newsfilter.io
Interview

Noam Chomsky: Deep Learning is Useful but It Doesn't Tell You Anything about Human Language

  • Deep learning is characterized by its opacity, making it difficult to formulate definitive theorems regarding its capabilities and limits.
  • The core mechanism of deep learning involves identifying patterns within massive datasets rather than deriving scientific understanding.
  • Classification of Utility
    • Deep learning is deemed effective as an engineering tool for practical applications (e.g., Google Translate).
    • It is distinguished from science because it does not elucidate the fundamental nature of the subject matter (e.g., human language).
  • Critique of Methodology
    • Scientific inquiry prioritizes "critical experiments" designed to test specific theoretical questions, whereas deep learning relies on the sheer volume of unguided data.
    • Success metrics for deep learning focus on performance on existing corpora (e.g., the Wall Street Journal) rather than performance on data that violates system rules.
    • Case Study: Structure Dependence
      • Neural networks may perform well on languages relying on "linear proximity," a mode that violates actual human linguistic structure.
      • This performance is characterized as a scientific failure because it suggests the model fails to discover the true nature of the system, successfully approximating incorrect patterns.
  • Relationship to Behavioral Science
    • Claims by some researchers that deep learning vindicates Skinnerian behaviorism (Terry Sejnowski) are rejected; the technology does not validate behavioral theories.
  • Potential for Scientific Insight
    • Neural networks may occasionally reveal previously unnoticed patterns within complex structures.
    • Analogy to Other Fields
      • This approach is compared to "corpus linguistics" and "paleoanthropology," where researchers must derive conclusions from available records when critical experiments are impossible.
      • While these are considered serious studies, they are inherently limited compared to methodologies allowing for direct, critical experimentation with living subjects.