How to Put the Pieces of AI Together Again

Aaron SlomanAAAI 2006

Since the 1970s, the field of Artificial Intelligence has undergone a process of extreme fragmentation. Researchers have retreated into specialized silos—vision, planning, natural language processing, and robotics—each with its own isolated conferences, journals, and benchmarks. While this specialization has led to incremental improvements in specific tools, Aaron Sloman argues that this "forward-chaining" approach does not guarantee progress toward the ultimate goal of integrated, human-like intelligence. This paper explores a proposed methodology for reintegrating these disparate pieces through a "backward-chaining" roadmap.

The Problem of Fragmentation and "Scaling Up"

A central critique in Sloman’s work is the "misdiagnosis" of AI failures. When systems fail to meet extravagant predictions, the field often blames specific tools or representations, leading to shifts from symbolic AI to connectionism or Bayesian methods. However, the real failure may lie in an inadequate analysis of long-term requirements.

Most current AI research focuses on "scaling up"—making existing systems bigger, faster, or more data-intensive. In contrast, biological intelligence "scales out." Humans exhibit altricial development, spending a long period in a relatively helpless state to play and explore. This allows them to decompose the world into reusable, "orthogonal" competences that can be combined in novel, flexible ways to solve unforeseen problems.

Methodology: Scenario-Based Backward Chaining

To move toward integrated intelligence, Sloman proposes a methodology called Scenario-Based Backward Chaining. Instead of starting with what current tools can do and trying to improve them, researchers should:

  1. Define the End Goal: Create detailed descriptions of machines performing complex, integrated tasks that are currently impossible, such as a domestic robot capable of assisting the elderly in a dynamic home environment.
  2. Work Backward: Identify the "partially ordered milestones" or dependencies required to reach that goal. This involves determining what a robot must understand about physics to use a tool, or what social cues it must process to interact safely with a human.
Figure 1: A roadmap for AI development, showing the path from current capabilities to human-like machines through a network of intermediate milestones.

As shown in Figure 1, this approach identifies a network of intermediate competences ordered by dependency and difficulty. It also helps researchers identify "tempting dead-ends"—approaches that might show quick progress on narrow benchmarks but do not lead toward the final goal of integrated intelligence.

The Grid of Requirements

To systematically generate these scenarios and identify missing dependencies, Sloman suggests a multi-dimensional grid of requirements. This grid allows researchers to define "orthogonal recombinable competences"—skills that can be learned once and applied across many different scenarios.

Figure 2: A grid for systematic requirement analysis, mapping various entity types against specific competences.

By looking at the intersections of entity types (concrete, abstract, mobile, etc.) and competence types (perceiving, pushing, thinking, etc.), researchers can define specific technical requirements. This shifts the focus from "whose algorithm is better" to "what are the shared dependencies we all need to solve."

Key Cognitive Insights

This framework highlights several qualitative shifts needed in AI research:

  • Kantian vs. Humean Causation: Current AI, particularly Bayesian and deep learning approaches, focuses on "Humean" causation—statistical correlations. To achieve human-level manipulation, robots need "Kantian" understanding: the ability to reason about structural, deterministic necessity (e.g., understanding why a gear must turn if its neighbor turns).
  • Vision as Process: Vision should not be viewed merely as a tool for identifying static structures. For a robot to manipulate objects, vision must perceive concurrent processes at different levels of abstraction in real-time.

Conclusion and Future Directions

The success of this roadmap depends on the AI community reaching a consensus on the problems to be solved, even if they continue to disagree on the algorithmic means. While designing detailed future scenarios is difficult, it is a necessary step to move beyond narrow AI. Future work must focus on "orthogonal recombinable competences" and deeper investigations into how biological systems learn to decompose and recombine skills to navigate a complex world.

Summary updated

BibTeX
@inproceedings{sloman2006how,
  author = {Aaron Sloman},
  title = {{How to Put the Pieces of AI Together Again}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure