About Hyperon+Neural-Symbolic Integration+SynerGAN

← Back to Neural-Symbolic Integration

Responsible: Ben Goertzel, Anatoly Belikov (semantic-vision team)

GitHub: singnet/semantic-vision (wiki documentation)

Status: Historical/Research. SynerGAN was designed as a neural-symbolic GAN architecture for OpenCog. The framework defined the theoretical integration of PLN and Pattern Mining into GAN training, with implementation milestones through compositional scene understanding. Current status of active development is unclear — the design predates the Hyperon transition.

Architecture Overview

SynerGAN extends the InfoGAN architecture by embedding symbolic probabilistic reasoning directly into the generative-discriminative game. Unlike standard GANs where both players are purely neural, SynerGAN integrates symbolic methods at both ends:

• Generative Player: Uses symbolic probabilistic methods (PLN) to sample latent variable values based on logical constraints, then feeds these into a neural generator to produce images/video
• Discriminative Player: Employs a neural network to extract perception states, applies symbolic Pattern Mining to recognize regularities among discretized "neural model predicates," then uses PLN inference to classify real vs. generated data

The key mechanism is SampleLink — a methodology for sampling latent variable values guided by logical constraints derived from the symbolic system. Continuous latent variables are discretized into neural model predicates that can be reasoned about symbolically. Dependencies among these predicates are discovered via Pattern Mining and refined through PLN inference.

Bidirectional Feedback

SynerGAN's defining innovation is bidirectional co-adaptation between neural and symbolic components. The generative symbolic system and discriminative symbolic system share abstract knowledge through AtomSpace: generative patterns help prune discriminative inference trees, and discriminative findings refine generative sampling distributions. The symbolic system can signal the neural network which "symbolic output neurons" produce uninterpretable patterns, driving network architecture adaptation. This creates genuine co-evolution rather than simple concatenation.

Compositional SynerGAN

For complex scenes, Compositional SynerGAN decomposes the problem hierarchically:

1. Cluster image regions/snippets into statistical groups
2. Train separate SynerGAN networks on each cluster
3. Iteratively refine clusters using learned network outputs as features
4. Compose multiple models hierarchically (e.g., glasses model within human head model)

This modular approach makes large-scale visual understanding tractable by reducing latent variable dimensionality per model. Transfer learning between compositional modules is a planned capability.

Implementation Milestones

• Milestone 1: Probabilistic Network InfoGANs (ProNetInfoGAN) — validate the probabilistic network approach on standard benchmarks (MNIST, SVHN, CelebA)
• Milestone 2: Full SynerGAN — integrate PLN and Pattern Mining into the GAN training loop
• Milestone 3: Compositional SynerGAN — hierarchical scene understanding with transfer learning

Extensions

The design extends beyond vision to multimodal perception (visual-auditory emotion recognition), action learning (robotic control via predictive modeling), and reinforcement learning through hierarchical composition of perception and action networks.

---

Related cards: Neural-Symbolic Integration · PLN · Pattern Mining · Meta-MeTTa Paper

(Provenance: github-wiki, singnet/semantic-vision wiki, 8 files)

Key References

• Goertzel et al. (2018), SynerGAN wiki — semantic-vision repository wiki (8 design documents)
• Goertzel (2025), Hyperon Whitepaper §7 — Neural-Symbolic Integration