Search

Increased explainability in machine learning is traditionally associated with lower performance, e.g. a decision tree is more explainable, but less accurate than a deep neural network. We argue that, in fact, increasing the explainability of a deep classifier can improve its generalization. In this chapter, we survey a line of our published work that demonstrates how spatial and spatiotemporal visual explainability can be obtained, and how such explainability can be used to train models that generalize better on unseen in-domain and out-of-domain samples, refine fine-grained classification predictions, better utilize network capacity, and are more robust to network compression.

By thoroughly revisiting the classic human action recognition paradigm, we analyzed different training/testing strategies, discovering that standard (cross-validating) testing strategies are not always the suitable validation procedures to assess an algorithm’s performance. As a consequence, we design a novel action recognition architecture, applying a “personalized” strategy to learn how any subject performs any action. We discover that it is advantageous to customize (i.e., personalize) the method to learn the actions carried out by each subject, rather than trying to generalize the actions executions across subjects. Leveraging on that, we propose an action recognition framework consisting of a two-stage classification approach where, given a test action, the subject is first identified before the actual recognition of the action takes place. Despite the basic, off-the-shelf descriptors and standard classifiers adopted, we score a favorable performance with respect to the state-of-the-art as to certify the soundness of our approach.

We present maps of the three-dimensional density (ρ), electrical conductivity (σ), and shear-wave speed (V S) structure of the mantle beneath Australia and surrounding ocean in the depth range of 100–800 km. These maps derived from stochastic inversion of seismic surface-wave dispersion data, thermodynamic modeling of mantle mineral phase equilibria, and laboratory-based conductivity models. Because composition and temperature act as fundamental parameters, we obtain naturally scaled maps of shear-wave speed, density, and electrical conductivity that depend only on composition, physical conditions (pressure and temperature), and laboratory measurements of the conductivity of anhydrous mantle minerals. The maps show that in the upper mantle ρ, σ and V S follow the continental-tectonic division that separates the older central and western parts of Australia from the younger eastern part. The lithosphere beneath the central and western cratonic areas appears to be relatively cold and Fe-depleted, and this is reflected in fast shear-wave speeds, high densities, and low conductivities. In contrast, the lithosphere underneath younger regions is relatively hot and enriched with Fe , which is manifested in slow shear-wave speeds, low densities, and high conductivities. This trend appears to continue to depths well below 300 km. The slow-fast shear-wave speed distribution found here is also observed in independent seismic tomographic models of the Australian region, whereas the coupled slow-fast shear-wave speed, low-high density, and high-low electrical conductivity distribution has not been observed previously. Toward the bottom of the upper mantle at 400 km depth marking the olivine → wadsleyite transformation (the “410-km” seismic discontinuity), the correlation between V S, ρ, and σ weakens. In the transition zone, V S, ρ, and σ are much less correlated indicating a significant compositional contribution to lateral heterogeneity. In particular, in the lower transition zone, σ and ρ appear to be governed mostly by variations in Fe/(Fe + Mg), whereas lateral variations in V S result from changes in (Mg + Fe)/Si and not, as observed in the upper mantle, from temperature variations. Lower mantle lateral variations in thermochemical parameters appear to smooth out, which suggests a generally homogeneous lower mantle in agreement with seismic tomographic images of the lower mantle. As a test of the regional surface-wave-based conductivity model, we computed magnetic fields of 24 h S q variations and compared these to observations. The comparison shows that while our predicted conductivity model improves the fit to observations relative to a one-dimensional model, amplitudes of the computed conductivity anomalies appear not to be large enough to enable these to be discriminated at present.