[Paper Review] Posthoc Interpretation via Quantization

Francesco Paissan, Cem Subakan+1 more|arXiv (Cornell University)|Mar 22, 2023

Multimodal Machine Learning Applications4 citations

TL;DR

This paper introduces Posthoc Interpretation via Quantization (PIQ), a method that uses vector quantization to create class-specific, discrete latent representations for interpreting deep learning classifier decisions. By forcing the model to focus on salient input regions through a discrete bottleneck, PIQ generates more interpretable, human-understandable explanations than existing methods, as validated by user studies on images and audio.

ABSTRACT

In this paper, we introduce a new approach, called Posthoc Interpretation via Quantization (PIQ), for interpreting decisions made by trained classifiers. Our method utilizes vector quantization to transform the representations of a classifier into a discrete, class-specific latent space. The class-specific codebooks act as a bottleneck that forces the interpreter to focus on the parts of the input data deemed relevant by the classifier for making a prediction. Our model formulation also enables learning concepts by incorporating the supervision of pretrained annotation models such as state-of-the-art image segmentation models. We evaluated our method through quantitative and qualitative studies involving black-and-white images, color images, and audio. As a result of these studies we found that PIQ generates interpretations that are more easily understood by participants to our user studies when compared to several other interpretation methods in the literature.

Motivation & Objective

• To address the lack of human-understandable explanations in deep neural network decisions, especially in high-stakes domains like healthcare.
• To develop a post-hoc interpretation method that generates intuitive, localized explanations highlighting relevant input regions.
• To improve interpretability by leveraging supervision from foundational models like Segment-Anything Model (SAM) for complex data.
• To evaluate interpretability not only quantitatively but also through human preference studies across diverse data modalities.
• To enable application of interpretation to audio and low-level vision data where saliency is not easily defined.

Proposed method

• PIQ uses vector quantization to map classifier latent representations into a discrete, class-specific latent space, forming a bottleneck that emphasizes relevant input features.
• Each class has its own codebook in the vector quantization layer, enabling the model to learn distinct, interpretable concepts per class.
• The method incorporates supervision from pretrained models—such as SAM—for real-world images, allowing high-quality concept learning without manual annotations.
• For audio and low-dimensional data, saliency is derived via thresholding; for complex data, foundational models provide supervision to guide interpretation.
• The interpretation decoder reconstructs input regions based on the quantized latent codes, producing visual or audio overlays that highlight decision-relevant parts.
• Training is optimized using the vector quantization objective from VQ-VAE, ensuring discrete, compact, and meaningful representations.

Experimental results

Research questions

• RQ1 Can vector quantization be used to create class-specific, discrete latent representations that improve the interpretability of deep neural network decisions?
• RQ2 How do PIQ-generated interpretations compare to existing methods like LIME, GradCAM, and L2I in terms of human preference across image and audio data?
• RQ3 Can PIQ effectively leverage foundational models (e.g., SAM) to generate high-quality interpretations for real-world images without fine-tuning?
• RQ4 Does PIQ maintain fidelity to the original data distribution while producing more interpretable explanations, especially in ambiguous or overlapping input cases?
• RQ5 How does PIQ perform on audio data, particularly in interpreting mixtures of sound events where source separation is challenging?

Key findings

• PIQ outperforms multiple baseline methods on quantitative metrics for black-and-white images, closely matching the original data distribution.
• In user studies on black-and-white images, color images, and audio, participants consistently preferred PIQ interpretations over those from LIME, VIBI, FLINT, L2I, and GradCAM.
• On audio mixtures from the ESC-50 dataset, PIQ received significantly higher average opinion scores (5-point scale) than both official and re-implemented L2I baselines.
• The user study results showed that PIQ interpretations were rated as more accurate and easier to understand, especially in complex scenarios like overlapping digits and mixed audio.
• PIQ successfully generated meaningful, localized explanations for real-world images using supervision from the Segment-Anything Model (SAM), demonstrating scalability to complex data.
• The method generalizes across modalities, producing interpretable outputs for both image and audio data without requiring modality-specific retraining.