Face Generation and Editing With StyleGAN: A Survey.

Our goal with this survey is to provide an overview of the state of the art deep learning methods for face generation and editing using StyleGAN. The survey covers the evolution of StyleGAN, from PGGAN to StyleGAN3, and explores relevant topics such as suitable metrics for training, different latent representations, GAN inversion to latent spaces of StyleGAN, face image editing, cross-domain face stylization, face restoration, and even Deepfake applications. We aim to provide an entry point into the field for readers that have basic knowledge about the field of deep learning and are looking for an accessible introduction and overview.

Decentralized control and local information for robust and adaptive decentralized Deep Reinforcement Learning.

Decentralization is a central characteristic of biological motor control that allows for fast responses relying on local sensory information. In contrast, the current trend of Deep Reinforcement Learning (DRL) based approaches to motor control follows a centralized paradigm using a single, holistic controller that has to untangle the whole input information space. This motivates to ask whether decentralization as seen in biological control architectures might also be beneficial for embodied sensori-motor control systems when using DRL. To answer this question, we provide an analysis and comparison of eight control architectures for adaptive locomotion that were derived for a four-legged agent, but with their degree of decentralization varying systematically between the extremes of fully centralized and fully decentralized. Our comparison shows that learning speed is significantly enhanced in distributed architectures-while still reaching the same high performance level of centralized architectures-due to smaller search spaces and local costs providing more focused information for learning. Second, we find an increased robustness of the learning process in the decentralized cases-it is less demanding to hyperparameter selection and less prone to becoming trapped in poor local minima. Finally, when examining generalization to uneven terrains-not used during training-we find best performance for an intermediate architecture that is decentralized, but integrates only local information from both neighboring legs. Together, these findings demonstrate beneficial effects of distributing control into decentralized units and relying on local information. This appears as a promising approach towards more robust DRL and better generalization towards adaptive behavior.

Using Tactile Sensing to Improve the Sample Efficiency and Performance of Deep Deterministic Policy Gradients for Simulated In-Hand Manipulation Tasks.

Deep Reinforcement Learning techniques demonstrate advances in the domain of robotics. One of the limiting factors is a large number of interaction samples usually required for training in simulated and real-world environments. In this work, we demonstrate for a set of simulated dexterous in-hand object manipulation tasks that tactile information can substantially increase sample efficiency for training (by up to more than threefold). We also observe an improvement in performance (up to 46%) after adding tactile information. To examine the role of tactile-sensor parameters in these improvements, we included experiments with varied sensor-measurement accuracy (ground truth continuous values, noisy continuous values, Boolean values), and varied spatial resolution of the tactile sensors (927 sensors, 92 sensors, and 16 pooled sensor areas in the hand). To facilitate further studies and comparisons, we make these touch-sensor extensions available as a part of the OpenAI Gym Shadow-Dexterous-Hand robotics environments.

The World as an External Memory: The Price of Saccades in a Sensorimotor Task.

Theories of embodied cognition postulate that the world can serve as an external memory. This implies that instead of storing visual information in working memory the information may be equally retrieved by appropriate eye movements. Given this assumption, the question arises, how we balance the effort of memorization with the effort of visual sampling our environment. We analyzed eye-tracking data in a sensorimotor task where participants had to produce a copy of a LEGO®-blocks-model displayed on a computer screen. In the unconstrained condition, the model appeared immediately after eye-fixation on the model. In the constrained condition, we introduced a 0.7 s delay before uncovering the model. The model disappeared as soon as participants made a saccade outside of the Model Area. To successfully copy a model of 8 blocks participants made saccades to the Model Area on average 7.9 times in the unconstrained condition and 5.2 times in the constrained condition. However, the mean duration of a trial was 2.9 s (14%) longer in the constrained condition even when taking into account the delayed visibility of the model. In this study, we found evidence for an adaptive shift in subjects' behavior toward memorization by introducing a price for a certain type of saccades. However, the response is not adaptive; it is maladaptive, as memorization leads to longer overall performance time.

EEG correlates of sensorimotor processing: independent components involved in sensory and motor processing.

Sensorimotor processing is a critical function of the human brain with multiple cortical areas specialised for sensory recognition or motor execution. Although there has been considerable research into sensorimotor control in humans, the steps between sensory recognition and motor execution are not fully understood. To provide insight into brain areas responsible for sensorimotor computation, we used complex categorization-response tasks (variations of a Stroop task requiring recognition, decision-making, and motor responses) to test the hypothesis that some functional modules are participating in both sensory as well as motor processing. We operationalize functional modules as independent components (ICs) yielded by an independent component analysis (ICA) of EEG data and measured event-related responses by means of inter-trial coherence (ITC). Our results consistently found ICs with event-related ITC responses related to both sensory stimulation and motor response onsets (on average 5.8 ICs per session). These findings reveal EEG correlates of tightly coupled sensorimotor processing in the human brain, and support frameworks like embodied cognition, common coding, and sensorimotor contingency that do not sequentially separate sensory and motor brain processes.

Systems, Subjects, Sessions: To What Extent Do These Factors Influence EEG Data?

Lab-based electroencephalography (EEG) techniques have matured over decades of research and can produce high-quality scientific data. It is often assumed that the specific choice of EEG system has limited impact on the data and does not add variance to the results. However, many low cost and mobile EEG systems are now available, and there is some doubt as to the how EEG data vary across these newer systems. We sought to determine how variance across systems compares to variance across subjects or repeated sessions. We tested four EEG systems: two standard research-grade systems, one system designed for mobile use with dry electrodes, and an affordable mobile system with a lower channel count. We recorded four subjects three times with each of the four EEG systems. This setup allowed us to assess the influence of all three factors on the variance of data. Subjects performed a battery of six short standard EEG paradigms based on event-related potentials (ERPs) and steady-state visually evoked potential (SSVEP). Results demonstrated that subjects account for 32% of the variance, systems for 9% of the variance, and repeated sessions for each subject-system combination for 1% of the variance. In most lab-based EEG research, the number of subjects per study typically ranges from 10 to 20, and error of uncertainty in estimates of the mean (like ERP) will improve by the square root of the number of subjects. As a result, the variance due to EEG system (9%) is of the same order of magnitude as variance due to subjects (32%/sqrt(16) = 8%) with a pool of 16 subjects. The two standard research-grade EEG systems had no significantly different means from each other across all paradigms. However, the two other EEG systems demonstrated different mean values from one or both of the two standard research-grade EEG systems in at least half of the paradigms. In addition to providing specific estimates of the variability across EEG systems, subjects, and repeated sessions, we also propose a benchmark to evaluate new mobile EEG systems by means of ERP responses.