Abstract:In this paper we derive an efficient algorithm to learn the parameters of structured predictors in general graphical models. This algorithm blends the learning and inference tasks, which results in a significant speedup over traditional approaches, such as conditional random fields and structured support vector machines. For this purpose we utilize the structures of the predictors to describe a low dimensional structured prediction task which encourages local consistencies within the different structures while learning the parameters of the model. Convexity of the learning task provides the means to enforce the consistencies between the different parts. The inference-learning blending algorithm that we propose is guaranteed to converge to the optimum of the low dimensional primal and dual programs. Unlike many of the existing approaches, the inference-learning blending allows us to learn efficiently high-order graphical models, over regions of any size, and very large number of parameters. We demonstrate the effectiveness of our approach, while presenting state-of-the-art results in stereo estimation, semantic segmentation, shape reconstruction, and indoor scene understanding.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to improve the efficiency of parameter learning in structured prediction, especially when dealing with high - order graph models. Traditional structured prediction methods (such as conditional random fields and structured support vector machines) need to perform inference tasks when optimizing parameters, which often involves comparing training labels with predicted labels, and the number of these labels is usually exponential. Therefore, traditional methods are very computationally expensive, especially when dealing with labels with complex structures.
### Main contributions of the paper
1. **Fusion of learning and inference**: The author proposes a new algorithm that fuses the learning and inference tasks together, thus significantly accelerating the parameter learning process. By utilizing the structural characteristics of the predictor, this algorithm can describe the structured prediction task in a low - dimensional space and ensure local consistency between different structures.
2. **Extension of the log - loss function**: In order to unify conditional random fields (CRFs) and structured support vector machines (structured SVMs), the author introduces a temperature parameter \(\epsilon\) to define the extended log - loss function:
\[
\bar{\ell}_{\epsilon\text{-log}}(w, x, y) \triangleq -\epsilon \log p(x,y)(\hat{y}; w, \epsilon)
\]
where,
\[
p(x,y)(\hat{y}; w, \epsilon) \propto \exp\left(\frac{\ell(y, \hat{y}) + \sum_k w_k \phi_k(x, \hat{y})}{\epsilon}\right)
\]
When \(\epsilon = 1\), the extended log - loss degenerates to the standard log - loss; when \(\epsilon \to 0\), it smoothly approximates the hinge loss.
3. **Decomposition of the upper bound**: The author derives a low - dimensional upper bound to approximate the extended log - loss and decomposes it into various regions in the graphical model. This decomposition enables the learning and inference tasks to be carried out simultaneously, thus improving the computational efficiency.
4. **Convex optimization guarantee**: Through convexity, the author ensures that the proposed algorithm can converge to the optimal solutions of the low - dimensional primal and dual programs. In addition, this method allows efficient learning of high - order graph models with regions of arbitrary size and a large number of parameters.
### Experimental results
The author verifies the effectiveness of this method in multiple computer vision tasks, including stereo estimation, semantic segmentation, shape reconstruction, and indoor scene understanding. The experimental results show that this method is not only more efficient than traditional methods but also reaches the state - of - the - art level in performance.
### Summary
The main objective of this paper is to improve the efficiency of parameter learning in structured prediction by fusing the learning and inference tasks, especially for high - order graph models with complex structures. By introducing the extended log - loss function and the low - dimensional decomposition upper bound, the author proposes an efficient algorithm and verifies its superiority in multiple practical applications.