Walking the Web of Concept-Class Relationships in Incrementally Trained Interpretable Models

TL;DR

This work proposes a novel method - MuCIL - that uses multimodal concepts to perform classification without increasing the number of trainable parameters across experiences, and obtains state-of-the-art classification performance compared to other concept-based models.

Abstract

Concept-based methods have emerged as a promising direction to develop interpretable neural networks in standard supervised settings. However, most works that study them in incremental settings assume either a static concept set across all experiences or assume that each experience relies on a distinct set of concepts. In this work, we study concept-based models in a more realistic, dynamic setting where new classes may rely on older concepts in addition to introducing new concepts themselves. We show that concepts and classes form a complex web of relationships, which is susceptible to degradation and needs to be preserved and augmented across experiences. We introduce new metrics to show that existing concept-based models cannot preserve these relationships even when trained using methods to prevent catastrophic forgetting, since they cannot handle forgetting at concept, class, and concept-class relationship levels simultaneously. To address these issues, we propose a novel method - MuCIL - that uses multimodal concepts to perform classification without increasing the number of trainable parameters across experiences. The multimodal concepts are aligned to concepts provided in natural language, making them interpretable by design. Through extensive experimentation, we show that our approach obtains state-of-the-art classification performance compared to other concept-based models, achieving over 2$\times$ the classification performance in some cases. We also study the ability of our model to perform interventions on concepts, and show that it can localize visual concepts in input images, providing post-hoc interpretations.

Figures (10)

• Figure 1: Illustration of our setting. Concepts introduced in an experience are shared among classes from other experiences. MuCIL focuses on the difficult challenge of capturing and preserving this web of class-concept relationships over multiple experiences.
• Figure 2: Overview of our setup and proposed architecture. Our architecture receives new classes and associated concepts across multiple experiences in a CL setting. We use pre-trained language and vision encoders to get embeddings of the input image, concepts, and classes. These are then used to create multimodal image-concept embeddings using our Multimodal Encoder. The multimodal concept embeddings are grounded to their concept anchors using a loss function and are used to predict both class labels and the presence/absence of corresponding concepts in the image.
• Figure 3: Visualization of Concept-Class Relationship Forgetting on ImageNet-100 across 10 experiences, for CBM and MuCIL. While the relationship between concepts and classes deteriorates over new experiences, our method maintains forgetting at the same level.
• Figure 4: Comparison of ACR matrices for (a) Standard CBM, (b) Label-Free CBM, and (c) MuCIL for CIFAR-100 over 10 experiences. Each row represents the ACR scores for the class sets of the corresponding experience, with each column representing the concept set of different experiences.
• Figure 5: Visual grounding of part-based concepts: Qualitative results for localizing visual concepts using MuCIL versus when localizing the same concepts using GradCAM on CBMs. More results provided in the Appendix (§ A3).