Tensor Logic "Unifies" AI Paradigms [Pedro Domingos]

The podcast explores Tensor Logic, a new language for AI, with Pedro Domingos, a computer science professor at the University of Washington. Domingos argues that Tensor Logic unifies symbolic AI, deep learning, kernel machines, and graphical models by combining tensor algebra and logic programming. He emphasizes its ability to perform automated reasoning and auto-differentiation, addressing the limitations of current systems like PyTorch. The discussion covers how Tensor Logic facilitates structure learning through gradient descent and predicate invention, enabling the discovery of new, explanatory relations in data. Domingos also addresses concerns about Turing completeness and the practical adoption of Tensor Logic, highlighting its potential to solve issues like hallucination and opacity in AI systems.

Outlines

Part 1: Introduction to Tensor Logic

Part 2: Core Properties and Technical Foundations

Part 3: Unifying AI Modalities and Structure Learning

Part 4: Symmetries, Physics, and Complexity

Part 5: Computational Universality and Efficiency

Part 6: Reasoning, Analogy, and Soundness

Part 7: Adoption, Education, and Future Outlook

Sign in to continue reading, translating and more.

Open full episode in Podwise

Machine Learning Street Talk

Part 1: Introduction to Tensor Logic

00:00Pedro Domingos' Vision: Unifying AI Paradigms with Tensor Logic

Pedro Domingos' Vision: Unifying AI Paradigms with Tensor Logic

00:58Tensor Logic: Deductive Reasoning, Hallucinations, and the Einstein Summation

Tensor Logic: Deductive Reasoning, Hallucinations, and the Einstein Summation

02:25Turing's Universal Machine and the Quest for Universal Induction

Turing's Universal Machine and the Quest for Universal Induction

Part 2: Core Properties and Technical Foundations

04:16Tensor Logic as the Language of AI: Properties and Goals

Tensor Logic as the Language of AI: Properties and Goals

05:24Key Properties of Tensor Logic: Automated Reasoning, Learning, and Scalability

Key Properties of Tensor Logic: Automated Reasoning, Learning, and Scalability

06:43Tensor Logic's Unifying Power and the Role of Calculus in Physics

Tensor Logic's Unifying Power and the Role of Calculus in Physics

07:53Einstein Summation: The Foundation of Tensor Logic and Deep Learning

Einstein Summation: The Foundation of Tensor Logic and Deep Learning

09:59Tensor Logic's Advantages: Syntax, Efficiency, and Unification of Symbolic and Numeric AI

Tensor Logic's Advantages: Syntax, Efficiency, and Unification of Symbolic and Numeric AI

12:26Tensor Logic: Generalizing Joins and Projections to Numeric Values

Tensor Logic: Generalizing Joins and Projections to Numeric Values

14:42Tensor Logic: Universality, Abstraction, and Practical Application

Tensor Logic: Universality, Abstraction, and Practical Application

Part 3: Unifying AI Modalities and Structure Learning

17:50Tensor Logic: Unifying AI Modalities and The Master Algorithm

Tensor Logic: Unifying AI Modalities and The Master Algorithm

19:10Tensor Logic: Scaffolding for The Master Algorithm

Tensor Logic: Scaffolding for The Master Algorithm

20:41Structure Learning and Predicate Invention in Tensor Logic

Structure Learning and Predicate Invention in Tensor Logic

22:25Tensor Logic: Gradient Descent and Architecture Discovery

Tensor Logic: Gradient Descent and Architecture Discovery

24:37Structuring Networks and Learning Topological Structure

Structuring Networks and Learning Topological Structure

25:55Tensor Logic: Encoding Multilayer Perceptrons and Structure Discovery

Tensor Logic: Encoding Multilayer Perceptrons and Structure Discovery

27:55Tensor Logic: Convolutional Networks and Inductive Bias

Tensor Logic: Convolutional Networks and Inductive Bias

Part 4: Symmetries, Physics, and Complexity

29:31Tensor Logic: Properties, Priors, and Transparency

Tensor Logic: Properties, Priors, and Transparency

31:23Geometric Deep Learning, Inductive Priors, and Tucker Decomposition

Geometric Deep Learning, Inductive Priors, and Tucker Decomposition

32:29Symmetries, Learning Bias, and Tensor Logic

Symmetries, Learning Bias, and Tensor Logic

34:47Closed vs. Open Domains and Computational Reducibility

Closed vs. Open Domains and Computational Reducibility

36:03Symmetries, Symmetry Breakings, and the Laws of Physics

Symmetries, Symmetry Breakings, and the Laws of Physics

37:33Handling Computational Irreducibility and Controlling Predictability

Handling Computational Irreducibility and Controlling Predictability

39:02Multiple Levels of Description and the Role of Tensor Logic

Multiple Levels of Description and the Role of Tensor Logic

40:32Tensor Logic as a Universal Language for Multiple Layers

Tensor Logic as a Universal Language for Multiple Layers

Part 5: Computational Universality and Efficiency

42:47Tensor Logic: A Language for Science and the Star T Hack

Tensor Logic: A Language for Science and the Star T Hack

44:01Star T: Computational Efficiency and Memory Management

Star T: Computational Efficiency and Memory Management

45:35Turing Completeness and the Universality of Tensor Logic

Turing Completeness and the Universality of Tensor Logic

47:05Proving Computational Universality and the Siegelman Construct

Proving Computational Universality and the Siegelman Construct

48:35Alternative Proofs of Turing Completeness

Alternative Proofs of Turing Completeness

50:23Tensor Logic Operators and Memory Operations

Tensor Logic Operators and Memory Operations

52:53Star T: Accumulators and Overwriting History

Star T: Accumulators and Overwriting History

55:00Autoregressive Transformers and Unbounded Rewrite Memory

Autoregressive Transformers and Unbounded Rewrite Memory

57:20Turing's Achievement: The Universal Machine

Turing's Achievement: The Universal Machine

59:32Transformers and the Ability to Learn from Small Examples

Transformers and the Ability to Learn from Small Examples

1:00:41The Universal Induction Machine

The Universal Induction Machine

Part 6: Reasoning, Analogy, and Soundness