PhD-Seminar Signal Processing

Seminar Chair: Wolfgang Utschick

Target Audience: Ph.D. candidates in the area of signal processing, open for all interested TUM members

Language: English

Offered in: Winter Term and Summer Term

Schedule in Winter Semester 2024/25

Description

Ph.D. candidates, MSV research staff, and invited guests present current results about topics that are relevant for research at MSV.

Seminar on 2020-07-16 10:00 (Zoom Session)

Machine Learning for ECG analysis

(Congyu Lou, Klinikum Rechts der Isar, Technische Universität München)

Physicians use ECG to diagnose a variety of cardiac abnormalities. However, manual interpretation is time-consuming and requires expertise. Therefore, it is needed to have a fast, and accurate algorithm that can classify cardiac abnormalities automatically. In this presentation, some researches on applying machine learning and deep learning algorithms to do heartbeats and short-period recording classification will be introduced. Among them, three representative algorithms that have been reproduced will be explained in detail, and their performances will be compared. Besides that, ideas to improve the models and a novel approach will be discussed.  

Seminar on 2020-06-12 15:00 (Zoom Session)

Modeling Complex Channels with Generative Adversarial Models

(Amir Hossein Rezaei, Technische Universität München)

In this presentation, a fully differentiable neural network is presented. Probabilistic and geometric shaping of the constellation are joined to Generative Adversarial Network (GAN) and Mutual Information Neural Estimation (MINE) networks. Combining all in a new model is the essence of this talk. The suggested model operates pretty well over a range of high SNRs. This model could be used for an arbitrary channel unlike conventional communication system designs where knowing the channel model beforehand is a must. The main concepts deployed in this work are: 1) Gumbel-Softmax trick which enables us to work with discrete probabilities and have geometric and probabilistic shaping of the constellation, 2) MINE which helps us observe and measure the mutual information when the underlying distributions are not known, and 3) GAN which helps us to model an unknown channel. Leveraging all these components in a single design, as demonstrated in the simulation, helps us get very close to the lower bound of capacity for Rayleigh channel.

Seminar on 2020-03-26 14:00 (cancelled)

Low Complexity Data Processing and Learning over Single-Agent and Distributed Networks

(Mohammad Hussein Nassralla, American University of Beirut, Lebanon)

In distributed learning, the data are processed locally to train a global model without the need to share the raw information, however just a few parameters are sent over the network. Local devices or agents use their own data and some shared parameters over the network to update the learning algorithm. Learning and estimation over networks usually consist of computation, communication, decision making, etc. Computation is usually accompanied with energy consumption, while communication is usually limited with privacy of sharing information, sparsity of link resources (bandwidth, speed, costs, etc.); in addition, decision making is usually concerned with high accuracy and good performance. This talk deals with designing algorithms for estimation, processing, adaptation and learning over single agents and over networks. Our target is to have algorithms that have low computational complexity, not greedy for bandwidth and link resources and good learning accuracy. The presentation is divided into two main parts: In the first part, we will summarize briefly our work on designing low computational complexity signal processing algorithms for big data and low power DSP applications. In the second part, we will introduce the topic of information processing, estimation and learning over networks and how collaboration among agents in a network can improve learning performance. Our work will be concentrated in proposing an algorithm that is capable of solving an optimization and learning problem in an efficient and distributed manner from streaming data through localized interactions among their agents.

Seminar on 2020-02-04 15:00

Efficient Globally Optimal Resource Allocation in Wireless Interference Networks

(Dr. Bho Matthiesen, Universität Bremen, Germany)

Resource allocation in interference networks is essential to obtain optimal performance and resource utilization. The resulting optimization problems are usually nonconvex and their solution requires, in general, significant computational resources. For this reason, practical systems employ heuristic algorithms with no or only weak optimality guarantees. However, evaluating the performance of such practical algorithms requires the globally optimal solution. State-of-the-art global optimization approaches mostly employ the monotonic optimization framework which has some major drawbacks, especially when dealing with fractional objectives or complicated feasible sets. In this talk, the mixed monotonic programming (MMP) framework is introduced that provides a novel approach to exploit hidden monotonicity in global optimization problems. This results in a branch-and-bound based algorithm with several orders of magnitude faster convergence than state-of-the-art methods. Convergence issues due to complicated feasible sets are discussed and the successive incumbent transcending (SIT) scheme is introduced as a remedy. In an application example, this SIT approach is combined with the MMP framework to solve a hierarchical resource allocation problem.

