Identifying the features of objects becomes a significant challenge when those features change in appearance. A fundamental trick of the human brain is "invariant" recognition: we know a coffee mug is a coffee mug whether it’s on the left side of the table, the right side, or turned upside down. For machine learning, however, unsupervised feature discovery is notoriously difficult when features don't stay put.
To address this, researchers have introduced the Transformed Indian Buffet Process (tIBP), a nonparametric Bayesian model that infers features capable of transforming across different instantiations. By mimicking human visual perception, the model demonstrates how we resolve visual ambiguity based on context.
The Problem with Standard Feature Discovery
Standard models like the Indian Buffet Process (IBP) are highly effective at figuring out which latent features compose an object. However, they generally assume those features appear in fixed positions. If a feature transforms—such as shifting spatially or rotating—the standard IBP interprets it as a completely new, distinct feature.
The tIBP bridges this gap by creating a model that can effectively "undo" these transformations, allowing it to find the underlying, invariant building blocks of visual scenes.
The Transformed Indian Buffet Process (tIBP)
The core innovation of the tIBP lies in its generative logic and its extension of the standard IBP framework.
- Transformation Parameters: While the standard IBP uses a binary matrix to track which objects possess which features, the tIBP adds a transformation parameter ($r_{nk}$) for every instantiation of a feature.
- Generative Logic: The model assumes that a visual scene is generated by selecting a set of "canonical" features, applying a specific transformation (like a spatial shift or rotation) to each, and then combining them.
- Inference: To deduce features from raw images, the model utilizes Gibbs sampling. A crucial technical detail is that the model marginalizes over possible transformations during sampling. This prevents it from getting stuck in local optima, allowing for an efficient search for the most likely features and their locations.
- Learning Invariance Types: The model is further extended with a latent variable ($t_k$) that allows it to learn which types of transformations are valid for a specific category. For instance, it can learn that in a certain context, "squares" might vary in size, but not in rotation.
Resolving Visual Ambiguity
Allowing features to transform introduces new kinds of ambiguity. The tIBP resolves these ambiguities using principles that closely align with human cognition.
Simplicity as a Tie-Breaker
When a scene is ambiguous, the tIBP follows a "minimum description length" logic. It prefers the simplest representation possible.
For example, if two bars frequently move together, the model must decide: are these two separate features that coincidentally move in sync, or are they a single "unitized" complex feature? The tIBP naturally favors treating them as a single unitized feature if their spatial relationship remains constant. In behavioral experiments, when the distance between two bars varied, both the tIBP model and human participants correctly switched to treating them as separate features.
Context-Dependent Invariance
The model also elegantly explains context-dependent perception, such as the famous Mach square/diamond illusion.
A square rotated 45 degrees is often perceived as a distinct shape—a diamond—because our brains have not "licensed" rotation as a standard transformation for that shape in everyday contexts. The tIBP mirrors this behavior. However, if the model is trained in a context where rotation is a common, valid transformation, it learns to treat the square and the diamond as the exact same invariant feature. In experiments testing this, the extended tIBP successfully learned whether a set of objects was "rotation-invariant" or "size-invariant," achieving a strong 0.68 correlation with human judgments.
Limitations and Future Directions
While the tIBP successfully extracts invariant features and aligns closely with human behavioral data, it has a few limitations:
- Computational Efficiency: The current inference method relies on Gibbs sampling, which becomes computationally expensive as the number of possible transformations (pixels, angles, scales) scales up.
- Simplified Priors: The model currently uses a discrete uniform prior for transformations. Incorporating more "perceptually realistic" priors—such as a human-like preference for vertical and horizontal orientations—could improve performance.
- Relative Geometry: The tIBP treats features as independent once transformed. Future iterations could explore how features interact structurally (e.g., "Feature A is always above Feature B"), moving the model closer to a comprehensive structural theory of shape.
Ultimately, the tIBP provides a flexible, nonparametric approach to unsupervised feature learning, proving that Bayesian models can successfully capture the nuanced, context-dependent ways humans recognize the building blocks of the visual world.
Summary updated