A size-reduced InP-based inverse-designed polarization rotator splitter is presented, being the first demonstration of a topology optimized passive waveguide component on InP. The manufactured device has extinction ratios >10dB for both polarizations over the entire C-band.
Communication constraints prevent the wide-spread adoption of Federated Learning systems. In this work, we investigate Federated Distillation (FD) from the perspective of communication efficiency by analyzing the effects of active distillation-data curation, soft-label quantization, and delta-coding techniques. We present Compressed Federated Distillation (CFD), an efficient FD method.
In collaboration with VI Systems and the University of Warsaw, HHI has demonstrated data rates of over 200 Gbit/s for the optical short haul. This was made possible by a new generation of VCSELs with over 30 GHz bandwidth in combination with multimode fibres.
LiFi exhibits characteristics that make it highly suitable as a wireless communication technology for industrial applications. As such, LiFi provides deterministic propagation, the potential for very high data density, use of license-free spectrum, and added physical layer security. In this work, we assess the requirements and challenges for LiFi in industrial wireless networks and propose to use distributed MIMO techniques to make LiFi capable of mastering them. We sketch essential capabilitie! s of the protocol- as well as physical layer and provide an outlook on the upcoming IEEE Std 802.15.13
Predicting the Binding of SARS-CoV-2 Peptides to the Major Histocompatibility Complex with Recurrent Neural Networks
Predicting the binding of viral peptides to the major histocompatibility complex with machine learning can potentially extend the computational immunology toolkit for vaccine development, and serve as a key component in the fight against a pandemic. In this work, we adapt and extend USMPep, a recently proposed, conceptually simple prediction algorithm based on recurrent neural networks. Most notably, we combine regressors (binding affinity data) and classifiers (mass spectrometry data) from qualitatively different data sources to obtain a more comprehensive prediction tool. We evaluate the performance on a recently released SARS-CoV-2 dataset with binding stability measurements. USMPep not only sets new benchmarks on selected single alleles, but consistently turns out to be among the best-performing methods or, for some metrics, to be even the overall best-performing method for this task.
With the goal of enabling optical wireless communications for mobile devices, we assess a physical layer based on high-bandwidth on-off keying modulation. This allows for amplifier designs that avoid operation in a resistive mode, reducing their energy usage to a fraction. Link-level simulations show that the investigated physical layer can deal with typical frontend limitations and can operate in challenging non-line-of-sight channels.
We present a hybrid animation approach that combines example-based and neural animation methods to create a simple, yet powerful animation regime for human faces. We introduce a light-weight auto-regressive network to transform our animation-database into a parametric model. During training, our network learns the dynamics of facial expressions, which enables the replay of annotated sequences from our animation database as well as their seamless concatenation in new order.
Radiation pattern of planar optoelectronic antennas for broadband continuous-wave terahertz emission
We measured and simulated the radiation pattern of continuous-wave terahertz emitters between 100 and 500 GHz. We could improve the radiation pattern by optimizing the connection between terahertz source and antenna: Unwanted side lobes were reduced by more than 10 dB and the beam angle was narrowed to 9° at 300 GHz. These improvements are relevant for wireless communication links and lensless terahertz imaging.
VR becomes popular in neurological and rehabilitation assessments and exercises for controlled simulation of complex environments that are difficult to setup physically in a laboratory. For such tasks, VR systems have to meet higher requirements than for an entertainment setup. The system needs to be suitable for sensitive/restricted users without limiting their VR interactions. To minimize the difference to the real-world interaction fidelity is desirable. So besides finding alternative metaphors for natural interaction, versatile research is done to en! able natural interaction in VR. For lack of better natural locomotion alternatives, teleportation is currently the most commonly used locomotion mechanism in VR. However, teleportation is not suitable for VR applications simulating reality, e.g., in search tasks during neurological tests.
This paper provides a timely overview of the field of explainable artificial intelligence (XAI). It explains the theoretical foundations of interpretability algorithms, outlines best practice aspects, demonstrates successful usage of XAI in selected application scenarios, and discusses challenges and possible future directions of this active emerging field.
