Gestalt Pattern Recognition with Arrays of Predetermined Neural Functions

TL;DR

This paper introduces a novel approach to pattern recognition using arrays of predetermined neural-like functions that perform ternary logical comparisons. The core idea is that these interconnected functions can directly sense the "shape" or "Gestalt" of a pattern, bypassing the need for complex feature extraction or learning. This work lays the groundwork for potentially simpler, more efficient, and mechanistic artificial systems capable of human-like pattern recognition.

What Problem Does This Paper Address?

The paper tackles the challenge of recognizing two-dimensional geometric patterns directly from their inherent shape, rather than relying on individual pixel characteristics or complex, often computationally expensive, feature extraction methods. Traditional approaches like the Perceptron, while capable of learning, often involve randomly connected arrays and require significant training. This work aims to demonstrate that a system with predetermined, logically connected neural-like units can achieve "Gestalt" recognition – perceiving the whole pattern as an integrated functional unit – without the need for self-learning or extensive preprocessing like normalization for size, orientation, or rotation. The authors also highlight the potential for creating more economical and efficient pattern recognition systems.

Methods

The core of the proposed method lies in an array of "predetermined neural functions." The paper focuses on a specific function: a logical comparator that performs a signed ternary logic comparison. This comparator takes two ternary input signals (representing values like -1, 0, +1) and outputs a ternary signal indicating whether the inputs are equal, or if not, the direction of the difference.

A key innovation is the concept of self-feedback within these arrays. By feeding the output of a comparator back into its reference input, the system can exhibit dynamic, yet deterministic, states. This feedback mechanism, combined with a ternary weigher (an additive function that determines the predominance of input signals), allows individual "neurons" to interact with their neighbors.

The paper explores different interconnection schemes for these arrays, primarily focusing on nearest orthogonal neighbors but also considering diagonal neighbors and weighted connections. The output of the array is analyzed through ratio detection, where the relative occurrences of different output states (e.g., +2, +1, 0, -1, -2) are found to be invariant for different patterns, even under transformations like rotation and scaling.

For empirical validation and exploration, the authors developed a digital simulation of these neural arrays, allowing for the testing of various configurations and input patterns without the expense of building physical hardware.

Datasets

Not discussed in the paper. The authors generated input patterns for testing and simulation rather than using pre-existing datasets.

Key Insights and Ideas

  • Gestalt Recognition via Predetermined Logic: The most significant insight is the demonstration that complex pattern recognition, akin to human Gestalt perception, can be achieved through arrays of simple, predetermined logical functions, rather than solely relying on learning.
  • Invariant Ratios as Signatures: The discovery that specific ratios of output states from the array are invariant to transformations (rotation, scaling) provides a powerful mechanism for pattern identification. This suggests a fundamental way the system encodes shape information.
  • Self-Feedback for Dynamic Behavior: The use of self-feedback in the logical comparator introduces dynamic, yet predictable, behavior, allowing the system to exhibit complex responses based on input timing and neighbor interactions.
  • Ternary Logic for Nuance: The adoption of signed ternary logic (-1, 0, +1) allows for more nuanced comparisons than binary logic, capturing not just equality but also the direction and magnitude of differences.
  • Mechanistic Basis for Gestalt: The work offers a potential mechanistic explanation for how biological systems might achieve Gestalt capabilities, suggesting that complex cognitive functions could emerge from the interaction of simpler, evolved neural components.

Key Results

The paper reports that empirical constructions and digital simulations have supported the theoretical foundation of their approach. While specific quantitative results (like accuracy percentages) are not detailed, the authors state that:

  • The system exhibits the ability to categorize two-dimensional geometric patterns directly by their shape.
  • Invariant ratios of output signals were observed, which are distinctive for different patterns and remain invariant under linear and rotational transformations, as well as size and shading changes.
  • The digital simulation successfully modeled the behavior of these logical neuron arrays, with processing times for a 9x9 matrix around 1.2 minutes (card input) or 30 seconds (direct CRT input).
  • About 50 patterns were tested using the simulation, with results generally matching expectations.

Limitations and Future Work

The authors acknowledge that the work is in its preliminary stages.

Limitations: * The findings are preliminary, and further work is needed to fully explore the scope and potential of the system. * The paper mentions that ambiguities can occur, especially with multiple patterns, and that the system does not possess the same ambiguity-separation capability as the human brain. * The initial exploration of diagonal neighbors and weighted connections yielded inconclusive results.

Future Work: * Further investigation into other predetermined neural-like functions beyond the logical comparator. * Exploring a wider range of interconnection schemes and their impact on pattern recognition. * Developing more sophisticated detectors for logical reduction of the array's output. * Refining the digital simulation and conducting more extensive testing with a larger variety of patterns and configurations. * Investigating the implementation of these functions using microelectronic techniques, magnetics, or cryotronic devices. * The authors express optimism that this work will lead to useful applications in areas like map reading, fingerprint analysis, and optical scanning.

Summary updated

BibTeX
@inproceedings{1969gestalt,
  title = {{Gestalt Pattern Recognition with Arrays of Predetermined Neural Functions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1969},
}
Figure