Introduction
The integration of deep learning with symbolic reasoning, known as Neuro-symbolic AI, has traditionally focused on discriminative tasks such as classification. However, the generative side of this field has remained underdeveloped. Standard generative models, like Variational Autoencoders (VAEs), often conflate perceptual learning with logical reasoning, embedding the rules of a specific task directly into their latent representations. This paper introduces VAEL (Variational Autoencoder Logic), a novel framework that decouples these components by bridging VAEs with Probabilistic Logic Programming (PLP).
The Challenge of Generative Neuro-Symbolic AI
Current generative models face a significant hurdle: a lack of modularity. For instance, if a VAE is trained to generate images of handwritten digits that sum to a specific value, the model’s latent space becomes biased toward the "addition" operation. If the task changes—for example, requiring the generation of digits that multiply to a certain value—the model typically requires complete retraining on a new dataset.
The authors identify three core requirements for a more robust framework:
- The ability to encode raw data into expressive symbolic representations.
- The capacity for high-level reasoning over these representations.
- The flexibility to generalize to new logical tasks using previously acquired perceptual knowledge.
The VAEL Framework
VAEL addresses these needs through an end-to-end differentiable architecture that combines neural perception with formal logic.
Hybrid Latent Space
The architecture utilizes a dual-layered latent space. It incorporates standard subsymbolic latent variables (the neural component) alongside a structured representation governed by a probabilistic logic program (the symbolic component). This allows the model to handle both the continuous nature of image data and the discrete nature of logical rules.
Logical Reasoning and Differentiability
The logic program serves as a bridge, defining the relationships between symbolic representations required to satisfy specific constraints. Crucially, despite the inclusion of discrete logical structures, the entire pipeline remains differentiable. This enables the neural components to learn the mapping from pixels to symbols while the logic guides the overall structural coherence of the generated output.
Modular Generalization
A defining feature of VAEL is its modularity. Because the reasoning logic is handled by an external program rather than being "baked" into the neural weights, users can swap logical rules (e.g., changing from sum(A, B, C) to mult(A, B, C)) while keeping the trained neural encoder and decoder intact.
Experimental Evaluation
The framework was evaluated using tasks involving images of multiple handwritten digits, primarily derived from the MNIST dataset.
Training and Task Switching
During training, the model received minimal supervision, provided only with the results of a logical operation (e.g., "these two images sum to 10"). The model was then tested on its ability to perform "unseen" generation tasks. This included generating images that satisfy multiplication or power operations—tasks the model was never explicitly trained on—by leveraging the perceptual knowledge gained during the initial addition-based training.
Key Insights and Results
The experiments yielded several significant findings:
- Decoupling Perception from Logic: By separating the "what" (identifying a digit) from the "how" (performing arithmetic), VAEL achieves a level of modularity absent in pure connectionist models.
- Knowledge Reuse and Zero-Shot Learning: Once the model learns to identify symbols within a noisy environment, that knowledge is "unlocked." VAEL can solve entirely new tasks at inference time simply by providing it with a new logic program.
- Data Efficiency: The inclusion of a logical prior acts as a data multiplier. The logic provides a template for the output, significantly reducing the amount of raw data required to learn complex mappings compared to standard deep learning approaches.
Conclusion and Future Work
VAEL represents a significant step forward as a general-purpose framework for integrating probabilistic logic programming into deep generative models. While the current work demonstrates success with arithmetic logic on digit datasets, future research is needed to explore scalability to high-dimensional symbolic spaces and more complex, non-mathematical logical reasoning. The authors position VAEL as a foundational "bridging" effort, opening new avenues for optimizing the interface between continuous latent spaces and discrete logic.
Summary updated