Pantypes: Diverse Representatives for Self-Explainable Models

TL;DR

Pantypes are introduced, a new family of prototypical objects designed to capture the full diversity of the input distribution through a sparse set of objects that can empower prototypical self-explainable models by occupying divergent regions of the latent space and thus fostering high diversity, interpretability and fairness.

Abstract

Prototypical self-explainable classifiers have emerged to meet the growing demand for interpretable AI systems. These classifiers are designed to incorporate high transparency in their decisions by basing inference on similarity with learned prototypical objects. While these models are designed with diversity in mind, the learned prototypes often do not sufficiently represent all aspects of the input distribution, particularly those in low density regions. Such lack of sufficient data representation, known as representation bias, has been associated with various detrimental properties related to machine learning diversity and fairness. In light of this, we introduce pantypes, a new family of prototypical objects designed to capture the full diversity of the input distribution through a sparse set of objects. We show that pantypes can empower prototypical self-explainable models by occupying divergent regions of the latent space and thus fostering high diversity, interpretability and fairness.

Figures (5)

• Figure 1: ProtoVAE (a) and PanVAE (b) visualizations of the latent space and decoded prototypes learned on MNIST after 30 epochs of training. Top: UMAP representations of the latent space with learned prototypes overlaid as squares. Bottom: Decoded prototypes of class '1' and '9'. One of the prototypes from PanVAE does not have the maximal similarity for any training image, indicated by a red cross. PanVAE has captured variations in the digit '1' pertaining to right-handedness (first '1' from the left), left-handedness (second '1' from the left) and a traditional writing style (third '1' from the left).
• Figure 2: Diversity control enabled by ProtoVAE and PanVAE. The figure shows the change in decoded prototype appearance as the respective diversity inducing losses are increased. The prototypes are shown for the FMNIST data of classes "sneaker", "bag" and "ankle boot" after 10 epochs of training. Figs. \ref{['fig:DivControl_orth1']} and \ref{['fig:DivControl_orth100']} show the difference between ProtoVAE prototypes with scale factor of 1 and 100 on the diversity loss $\mathcal{L}_\mathrm{orth}$. Figs. \ref{['fig:DivControl_vol1']} and \ref{['fig:DivControl_vol100']} show the difference between PanVAE pantypes with scale factor of 1 and 100 on the diversity loss $\mathcal{L}_\mathrm{vol}$.
• Figure 3: Evolution of prototype DB scores for PanVAE and ProtoVAE on MNIST, FMNIST and QuickDraw. Data points indicate mean values and associated standard deviations over three runs.
• Figure 4: Prototype coverage in UMAP space from 20 epochs of training on FMNIST with 5 prototypes for the "bag" class for ProtoVAE (a) and PanVAE (b). Top: UMAP representations of the latent space with learned prototypes overlaid as red squares. The prototype convex hull in UMAP space is shown as a red outline around the prototypes and the full class dataspace convex hull is shown as a blue outline around the data. A sample of the 100 closest observations to each prototype is shown as black datapoints. The convex hull of the sampled observations is shown as a black outline. The PanVAE sample convex hull covers 77% of the volume of the full class convex hull, whereas the ProtoVAE sample convex hull covers 33%. Bottom: Decoded prototypes.
• Figure 5: Face prototypes learned on the UTK Face dataset. The learned prototypes are shown for ProtoVAE in (a) and for PanVAE in (b). PanVAE has captured variations in race as well as other unseen features such as facial hair in males. The ProtoVAE males all have somewhat neutral expressions with shut mouths while most of the females have slight smiles. The PanVAE males and females all exhibit large variations in expression from full smiles with visible teeth to neutral expressions without visible teeth.