Gerd Ascheid

This work focuses on using Deep Reinforcement Learning (DRL) to control an autonomous vehicle in the hyper-realistic urban simulation LGSVL. Classical control systems such as MPC maneuver vehicles based on a given trajectory, current velocity, position, distances, and more. Our approach does not pass this information to the DRL agent but only images provided by the camera. Current DRL efforts also exploit similar approaches for autonomous driving, but they are only suitable for small, simple tasks using simple simulations. Our approach consists of two differently trained neural networks (NN), a perceptual NN for representation learning and an actor NN for selecting the correct action. The perception NN will be trained via representation and self-supervised learning to strengthen our DRL agent's understanding of the scene. It can recognize temporal information and the dynamics of a complex environment. This work shows the importance of decoupling the perception and decision (actor) model for autonomous driving. All in all, we could drive autonomously in a hyper-realistic urban simulation using our modular DRL framework. Moreover, our approach also provides a solution for other similar tasks in the field of robotics based on images.

Machine Learning (ML) techniques are applied to solve automation tasks in more and more scenarios of our daily life. For instance, Keyword Spotting applications are of interest for Internet-of-things devices in a smart home or for human-machine interaction. Different ML techniques, such as Neural Networks including Long-Short Term Memories, recently achieve reasonable accuracy above 95 % [1]. Therefore, we propose an accelerator for the Keyword spotter as a specialized coprocessor in order to retain the flexibility to fit different network structures, while the energy efficiency is increased.

Autonomous vehicles have become a hot topic in recent years due to advances in deep learning based signal processing, computational and sensor performance, and connectivity. A common way to classify the autonomy of the vehicle is the 6-level scale defined by the SAE International [15]. Traditional vehicles with no autonomous or support systems are designated a driving automation level of 0, while a fully autonomous vehicle, capable of navigating in all circumstances and not requiring human supervision, corresponds to level 5. While support systems, such as lane following, parking assistance or systems keeping a safe distance from leading vehicles are developed as standalone comfort systems, autonomous vehicles above level 3 require a holistic approach to the topic. Today, even the most advanced autonomous vehicles are not reaching level 5 autonomy.

In this letter, the probability of reaching consensus in wireless multi-agent systems (WMASs) is introduced. Starting with the connectivity probability of random graphs, and considering link outages between agents, we obtain the probability expression that a WMAS with N agents reaches consensus. We validate the obtained exact expression via simulation results. Furthermore, we show that the corresponding diversity order is N -1, hence increasing the number of agents will contribute towards the likelihood of consensus, despite the increased communication need between each agent. Finally, we observe that the consensus speed is highly dependent on the signal-to-noise ratio (SNR) level and the number of agents in a WMAS.

The aim of this work was to develop and evaluate the reinforcement learning algorithm VentAI, which is able to suggest a dynamically optimized mechanical ventilation regime for critically-ill patients. We built, validated and tested its performance on 11,943 events of volume-controlled mechanical ventilation derived from 61,532 distinct ICU admissions and tested it on an independent, secondary dataset (200,859 ICU stays; 25,086 mechanical ventilation events). A patient “data fingerprint” of 44 features was extracted as multidimensional time series in 4-hour time steps. We used a Markov decision process, including a reward system and a Q-learning approach, to find the optimized settings for positive end-expiratory pressure (PEEP), fraction of inspired oxygen (FiO₂) and ideal body weight-adjusted tidal volume (Vt). The observed outcome was in-hospital or 90-day mortality. VentAI reached a significantly increased estimated performance return of 83.3 (primary dataset) and 84.1 (secondary dataset) compared to physicians’ standard clinical care (51.1). The number of recommended action changes per mechanically ventilated patient constantly exceeded those of the clinicians. VentAI chose 202.9% more frequently ventilation regimes with lower Vt (5–7.5 mL/kg), but 50.8% less for regimes with higher Vt (7.5–10 mL/kg). VentAI recommended 29.3% more frequently PEEP levels of 5–7 cm H₂O and 53.6% more frequently PEEP levels of 7–9 cmH₂O. VentAI avoided high (>55%) FiO₂ values (59.8% decrease), while preferring the range of 50–55% (140.3% increase). In conclusion, VentAI provides reproducible high performance by dynamically choosing an optimized, individualized ventilation strategy and thus might be of benefit for critically ill patients.

