Introduction
This paper introduces Markov Logic Mixtures of Gaussian Processes (MLxGP), a novel framework that enables machines to "read" and interpret regression data using probabilistic logic. By intertwining Gaussian Processes (GPs) with Markov Logic Networks (MLNs), the model can simultaneously perform non-parametric regression and logical reasoning. This allows background knowledge to improve numerical predictions while raw data is used to refine logical beliefs.
The Challenge: Machine Reading Regression Data (MR²D)
Traditional regression models, such as Gaussian Processes, are highly effective at mapping continuous inputs to outputs but often function as "black boxes" lacking semantic context. Conversely, probabilistic logical models like Markov Logic Networks excel at reasoning about complex relationships but struggle with raw, non-linear continuous data.
The authors address this gap through the concept of Machine Reading Regression Data (MR²D). For example, a robot observing a human setting a table sees a stream of hand-height coordinates. Without context, this is merely a sequence of numbers; with logic, the robot can infer that a high hand position moving toward a cupboard likely indicates the action of grabbing a cup. The challenge lies in creating a unified model where numerical data and logical rules inform one another.
Methods: The MLxGP Framework
The core innovation of MLxGP is the combination of two powerful statistical tools:
- Mixtures of Gaussian Processes (MGP): Rather than using a single GP to model a complex signal, the data is partitioned among several "experts" (individual GPs). A gating function determines which expert is responsible for a given data point.
- Markov Logic Networks (MLN): This formalism attaches weights to first-order logic formulas, creating a template for a Markov network that handles uncertainty in logical rules.
In the MLxGP framework, the gating function is directly connected to an MLN. The MLN provides a declarative prior, using background knowledge (e.g., "an object cannot be in two places at once") to influence which GP expert should be active at any given time.
Learning and inference are performed using an Expectation-Maximization (EM) algorithm. In the E-step, the model computes the probability that a data point belongs to a specific expert by combining the GP's likelihood with the MLN's logical beliefs. In the M-step, the hyperparameters of the GP experts are optimized based on these assignments.
Experimental Evaluation
The model was evaluated across several distinct domains:
- TUM Kitchen Dataset: Real-world motion capture data of a subject setting a table, involving hand heights and logical labels for objects.
- 1D Terrain Model: A synthetic dataset testing how logic can assist regression in areas with missing data (e.g., predicting terrain shape behind an obstacle).
- Friends-Coffee: A synthetic dataset relating kitchen activity to social relationships.
- Motorcycle Dataset: A classic benchmark where MLxGP was used to automatically partition head acceleration data into physical phases (riding, impact, and hitting the ground).
Key Results and Insights
The primary insight of this work is the bidirectional flow of information. Regression data helps the logic disambiguate states, while logic helps the regression "fill in the blanks" where data is noisy or absent. Unlike standard mixtures of experts, MLxGP experts have semantic meaning defined by the logic (e.g., "Expert 1" represents the state of "carrying a plate").
- Logical Accuracy: On the Kitchen task, MLxGP achieved a Conditional Marginal Log-Likelihood (CMLL) of -1.07, significantly outperforming the standalone MLN (-57.81). This demonstrates that "reading" regression data is essential for accurate logical inference.
- Regression Performance: In the Terrain dataset, MLxGP achieved a much lower error (NRMSE of 0.04) compared to standard GP regression (0.10) by using logic to predict terrain shape in unobserved areas.
- Interpretability: The model successfully partitioned the Motorcycle benchmark into its three physical phases without sacrificing the accuracy of the fit, providing a semantically meaningful decomposition of the data.
Conclusion
MLxGP represents a significant step toward machines that can reason about the continuous world. While current work uses manually fixed logical weights, future research aims to learn these weights jointly with GP hyperparameters and scale the system to massive robotics datasets using more efficient approximate inference techniques.
Summary updated