From Pixels to Perception: Interpretable Predictions via Instance-wise Grouped Feature Selection (2505.06003v2)

Abstract: Understanding the decision-making process of machine learning models provides valuable insights into the task, the data, and the reasons behind a model's failures. In this work, we propose a method that performs inherently interpretable predictions through the instance-wise sparsification of input images. To align the sparsification with human perception, we learn the masking in the space of semantically meaningful pixel regions rather than on pixel-level. Additionally, we introduce an explicit way to dynamically determine the required level of sparsity for each instance. We show empirically on semi-synthetic and natural image datasets that our inherently interpretable classifier produces more meaningful, human-understandable predictions than state-of-the-art benchmarks.

An Overview of "From Pixels to Perception: Interpretable Predictions via Instance-wise Grouped Feature Selection"

The paper by Moritz Vandenhirtz and Julia E. Vogt offers a novel approach to enhance interpretability in machine learning models, particularly in image classification tasks. The authors present "From Pixels to Perception" (P2P), a method focused on achieving inherently interpretable predictions through instance-wise sparsification of input images, shifting feature selection from the pixel level to more perceptually meaningful regions.

Methodology

P2P leverages the concept of sparse feature selection, but rather than operating directly on pixel-level granularity, it introduces a paradigm shift by focusing on semantically interpretable regions within images. The key innovation is the segmentation of images into perceptually coherent atomic units, thus aiming for a level of abstraction that aligns more closely with human cognition. This technique is implemented using superpixel algorithms, such as SLIC, to form these atomic regions.

The authors propose a probabilistic model utilizing the Gumbel-Softmax trick for sampling a binary mask that determines active regions for every input. This binary mask ensures that selected regions are treated discretely, thereby influencing the model's predictions based on perceptual significance rather than individual pixel values.

Another notable feature of the proposed approach is its dynamic thresholding mechanism that adapts the sparsity level based on the model's prediction confidence. This enables the model to dynamically query more information when necessary to reach a pre-defined confidence threshold, thereby optimizing both interpretability and prediction fidelity.

Empirical Evaluation

The empirical evaluation conducted on datasets such as CIFAR-10, ImageNet, COCO-10, and semi-synthetic BAM datasets demonstrates that P2P achieves competitive classification accuracy while providing more interpretable decisions compared to existing methods. Quantitatively, P2P accomplishes strong performance metrics close to the upper-bound black-box models while significantly reducing the amount of visual information processed by the model.

Furthermore, through insertion and deletion metrics, P2P exhibited superior fidelity, underscoring that its predictions are genuinely grounded in the highlighted regions of the images, effectively validating the inherent interpretability claim of the method.

Discussion of Results

The results emphasize the importance of aligning machine learning inputs with human perceptual understanding. By introducing semantic grouping in feature selection, P2P addresses several challenges present in traditional pixel-level approaches, such as the risk of models making predictions based on irrelevant pixel patterns.

P2P's approach shows that machine learning models can be both highly accurate and interpretable, which is crucial for applications in high-stakes domains where understanding model decisions is as vital as their predictive power.

Future Directions

The technique opens several avenues for future research:

1. Expansion to Other Modalities: While the focus has been on image data, the concept of perceptually meaningful segmentation could be extended to other data types, such as audio or text.
2. Integration with Other Interpretability Methods: Combining P2P with complementary interpretability tools could enhance both local and global understanding of model behavior.
3. Scalability and Efficiency: Further work could explore optimizing the computational efficiency of generating perceptually meaningful regions, especially for large-scale datasets.

Conclusion

Overall, the paper provides valuable insights and a substantive contribution to the field of interpretability. By combining machine learning with principles of human perception, P2P represents a step towards models that are not only predictive but also transparent and trustworthy. This work underscores the potential for advanced feature selection methods to transform our approach to machine learning, especially in contexts where interpretability is paramount.