Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence

Gary Marcus

TL;DR

The paper argues that robust AI—systems that can reliably apply knowledge across diverse, open-ended contexts—will require more than scale and data. It proposes a four-step program centered on hybrid neurosymbolic architectures, large-scale knowledge bases with innate priors, explicit reasoning, and rich cognitive models to ground understanding and inference. It critiques purely data-driven approaches (e.g., GPT-2/Meena) for their brittle generalization and lack of causal, temporal reasoning, and it surveys existing hybrids and knowledge-structuring efforts as proof of concept. If realized, this framework could yield AI that is safer, more interpretable, and capable of trustworthy interaction in homes, roads, healthcare, and industry—the kind of robust intelligence needed for real-world deployment.

Abstract

Recent research in artificial intelligence and machine learning has largely emphasized general-purpose learning and ever-larger training sets and more and more compute. In contrast, I propose a hybrid, knowledge-driven, reasoning-based approach, centered around cognitive models, that could provide the substrate for a richer, more robust AI than is currently possible.

The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence

TL;DR

The paper argues that robust AI—systems that can reliably apply knowledge across diverse, open-ended contexts—will require more than scale and data. It proposes a four-step program centered on hybrid neurosymbolic architectures, large-scale knowledge bases with innate priors, explicit reasoning, and rich cognitive models to ground understanding and inference. It critiques purely data-driven approaches (e.g., GPT-2/Meena) for their brittle generalization and lack of causal, temporal reasoning, and it surveys existing hybrids and knowledge-structuring efforts as proof of concept. If realized, this framework could yield AI that is safer, more interpretable, and capable of trustworthy interaction in homes, roads, healthcare, and industry—the kind of robust intelligence needed for real-world deployment.

Abstract

Recent research in artificial intelligence and machine learning has largely emphasized general-purpose learning and ever-larger training sets and more and more compute. In contrast, I propose a hybrid, knowledge-driven, reasoning-based approach, centered around cognitive models, that could provide the substrate for a richer, more robust AI than is currently possible.

Paper Structure

This paper contains 23 sections, 7 figures, 1 table.

Table of Contents

  1. Towards robust artificial intelligence
  2. A hybrid, knowledge-driven, cognitive-model-based approach
  3. Hybrid architecture
  4. Symbolic operations over variables offer the only known solution, but the solution is partial
  5. Hybrids are often effective
  6. Common objections to hybrid models and symbol-manipulation
  7. Determining whether a given system is a hybrid system is not always trivial
  8. Summary
  9. Large-scale knowledge, some of which is abstract and causal
  10. What kind of knowledge will robust artificial intelligence require?
  11. Case Study: Containers
  12. Innate Frameworks for Knowledge
  13. Reasoning
  14. Cognitive models
  15. Discussion
  16. ...and 8 more sections

Figures (7)

  • Figure 1: snowplow 0.92
  • Figure 2: Multilayer perceptron trained on the identity function
  • Figure 3: Yarn feeder. Ball of yarn stays in large opening, and remains there, even as a single strand of yarn is pulled out.
  • Figure 4: Yarn feeder 2, pixel-by-pixel rather different.
  • Figure 5: Framework for knowledge about containers, drawn from (Davis et al., 2017)
  • ...and 2 more figures