It's About Time

Neil Madden, Brian LoganIJCAI 2005

Abstract

Human perception often involves "backwards referral in time," where subsequent stimuli influence the interpretation of prior events. This paper introduces the Temporal Abstraction Network (TAN), a decentralized cognitive architecture that models these phenomena using a hierarchy of processes with time-limited buffers. By simulating the "cutaneous rabbit" illusion, the authors demonstrate that complex temporal perceptions can emerge from parallel "multiple drafts" without requiring a central executive or biologically implausible processing delays.

Introduction

Human perception is not a live, frame-by-frame video feed. In many instances, what happens now fundamentally alters how we perceived what happened a few milliseconds ago. This "backwards referral in time" poses a challenge for traditional serial models of cognition, which often assume the brain must "wait" for all data to arrive before presenting it to a central conscious observer. Such delays are biologically implausible.

This paper proposes a computational framework for Daniel Dennett’s Multiple Drafts theory, suggesting that consciousness is a parallel, distributed process where multiple interpretations (drafts) of sensory data are constantly being revised.

The Temporal Abstraction Network (TAN)

The authors propose the Temporal Abstraction Network (TAN) as a mechanism for this distributed processing. The architecture is characterized by several key features:

  • Decentralized Processes: The system lacks a central "director"; instead, it consists of independent, parallel processes.
  • Time-Limited Buffers: Each process maintains an input buffer with a specific capacity and duration. The duration determines how long an item persists before being forgotten, naturally grouping temporally proximate events.
  • Production Rules and Inference: Processes use local rules to identify patterns within their buffers and generate inferences.
  • Bus Architecture: Conclusions are broadcast to an output bus, feeding into the input buffers of other processes in the hierarchy.
  • Temporal Abstraction: Processes higher in the hierarchy possess buffers spanning longer durations, allowing them to maintain abstract conclusions while low-level sensory details are discarded.
The Temporal Abstraction Network (TAN) architecture, illustrating the hierarchical flow from low-level sensory data to high-level inferences through buffers and feedback loops.

Modeling the "Cutaneous Rabbit" Illusion

The TAN model was tested against the "cutaneous rabbit" experiments (Geldard and Sherrick, 1972). In this illusion, a rapid sequence of taps at different locations (e.g., wrist then elbow) is perceived as a sequence of taps "hopping" up the arm.

The model processes this through three levels:

  1. Low-level: Taps occurring within 20ms are merged or overwritten in the buffer.
  2. Intermediate-level: Taps within 200ms are grouped into a "run" or sequence. A feedback loop allows the end point of a sequence to influence the perceived location of earlier taps.
  3. High-level: The system detects the completion of a run and generates a final report.

Crucially, the "revision" of perception occurs because new data arrives while old data is still in the buffer, not because the system was waiting to process the information.

Key Results

The TAN model successfully replicates qualitative findings from sensory saltation experiments:

  • Perceptual Grouping: It simulates how 15 rapid taps (ISI < 20ms) are perceived as a smaller number of discrete events.
  • Temporal Thresholds: It accounts for the breakdown of the illusion when the interval between taps exceeds the buffer duration (~200ms).
  • Spatial Displacement: The model demonstrates how the final tap in a sequence can retroactively influence the perceived spatial distribution of previous taps.

Discussion and Future Work

The TAN architecture demonstrates that "forgetting" is a functional feature of temporal processing, and that global coherence can emerge from local integration without a central executive.

Future work aims to extend this model to action selection, exploring how these "multiple drafts" translate into agent behavior. Additionally, the authors intend to apply the architecture to Libet’s experiments regarding the timing of conscious will, further probing the relationship between brain activity and conscious experience.

Summary updated

BibTeX
@inproceedings{madden2005it,
  author = {Neil Madden and Brian Logan},
  title = {{It's About Time}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2005},
}
Figure