A Dynamical Model of Priming and Repetition Blindness

Daphne Bavelier, Michael I. JordanNeurIPS 1992

Human visual word recognition exhibits surprisingly non-linear behavior. When a person sees two identical words in quick succession, their brain’s response depends heavily on the precise timing between the presentations. Researchers Daphne Bavelier and Michael I. Jordan propose a signal-processing model that explains why our ability to recognize repeated words changes so drastically based on these temporal constraints.

By treating word recognition as a dynamical system with adaptive thresholds and baselines, their model unifies three distinct psychological phenomena into a single mathematical framework:

  1. Masked Priming (<40ms): The viewer does not consciously perceive the first word, but its presence makes them faster at recognizing the second.
  2. Repetition Blindness (~100ms): The viewer consciously sees the first word, but often completely fails to notice the second identical word.
  3. Classical Repetition Priming (>250ms): The viewer sees both words, and the first one makes the second one easier to process.

The primary challenge in cognitive psychology has been explaining how the same perceptual system can be "facilitated" (primed), then "inhibited" (blind), and then "facilitated" again simply by altering the presentation duration. Bavelier and Jordan aim to move beyond descriptive psychology to a formal, dynamical model that naturally captures this U-shaped response.

A Signal-Processing Approach to Word Recognition

The authors built a model grounded in Signal Detection Theory and Dynamical Systems. It consists of two main components: a geometric representation of words and a cascaded processing pipeline.

1. Geometric Representation

Rather than using arbitrary discrete IDs for words, the model utilizes a continuous vector space:

  • Similarity: An inter-word similarity matrix was calculated using dynamic programming, based on letter-level visual similarity.
  • Mapping: This matrix was converted into a vector space using Multidimensional Scaling (MDS).
  • Geometry: Word vectors were projected onto a semi-hypersphere. In this space, the origin represents a "blank" stimulus, and the magnitude of the vector represents the signal energy.

2. The Processing Pipeline

The model processes input through a cascaded filter system designed to handle noisy, continuous signals.

Model Architecture and Processing Stages
  • Perceptual Stage: A bank of noisy linear filters smooths the incoming visual input signal.
  • Detection Stage: A matched filter—a standard engineering mechanism for detecting known signals in noise—projects the smoothed output onto the lexicon to generate "word activation" signals.
  • Decision Logic: A binary decision for word recognition occurs only if the word's activation minus a variable baseline ($\mu$) exceeds a variable threshold ($\theta$).

Emergent Dynamics: Thresholds and Baselines

The most significant insight of the paper is that the "U-shaped" behavioral response does not require three separate cognitive modules. Instead, masked priming, repetition blindness, and classical priming emerge naturally from the interaction between the filters, the baseline, and the threshold.

Dynamic Threshold and Baseline Interaction
  • Temporal Superposition (Masked Priming): At very short intervals (<40ms), the signals from the two words physically overlap within the perceptual filters. This boosts the total activation before the system has a chance to reset, leading to facilitation.
  • Baseline as a "Reset" Mechanism (Repetition Blindness): When a word is successfully detected, the baseline ($\mu$) for that word and its orthographic neighbors is temporarily raised. If a second instance of the word arrives before this baseline resets (around 100ms), the new activation fluctuation isn't large enough to trigger a second detection. The system is effectively "blind" to the repetition.
  • Threshold as Adaptation (Classical Priming): The threshold ($\theta$) decreases during activation, acting as a form of adaptation. In long-duration trials (>250ms), the baseline has reset, but this lowered threshold remains, making subsequent detection of the same word easier.

Validation and Key Results

Rather than using a modern machine learning dataset, the model was calibrated and evaluated against established behavioral benchmarks from psychological literature, including foundational studies by Kanwisher (1987) on repetition blindness and Humphreys et al. (1987) on masked priming.

The model successfully replicates the qualitative transition from facilitation to inhibition and back to facilitation. Furthermore, because it uses a vector space based on letter similarity, it correctly predicts that "orthographic neighbors" (words that look visually similar, like BRING and BLING) will also interfere with or prime each other. It also naturally accounts for the recency effect—why the last word in a list is remembered better—because it is followed by a "blank" (the origin of the vector space), leaving its residual activation undisrupted.

Limitations and Future Work

While the dynamical equations are inspired by signal processing, the authors acknowledge this is a functional model rather than a literal map of neural circuitry. Additionally, parameters for the perceptual filter and threshold were specifically tuned to match existing behavioral data. While repetition blindness emerged organically without specific tuning, the model's predictive power on entirely new, untested phenomena remains an open question. Finally, the model currently focuses strictly on orthographic (visual) similarity and does not account for semantic (meaning-based) priming or the role of sentence context in word recognition.

Summary updated

BibTeX
@inproceedings{bavelier1992dynamical,
  author = {Daphne Bavelier and Michael I. Jordan},
  title = {{A Dynamical Model of Priming and Repetition Blindness}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {1992},
}
Figure