Recognition of All Categories of Entities by AI

Hiroshi Yamakawa, Yutaka MatsuoarXiv 2022 · OpenReview

Abstract

Human-level AI will have significant impacts on human society. However, estimates for the realization time are debatable. To arrive at human-level AI, artificial general intelligence (AGI), as opposed to AI systems that are specialized for a specific task, was set as a technically meaningful long-term goal. But now, propelled by advances in deep learning, that achievement is getting much closer. Considering the recent technological developments, it would be meaningful to discuss the completion date of human-level AI through the "comprehensive technology map approach," wherein we map human-level capabilities at a reasonable granularity, identify the current range of technology, and discuss the technical challenges in traversing unexplored areas and predict when all of them will be overcome. This paper presents a new argumentative option to view the ontological sextet, which encompasses entities in a way that is consistent with our everyday intuition and scientific practice, as a comprehensive technological map. Because most of the modeling of the world, in terms of how to interpret it, by an intelligent subject is the recognition of distal entities and the prediction of their temporal evolution, being able to handle all distal entities is a reasonable goal. Based on the findings of philosophy and engineering cognitive technology, we predict that in the relatively near future, AI will be able to recognize various entities to the same degree as humans.

Introduction: The Quest for Human-Level AI

Predicting the arrival of Artificial General Intelligence (AGI) is notoriously difficult because there are infinite possible tasks to measure. Previous attempts to create a "technology map" for AI have been either too narrow, listing specific tasks, or too coarse, such as the "System 1 vs. System 2" distinction.

The authors argue that the core of intelligence is modeling the world. To do this, an agent must be able to recognize "distal entities"—things that exist out in the world—and predict how they change over time. This paper addresses the lack of a granular, exhaustive framework to track whether an AI can truly perceive the world with the same breadth and depth as a human.

The Ontological Sextet as a Technology Map

The authors introduce a framework based on the Ontological Sextet, derived from the work of philosopher Barry Smith. This sextet categorizes everything into six buckets based on two primary axes:

  1. Continuant vs. Occurrent: Things that endure over time (like a cat) vs. things that happen over time (like a walk).
  2. Particular vs. Universal: Specific instances (this specific cat) vs. general kinds (the concept of "cat").
The Ontological Sextet framework categorizes entities into six buckets based on their endurance over time and their specificity.

By mastering "entification"—the ability to recognize and model objects, qualities, and processes—AI can move beyond specialized tasks toward a generalized understanding of the world.

The Entification Pipeline

To bridge the gap between philosophy and engineering, the paper proposes a mechanism called Entification, which consists of a four-stage pipeline:

  • Case Set Sampling: Dynamically selecting relevant data from sensors.
  • Single-out Process: Segmenting an entity from its background.
  • Alignment Process: Correcting for distortions (like perspective or lighting) to find the invariant "description" of an entity.
  • Encoding Map: Compressing that description to identify its "kind" (universal) and its "qualities" (particulars).
The detailed Entification pipeline illustrates the process from raw observation to compressed encoded representation.

The authors posit that intelligence is essentially the ability to compress the vast, redundant information of the world into minimalistic representations to increase predictability. Modern AI components, such as attention mechanisms in Transformers, can be mapped to these stages (e.g., performing "Case Set Sampling" and "Single-out" functions).

Current State and the Path to 2030

Modern Deep Learning has largely mastered "Substantial Particulars." We can identify objects and even generate them from text. We are currently in the "disentanglement" phase, learning to separate an object's identity from its qualities (e.g., separating the "shape" of a car from its "color").

The most significant technical challenge remaining is "Occurrents" (processes). Recognizing a process requires defining temporal boundaries (when does a "collision" start and end?) and handling much higher computational loads. However, based on the current trajectory and the lack of critical obstacles in the entification of the remaining categories, the authors predict that AI will be capable of recognizing nearly any kind of existence—reaching a human level—by 2030.

Conclusion and Future Challenges

While the roadmap is clear, several hurdles remain. AI still struggles with complex inheritance hierarchies (e.g., understanding that a "poodle" is a "dog," which is a "mammal") without explicit labels. Future work must focus on unsupervised hierarchy discovery and better ways to abstract time. Fully mastering the "Process" category is the gateway to understanding causal relationships, which is essential for creative problem-solving and true AGI.

Summary updated

Abstract

Human-level AI will have significant impacts on human society. However, estimates for the realization time are debatable. To arrive at human-level AI, artificial general intelligence (AGI), as opposed to AI systems that are specialized for a specific task, was set as a technically meaningful long-term goal. But now, propelled by advances in deep learning, that achievement is getting much closer. Considering the recent technological developments, it would be meaningful to discuss the completion date of human-level AI through the "comprehensive technology map approach," wherein we map human-level capabilities at a reasonable granularity, identify the current range of technology, and discuss the technical challenges in traversing unexplored areas and predict when all of them will be overcome. This paper presents a new argumentative option to view the ontological sextet, which encompasses entities in a way that is consistent with our everyday intuition and scientific practice, as a comprehensive technological map. Because most of the modeling of the world, in terms of how to interpret it, by an intelligent subject is the recognition of distal entities and the prediction of their temporal evolution, being able to handle all distal entities is a reasonable goal. Based on the findings of philosophy and engineering cognitive technology, we predict that in the relatively near future, AI will be able to recognize various entities to the same degree as humans.
BibTeX
@misc{yamakawa2022recognition,
  author = {Hiroshi Yamakawa and Yutaka Matsuo},
  title = {{Recognition of All Categories of Entities by AI}},
  year = {2022},
  eprint = {2208.06590},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.06590},
}
Figure