Chapter 15. Human-Centered Concept Explanations for Neural Networks

Understanding complex machine learning models such as deep neural networks with explanations is crucial in various applications. Many explanations stem from the model perspective, and may not necessarily effectively communicate why the model is making its predictions at the right level of abstraction. For example, providing importance weights to individual pixels in an image can only express which parts of that particular image is important to the model, but humans may prefer an explanation which explains the prediction by concept-based thinking. In this work, we review the emerging area of concept based explanations. We start by introducing concept explanations including the class of Concept Activation Vectors (CAV) which characterize concepts using vectors in appropriate spaces of neural activations, and discuss different properties of useful concepts, and approaches to measure the usefulness of concept vectors. We then discuss approaches to automatically extract concepts, and approaches to address some of their caveats. Finally, we discuss some case studies that showcase the utility of such concept-based explanations in synthetic settings and real world applications.

Concept Discovery for The Interpretation of Landscape Scenicness

In this paper, we study how to extract visual concepts to understand landscape scenicness. Using visual feature representations from a Convolutional Neural Network (CNN), we learn a number of Concept Activation Vectors (CAV) aligned with semantic concepts from ancillary datasets. These concepts represent objects, attributes or scene categories that describe outdoor images. We then use these CAVs to study their impact on the (crowdsourced) perception of beauty of landscapes in the United Kingdom. Finally, we deploy a technique to explore new concepts beyond those initially available in the ancillary dataset: Using a semi-supervised manifold alignment technique, we align the CNN image representation to a large set of word embeddings, therefore giving access to entire dictionaries of concepts. This allows us to obtain a list of new concept candidates to improve our understanding of the elements that contribute the most to the perception of scenicness. We do this without the need for any additional data by leveraging the commonalities in the visual and word vector spaces. Our results suggest that new and potentially useful concepts can be discovered by leveraging neighbourhood structures in the word vector spaces.

NRF2 regulates core and stabilizing circadian clock loops, coupling redox and timekeeping in Mus musculus

Diurnal oscillation of intracellular redox potential is known to couple metabolism with the circadian clock, yet the responsible mechanisms are not well understood. We show here that chemical activation of NRF2 modifies circadian gene expression and rhythmicity, with phenotypes similar to genetic NRF2 activation. Loss of Nrf2 function in mouse fibroblasts, hepatocytes and liver also altered circadian rhythms, suggesting that NRF2 stoichiometry and/or timing of expression are important to timekeeping in some cells. Consistent with this concept, activation of NRF2 at a circadian time corresponding to the peak generation of endogenous oxidative signals resulted in NRF2-dependent reinforcement of circadian amplitude. In hepatocytes, activated NRF2 bound specific enhancer regions of the core clock repressor gene Cry2, increased Cry2 expression and repressed CLOCK/BMAL1-regulated E-box transcription. Together these data indicate that NRF2 and clock comprise an interlocking loop that integrates cellular redox signals into tissue-specific circadian timekeeping.