Abstract
Semantic attributes have been proposed to bridge the semantic gap between low-level feature representation and high-level semantic understanding of visual objects. Obtaining a good representation of semantic attributes usually requires learning from high-dimensional low-level features, which not only significantly increases the time and space requirement but also degrades the performance due to numerous irrelevant features. Since multiattribute prediction can be generalized as an multitask learning problem, sparse-based multi-task feature selection approaches have been introduced, utilizing the relatedness among multiple attributes. However, such approaches either do not investigate the pattern of the relatedness among attributes, or require prior knowledge about the pattern. In this paper, we propose a novel feature selection approach which embeds attribute correlation modeling in multi-attribute joint feature selection. Experiments on both synthetic dataset and multiple public benchmark datasets demonstrate that the proposed approach effectively captures the correlation among multiple attributes and significantly outperforms the state-of-the-art approaches.
Introduction
In computer vision, semantic attributes serve as an intermediate layer between raw pixel data (low-level features) and high-level object categories. For instance, instead of directly identifying a "shoe," a model might first detect attributes like "red," "high-heel," or "formal." This approach is particularly useful for zero-shot learning, where a system must recognize objects it has never seen before based solely on a textual description of their attributes.
However, learning these attributes is challenging due to the "curse of dimensionality." Raw image data contains thousands of features, many of which are irrelevant to a specific attribute. Furthermore, attributes are not independent; for example, "high-heel" and "formal" often share similar shape-based features, whereas "red" is primarily determined by color information.
The Problem with Current Feature Selection
Existing methods for multi-attribute feature selection often fall into two categories. Some assume that all attributes share the same subset of low-level features, which is rarely true in practice. Others require human experts to manually group related attributes together before training, a process that is both labor-intensive and potentially biased. There is a clear need for a method that can automatically discover these underlying relationships to improve prediction accuracy and efficiency.
Proposed Methodology: Clustering-Based Joint Feature Selection
The researchers at Arizona State University proposed a framework that integrates attribute correlation modeling directly into the feature selection process. Their approach uses a multi-task learning objective with three key components:
- Clustering Regularization: The model automatically partitions attributes into clusters using a k-means-style approach. This allows the model to identify which attributes are related (e.g., texture-based vs. color-based) without human intervention.
- Group-Sparsity ($\ell_{2,1}$-norm): This mathematical constraint encourages attributes within the same cluster to share features (intra-group sharing) while forcing different clusters to compete for features (inter-group competition). This effectively filters out noise and irrelevant data.
- Alternating Structure Optimization (ASO): To solve the complex optimization problem, the authors developed an algorithm that toggles between updating feature weights and refining cluster assignments until the model reaches convergence.
Experimental Results
The proposed method was rigorously tested against state-of-the-art algorithms across several domains:
- General Feature Selection: On benchmark datasets like USPS (digits) and YaleB (faces), the model consistently outperformed standard methods. For example, it achieved 93.15% accuracy on USPS, significantly higher than the 86.25% achieved by standard Multi-Task Feature Selection (MTFS).
- Attribute Prediction: Using the aPascal/aYahoo and Animals with Attributes (AwA) datasets, the model showed a 2% to 4% improvement over previous leaders like DSVA in predicting attributes for unseen object categories.
- Zero-Shot Learning: The method demonstrated a significant advantage in zero-shot learning tasks on the SUN and AwA datasets, proving its ability to generalize high-level concepts to entirely new objects.
Conclusion and Future Work
By automatically discovering the latent structure of attribute correlations, this new feature selection approach provides a more accurate and efficient bridge for semantic understanding. While the model currently requires an estimate for the number of clusters ($k$), it represents a major step forward in automating the learning of complex visual concepts. Future research will focus on refining the selection of $k$ and optimizing the initial computational steps for even larger datasets.
Summary updated