Prototype Representations¶
Definition¶
Related to RBF Networks, but the main distinction is that they use any similarity measure \(k: \mathcal{X} \times \mathcal{X} \to \mathbb{R}\) to produce the feature:
where \(k\) is any similarity function.
Radial Basis Functions are an example of a similarity function.
Also, the centers \(\mathbf{c_{j}}\) are typically subselected from the training dataset; and are called prototypical instances In RBF Networks, choosing centers is an open step
More formally, we choose p centers from our dataset, without replacement, and if n (the number of samples in the dataset) is sufficiently small, then we choose p = n; otherwise, we choose p to be a diverse set. Here, diverse set means a set representative of the overall distribution.
References¶
Highly related to Radial Basis Function Networks (RBF)