concept Discovery and their Interlayer Connections in Deep Models (CVPR 2024 · similarity 0.87)