ConvAix is an application-specific instruction-set processor (ASIP) that enables the energy-efficient processing of convolutional neural networks (CNNs) while retaining substantial flexibility through its instruction-set architecture (ISA) based design. By utilizing a combination of data-level parallelism (DLP), instruction-level parallelism (ILP), and subword parallelism, the proposed design offers sufficient processing power for the execution of state-of-the-art CNNs in real-time. ConvAix’s arithmetic logic units (ALUs) are C-programmable, thereby offering the degree of flexibility required to implement many different convolution layer types, e.g. depthwise-separable convolutions and residual blocks, as well as fully-connected and pooling layers. It comprises a total of 256ALUs and leverages low-precision computations down to 4bits. Furthermore, it exploits sparsity in feature maps and weights via zero-guarding of redundant computations to maximize its energy efficiency. The processor was implemented in a modern 28nm CMOS technology operating at 1V supply voltage with a resulting clock frequency of 513MHz. The final design offers a precision-dependent peak throughput between 263GOP/s (int16) and 1.1TOP/s (int4), while consuming between 972mW and 340mW of power, resulting in effective energy-efficiencies ranging from 176GOP/s/W to 2TOP/s/W. Well-known CNNs, such as AlexNet, MobileNet, and ResNet-18, are simulated based on the placed and routed netlist, achieving between 233 (AlexNet) and 69 (ResNet-18) frames-per-second for a batch-size of 1, including times for off-chip transfers.

Cyber-physical systems (CPS) are state-of-the-art communication environments that offer various applications with distinct requirements. However, security in CPS is a nonnegotiable concept, since without a proper security mechanism the applications of CPS may risk human lives, the privacy of individuals, and system operations. In this paper, we focus on PHY-layer security approaches in CPS to prevent passive eavesdropping attacks, and we propose an integration of physical layer operations to enhance security. Thanks to the McEliece cryptosystem, error injection is firstly applied to information bits, which are encoded with the forward error correction (FEC) schemes. Golay and Hamming codes are selected as FEC schemes to satisfy power and computational efficiency. Then obtained codewords are transmitted across reliable intermediate relays to the legitimate receiver. As a performance metric, the decoding frame error rate of the eavesdropper is analytically obtained for the fragmentary existence of significant noise between relays and Eve. The simulation results validate the analytical calculations, and the obtained results show that the number of low-quality channels and the selected FEC scheme affects the performance of the proposed model.

The lack of technical resources and the high cost of establishing a semiconductor foundry has forced most integrated circuit design houses to rely on outsourcing part of their design and fabrication services to off-site companies. The involvement of external parties has given rise to major security threats, ranging from intellectual property piracy to the insertion of malicious circuits. Logic encryption has emerged as a popular mitigation technique against these threats. In recent years, a vast amount of logic encryption algorithms has been proposed. However, existing approaches strongly focus on isolated circuit components without taking the complexity and modular structure of modern circuit designs into account. In this paper, we propose Inter-Lock, a novel logic encryption framework tailored towards scaling logic encryption to larger designs by leveraging their complexity and exploiting multi-module interdependencies to exponentially enhance the security of sequential circuits. To showcase the applicability of the approach, we present an extensive evaluation on a real-life 32-bit RISC-V core together with the analysis of the security-cost trade-off. Our recommended Inter-Lock configuration of the core features a 1024-bit key and 100% functional corruption for 13.2% area overhead, less than 1% power overhead and 22.2% delay penalty in the worst case.