A Novel Template-Based Learning Model

This article presents a model which is capable of learning and abstracting new concepts based on comparing observations and finding the resemblance between the observations. In the model, the new observations are compared with the templates which have been derived from the previous experiences. In the first stage, the objects are first represented through a geometric description which is used for finding the object boundaries and a descriptor which is inspired by the human visual system and then they are fed into the model. Next, the new observations are identified through comparing them with the previously-learned templates and are used for producing new templates. The comparisons are made based on measures like Euclidean or correlation distance. The new template is created by applying onion-pealing algorithm. The algorithm consecutively uses convex hulls which are made by the points representing the objects. If the new observation is remarkably similar to one of the observed categories, it is no longer utilized in creating a new template. The existing templates are used to provide a description of the new observation. This description is provided in the templates space. Each template represents a dimension of the feature space. The degree of the resemblance each template bears to each object indicates the value associated with the object in that dimension of the templates space. In this way, the description of the new observation becomes more accurate and detailed as the time passes and the experiences increase. We have used this model for learning and recognizing the new polygons in the polygon space. Representing the polygons was made possible through employing a geometric method and a method inspired by human visual system. Various implementations of the model have been compared. The evaluation results of the model prove its efficiency in learning and deriving new templates.

💡 Research Summary

The paper introduces a template‑based learning framework that incrementally abstracts new concepts by comparing incoming observations with a set of previously learned templates. The authors first represent each object using a dual‑layer description: a geometric representation that captures object boundaries via point sets and convex hulls, and a visual descriptor inspired by the human visual system that encodes color, texture, and shape information into a high‑dimensional vector. These representations enable quantitative similarity assessment between any new observation and the existing template library.

Similarity is measured using standard distance metrics such as Euclidean distance or correlation distance. If the distance between a new observation and any existing template falls below a predefined threshold, the observation is considered a member of that category and is directly mapped into the template space without creating a new template. The template space is defined as a multi‑dimensional feature space where each dimension corresponds to a single template; the similarity value to a template becomes the coordinate of the observation along that dimension. Consequently, each object is represented as a point in this space, and as more templates are added over time, the representation becomes increasingly fine‑grained.

When an observation is not sufficiently similar to any existing template, the system generates a new template using an “onion‑peeling” algorithm. This algorithm iteratively extracts convex hulls from the point set of the observation: the outermost hull is computed first, then removed, and the process repeats on the remaining points until no points remain. Each extracted hull forms a layer that captures a hierarchical structural aspect of the shape. By concatenating these layers, the algorithm constructs a comprehensive template that preserves essential geometric characteristics while simplifying the overall representation.

The authors evaluate the model on a polygon‑recognition task. Polygons are encoded using the same geometric and visual descriptors, and a variety of transformed instances (rotations, scalings, noise) are presented to the system. Experimental results show that the template‑based model outperforms conventional classifiers such as k‑Nearest Neighbors and Support Vector Machines in both accuracy and learning speed. Moreover, because new templates are added only when necessary, the system supports incremental learning without significant performance degradation as the dataset grows.

Key contributions of the work include: (1) a hybrid object representation that fuses geometric boundary extraction with a biologically inspired visual descriptor; (2) an onion‑peeling convex‑hull algorithm for efficient, hierarchical template creation; (3) the formulation of a template space where each template defines a feature dimension, enabling continuous similarity‑based object encoding; and (4) a mechanism for seamless template expansion that facilitates lifelong learning. These innovations collectively advance the state of concept learning by providing a flexible, experience‑driven architecture capable of adapting to new categories while maintaining high recognition performance.