Pedro Domingos: Tensor Logic Unifies AI Paradigms

What It Covers

Pedro Domingos presents Tensor Logic, a unified programming language for AI that combines tensor algebra from deep learning with logic programming from symbolic AI, enabling both automated reasoning and gradient descent learning within a single framework.

Key Questions Answered

• •Tensor Logic Unification: Einstein summation operations and logic programming rules are mathematically identical constructs operating on different data types (real numbers versus booleans). This insight enables expressing neural networks, symbolic reasoning, kernel machines, and graphical models within one language using only tensor equations.
• •Zero-Temperature Deduction: Setting temperature parameters to zero in Tensor Logic enables guaranteed sound deductive reasoning in embedding space without hallucinations. Random high-dimensional vector embeddings approximate identity matrices through dot products, allowing pure logical inference while learned embeddings enable analogical reasoning at higher temperatures for structure mapping.
• •Predicate Invention via Gradient Descent: Structure learning happens automatically through gradient descent using Tucker decomposition on tensor equations. The system discovers new predicates and relations not present in training data, similar to how matrix factorization reveals latent factors, enabling representation discovery comparable to scientific concept formation.
• •Universal Induction Challenge: Turing machines provide universal deduction but AI requires universal induction—a learning equivalent that can generalize from small examples to arbitrary problem sizes. Tensor Logic approaches this by enabling programs that learn addition from elementary examples yet apply to numbers of any length through compositional structure.
• •Adoption Strategy Through Education: Tensor Logic can preprocess into existing Python frameworks, allowing incremental adoption without rewriting codebases. Teaching AI courses with this single unified language instead of multiple frameworks reduces cognitive overhead, creating a generation of developers who prefer its declarative-procedural dual semantics for production systems.