In this work, we address the task of unconditional head motion generation to animate still human faces in a low-dimensional semantic space from a single reference pose. Different from traditional audio-conditioned talking head generation that seldom puts emphasis on realistic head motions, we devise a GAN-based architecture that learns to synthesize rich head motion sequences over long duration while maintaining low error-accumulation levels. In particular, the autoregressive generation of incremental outputs ensures smooth trajectories, while a multi-scale discriminator on input pairs drives generation toward better handling of high- and low-frequency signals and less mode collapse. We experimentally demonstrate the relevance of the proposed method and show its superiority compared to models that attained state-of-the-art performances on similar tasks.

Convolutional neural networks (CNNs) have recently achieved outstanding performance for hyperspectral (HS) and multispectral (MS) image fusion. However, CNNs cannot explore the long-range dependence for HS and MS image fusion because of their local receptive fields. To overcome this limitation, a transformer is proposed to leverage the long-range dependence from the network inputs. Because of the ability of long-range modeling, the transformer overcomes the sole CNN on many tasks, whereas its use for HS and MS image fusion is still unexplored. In this article, we propose a spectral-spatial transformer (SST) to show the potentiality of transformers for HS and MS image fusion. We devise first two branches to extract spectral and spatial features in the HS and MS images by SST blocks, which can explore the spectral and spatial long-range dependence, respectively. Afterward, spectral and spatial features are fused feeding the result back to spectral and spatial branches for information interaction. Finally, the high-resolution (HR) HS image is reconstructed by dense links from all the fused features to make full use of them. The experimental analysis demonstrates the high performance of the proposed approach compared with some state-of-the-art (SOTA) methods.

Full-waveform inversion (FWI) in the North Sea has demonstrated its imaging power starting from low-resolution models obtained by traveltime tomography, enriching them with geologically interpretable fine-scale details. However, building a traveltime-based kinematically accurate starting model for FWI is a time-consuming and rather subjective process requiring phase identification and selection. The two main problems faced by FWI starting from noninformative initial models are the susceptibility to cycle skipping and a lack of sensitivity to low wavenumbers in the deep subsurface not sampled by turning waves. On a North Sea ocean-bottom cable 3D data set, a novel [Formula: see text] building methodology is applied that addresses those issues by jointly inverting reflections and refractions (joint full-waveform inversion [JFWI]) using a robust misfit function in the vertical traveltime domain (pseudotime). Pseudotime addresses reflectivity-velocity coupling and attenuates phase ambiguities at short offsets, whereas a graph-space optimal transport (GSOT) objective function with dedicated data windowing averts cycle skipping at intermediate-to-long offsets. A fast and balanced reflectivity reconstrution is obtained prior to JFWI thanks to an asymptotic-preconditioned impedance waveform inversion ([Formula: see text]WI). Starting from a linearly increasing one-dimensional model, GSOT-pseudotime JFWI is effective at obtaining a meaningful P-wave velocity macromodel down to depths sampled by reflections only, without phase identification and picking. P-wave FWI, starting from the JFWI-based model, injects the high wavenumbers missing in the JFWI solution, attaining apparent improvements in shallow and deep model reconstruction and imaging compared with the previous studies in the literature, and a satisfactory prediction of the ground-truth logs.

Extracting, processing, and delivering energy requires energy itself, which reduces the net energy available to society and yields considerable socioeconomic implications.

Regularization is a central topic in the study of the solutions of ill-posed inverse problems. High-resolution seismic imaging using full-waveform inversion (FWI) belongs to this category of problems. Regularization through anisotropic diffusion, a technique that emerged in the field of image processing, is an interesting alternative to conventional regularization strategies. Exploiting the structural information of a given image, it has the capability to smooth this image along its main structures. The main difficulty is how to design the anisotropic diffusion operator. The concept of coherence enhancing proposed in 2D is extended in 3D and applied so as to filter and enhance the structural coherence of the model updates within an FWI algorithm. The benefits of this strategy are investigated on a 2D synthetic experiment before considering the multiparameter inversion of a 3D field data set from the North Sea up to 10 Hz. From this data, the vertical velocity and the density are simultaneously reconstructed. Compared with a conventional nonstationary Gaussian regularization strategy, the models obtained using the coherence-enhancing anisotropic diffusion strategy indicate an enhanced resolution, especially for the density model. The high-resolution reflectivity image computed from the impedance volume clearly illustrates the benefit this filtering approach can deliver in terms of structural interpretation.

