Published Mar 7, 2022

#68 DR. WALID SABA 2.0 - Natural Language Understanding [UNPLUGGED]

Explore the potential of hybrid AI models with Dr. Walid Saba as he critiques the limitations of neural networks and emphasizes the need for symbolic logic in advancing true natural language understanding and AI capabilities.
Episode Highlights
Machine Learning Street Talk (MLST) logo

Popular Clips

Episode Highlights

  • Cognitive Templates

    challenges the notion that cognitive templates can be learned through data-driven approaches. He argues that these templates are intrinsic and universal, precluding the possibility of learning them from individual experiences 1. This perspective suggests that children, by the age of three, have already mastered these templates, indicating their presence in our DNA rather than being acquired through learning.

    Walid asserts that everyone acquires exactly the same abstract categories or templates. It's like they're a gift to us from physical reality itself.

    ---

    These templates, according to Walid, are sophisticated and operate on a logic that is often ignored in AI research today 2.

       

    Learning Paradigms

    The current machine learning paradigms face limitations in effectively acquiring and applying knowledge. argues that the existing models fail to explain how children learn complex templates in a short time, suggesting the need for new learning paradigms 3. He criticizes the reliance on bottom-up data approaches, emphasizing that some knowledge cannot be learned differently or through sensory input alone.

    We're not allowed to learn this stuff. Not only we cannot. And we don't have enough time learning the ML paradigm that we have now.

    ---

    This critique extends to the handling of infinity in computational models, where the distinction between tractable and intractable infinities plays a crucial role in learnability 4.

Related Episodes