A new concept for spatially resolved coherent detection with vertically illuminated photodetectors targeting ranging applications
This paper proposes a novel approach for coherent detection with double-side vertically illuminated photodetectors. Signal and local oscillator are injected collinearly from opposite sides of the photodetector chip. The concept inherently provides angular selective detection and can be used for developing compact, solid-state receiver modules for coherent light detection and ranging (LiDAR).
This paper gives a comprehensive overview over classic shallow and novel deep approaches to anomaly detection. We identify the common underlying principles and provide an empirical assessment of major existing methods that are enriched by the use of recent explainability techniques. We present specific worked-through examples together with practical advice and discuss open challenges.
This paper proposes a novel criterion for CNN pruning inspired by neural network interpretability: The most relevant units, i.e. weights or filters, are automatically found using their relevance scores obtained from concepts of explainable AI (XAI). By exploring this idea, we connect the lines of interpretability and model compression research.
Time-domain spectroscopy with terahertz frequencies typically requires complex and bulky systems. Here, the authors present an optoelectronics-based, frequency-domain terahertz sensing technique which offers competitive measurement performance in a much simpler system.
Inferring respiratory and circulatory parameters from electrical impedance tomography with deep recurrent models
Electrical impedance tomography (EIT) is a non-invasive imaging modality that allows a continuous assessment of changes in regional bioimpedance of different organs. One of its most common biomedical applications is monitoring regional ventilation distribution in critically ill patients treated in intensive care units. In this work, we put forward a proof-of-principle study that demonstrates how one can reconstruct synchronously measured respiratory or circulatory parameters from the EIT image sequence using a deep learning model trained in an end-to-end fashion. For this purpose, we devise an architecture with a convolutional feature extractor whose output is processed by a recurrent neural network. We demonstrate that one can accurately infer absolute volume, absolute flow, normalized airway pressure and within certain limitations even the normalized arterial blood pressure from the EIT signal alone, in a way that generalizes to unseen patients without prior calibration. As an outlook with direct clinical relevance, we furthermore demonstrate the feasibility of reconstructing the absolute transpulmonary pressure from a combination of EIT and absolute airway pressure, as a way to potentially replace the invasive measurement of esophageal pressure. With these results, we hope to stimulate further studies building on the framework put forward in this work.
Externally modulated DFB lasers (EML) and vertically illuminated photodetectors are presented. Because of their excellent high-speed behavior and operation wavelength of 1310 nm, the devices are of interest for intra-data center communication. Since the EML and the photodetector chips are compatible with current systems, these devices are candidates for upgrading existing transceivers to higher baud rates. Therefore, a proof of concept for 100 GBaud data transmission with the presented components is demonstrated. Even without predistor! tion, the experiments show clearly open eye diagrams.
This paper proposes a novel, simple yet effective defense strategy for adversarial attacks on deep learning models. Our algorithm, called MALA for DEfense (MALADE), is applicable to any existing classifier, providing robust defense as well as off-manifold sample detection. In our experiments, MALADE exhibited state-of-the-art performance against various elaborate attacking strategies.
Inverse design is rapidly gaining popularity for automated design of photonic components. Two methods to improve it for large passive waveguide structures are developed: Adaptive Threshold Binarization and Hybrid Optimization. To demonstrate their capability, inverse design is applied to an InP waveguide platform for the first time. As an example, a polarizer with a PER of -19.4 dB is presented.
In this paper we present an example-based approach for visual speech synthesis. We combine the advantages of deep generative models and classical animation approaches to create a real-time capable facial animation framework based on volumetric captures.
Second harmonic generation is an efficient way to create coherent radiation at wavelengths that are not accessible with standard laser sources. In this work we demonstrate second harmonic generation from 1550 nm to 775 nm in a polymer photonic integrated circuit via the hybrid integration of a periodically poled lithium niobate crystal. The bulk crystal is inserted in an on-chip free-space section between two waveguide couple! d GRIN lenses. Fiber to fiber conversion efficiencies were 0.03 %/W for a continuous wave laser source and 100 %/W for a femtosecond laser source. Furthermore, third and fourth harmonic light at 517 nm and 388 nm was observed.