Abstract Neuroscience studies face challenges in gathering large datasets, which limits the use of machine learning (ML) approaches. One possible solution is to incorporate additional data from large public datasets; however, data collected in different contexts often exhibit systematic differences called dataset shifts. Various factors, for example, site, device type, experimental protocol, or social characteristics, can lead to substantial divergence of brain signals that can hinder the success of ML across datasets. In this work, we focus on dataset shifts in recordings of brain activity using MEG and EEG. State-of-the-art predictive approaches on magneto- and electroencephalography (M/EEG) signals classically represent the data by covariance matrices. Model-based dataset alignment methods can leverage the geometry of covariance matrices, leading to three steps: re-centering, re-scaling, and rotation correction. This work explains theoretically how differences in brain activity, anatomy, or device configuration lead to certain shifts in data covariances. Using controlled simulations, the different alignment methods are evaluated. Their practical relevance is evaluated for brain age prediction on one MEG dataset (Cam-CAN, n = 646) and two EEG datasets (TUAB, n = 1385; LEMON, n = 213). Among the same dataset (Cam-CAN), when training and test recordings were from the same subjects but performing different tasks, paired rotation correction was essential (δR2=+0.13 (rest-passive) or +0.17 (rest-smt)). When in addition to different tasks we included unseen subjects, re-centering led to improved performance (δR2=+0.096 for rest-passive, δR2=+0.045 for rest-smt). For generalization to an independent dataset sampled from a different population and recorded with a different device, re-centering was necessary to achieve brain age prediction performance close to within dataset prediction performance. This study demonstrates that the generalization of M/EEG-based regression models across datasets can be substantially enhanced by applying domain adaptation procedures that can statistically harmonize diverse datasets.

Transfer of learning is influenced by transfer distance, i.e., the degree of correspondence of contexts as well as of the content to be transferred between a main task and a transfer task (Barnett & Ceci, 2002; Klahr & Chen, 2011). A long tradition of research has used either similarity of tasks or similarity of contexts designs to test transfer distance in problem-solving in children. To our knowledge, no study to date has experimentally varied the degree of similarities between tasks and between contexts jointly, which is the aim of this study. In the present study, we varied the degree of similarity between tasks, by presenting different versions of the Tower of Hanoi (ToH), and between digital learning contexts, by presenting different graphical interfaces. One hundred children aged 6 to 7 years old were divided into four conditions, High similarity of tasks/High similarity of contexts (H/H); High similarity of tasks/Low similarity of contexts (H/L); Low similarity of tasks/High similarity of contexts (L/H); Low similarity of tasks/Low similarity of contexts (L/L). Inhibitory control was also measured, as it may be involved in transfer of learning (Clerc, Miller, & Cosnefroy, 2014; Thibaut, French, Vezneva, Gérard, & Glady, 2011). Transfer of the solving procedure was accompanied by a performance decrement in the L/L and H/L conditions, whereas no such decrement was observed in the L/H and H/H conditions, showing the role of the similarity of digital contexts. Furthermore, inhibition scores predicted transfer performance. Results are discussed considering the theoretical framework of similarities between tasks and between contexts, and the role of inhibition.

Human trade and movements are central to biological invasions worldwide. Human activities not only transport species across biogeographical barriers, but also accelerate their post‐introduction spread in the landscape. Thus, by constraining human movements, the spatial structure of road networks might greatly affect the regional spread of invasive species. However, few invasion models have accounted for the topology of road networks so far, and its importance for explaining the regional distribution of invasive species remains mostly unexplored. To address this issue, we developed a spatially explicit and mechanistic human‐mediated dispersal model that accounts and tests for the influence of transport networks on the regional spread of invasive species. Using as a model the spread of the invasive ant Lasius neglectus in the middle Rhône valley (France), we show that accounting for the topology of road networks improves our ability to explain the current distribution of the invasive ant. In contrast, we found that using human population density as a proxy for the frequency of transport events decreases models' performance and might thus not be as appropriate as previously thought. Finally, by differentiating road networks into sub‐networks, we show that national and regional roads are more important than smaller roads for explaining spread patterns. Overall, our results demonstrate that the topology of transport networks can strongly bias regional invasion patterns and highlight the importance of better incorporating it into future invasion models. The mechanistic modelling approach developed in this study should help invasion scientists explore how human‐mediated dispersal and topography shape invasion dynamics in landscapes. Ultimately, our approach could be combined with demographic, natural dispersal and environmental suitability models to refine spread scenarios and improve invasive species monitoring and management at regional to national scales.Keywords: biological invasions, human‐mediated dispersal, road network, secondary spread, spatially explicit model, stochastic jump model

This paper introduces a novel approach to the fine alignment of images in a burst captured by a handheld camera. In contrast to traditional techniques that estimate two-dimensional transformations between frame pairs or rely on discrete correspondences, the proposed algorithm establishes dense correspondences by optimizing both the camera motion and surface depth and orientation at every pixel. This approach improves alignment, particularly in scenarios with parallax challenges. Extensive experiments with synthetic bursts featuring small and even tiny baselines demonstrate that it outperforms the best optical flow methods available today in this setting, without requiring any training. Beyond enhanced alignment, our method opens avenues for tasks beyond simple image restoration, such as depth estimation and 3D reconstruction, as supported by promising preliminary results. This positions our approach as a versatile tool for various burst image processing applications.