Seminar on 2019-11-29 15:00

Neural Network Priors for Inverse Imaging Problems

(Professor Paul Hand, Northeastern University, Massachusetts, USA)

Recovering images from very few measurements is an important task in medical, biological, astronomical, and other image problems. Doing so requires assuming a model of what makes some images natural. Such a model is called an image prior. Classical priors such as sparsity have led to the speedup of Magnetic Resonance Imaging by a factor of 10x in certain cases. With the recent developments in machine learning, neural networks have been shown to provide efficient and effective priors for inverse problems arising in imaging. They can, for example, achieve comparable recovery quality as classical sparsity priors with 10x less data. In this talk, we will discuss three different philosophies of neural network image priors. First are generative models which describe desired images as belonging to an explicitly parameterized low-dimensional manifold. Second are invertible priors, which learn a probability distribution over all images. Third are unlearned priors, which include neural networks which require no training at all. For all these methods, we will see how to exploit the prior in a tractable optimization problem. We will present empirical performance and discuss relevant theoretical results and opportunities. 

Seminar on 2019-11-11 13:00

Computation, Estimation, Optimization and Learning Over Networks

(Mohammad Hussein Nassralla, American University of Beirut, Lebanon)

In this work, an algorithm for estimation, adaptation and learning over a single device and over a networks is presented. The proposed method has the properties of low computational complexity and good performance. Several aspects of complexity and convergence are discussed and a lightweight learning algorithm is presented that can be implemented locally at each agent of the network. The main contribution of this work is in proposing a low computational gradient descent algorithm.

Seminar on 2019-07-16 15:00

Coded Caching Over Wireless Channels 

(Professor Petros Elia, Communication Systems Department, Eurecom, Sophia-Antipolis, France)

This presentation will discuss coded caching as a means of achieving unparalleled performance in a variety of multi-antenna wireless settings. The presentation will advocate for a joint and careful treatment of caching and PHY can far outperform approaches that simply combine existing PHY and existing caching schemes. The presentation can be of interest to any coded caching specialist who wishes to see how "things change" when considering wireless channels, as well as of interest to PHY specialists who can see how existing PHY multi-antenna schemes can be changed in order to allow a splash of caching to provide substantial gains. In many cases, the results will be supported by novel information-theoretic converses that prove DoF optimality.

Seminar on 2019-07-10 11:00

Cyber Attacks on Internet of Things Sensor Systems for Inference

(Professor Rick S. Blum, Electrical and Computer Engineering Department, Lehigh University, Pennsylvania, USA)

The Internet of Things (IoT) improves pervasive sensing and control capabilities via the aid of modern digitial communication, signal processing and massive deployment of sensors. The employment of low-cost and spatially distributed IoT sensor nodes with limited hardware and battery power, along with the low required latency to avoid unstable control loops, presents severe security challenges. Attackers can modify the data entering or communicated from the IoT sensors which can have serious impact on any algorithm using this data for inference. In this talk we describe how to provide tight bounds (with sufficient data) on the performance of the best algorithms trying to estimate a parameter from the attacked data and communications under any assumed statistical model describing how the sensor data depends on the parameter before attack. The results hold regardless of the estimation algorithm adopted which could employ deep learning, machine learning, statistical signal processing or any other approach. Example algorithms that achieve performance close to these bounds are illustrated. Attacks that make the attacked data useless for reducing these bounds are also described. These attacks provide a guaranteed attack performance in terms of the bounds regardless of the algorithms the estimation system employs. References are supplied which provide various extensions to all the specific results presented and a brief discussion of applications to IEEE 1588 for clock synchronization is provided.

Seminar on 2019-06-24 16:00

Unsupervised Learning of Image Features Based on Relational Clustering

(Professor Michael Botsch, Technische Hochschule Ingolstadt)

A method for unsupervised learning of image features based on relational clustering methods is presented. There, image features are small characteristic pieces, called filters, of an image data set. The method can be summarized as follows. In a first step, all possible filters are extracted out of the image data set. Then some pre-processing and pre-clustering is applied to the filter data set. A reduced version of the filter data set is then clustered with relational clustering methods in order to determine the final filters. By convolving the final filters with the origin image data set, a feature extraction is performed. Moreover, an optional pooling step is part of the feature extraction principal. The method of learning and extracting features can be applied in a stacked way, such that the output of one layer is the input to another layer. Stacking enables deep architectures which can be used to extract high-level image features. The architectures feature extraction capabilities are evaluated using a random forest classifier. Therefore, clustering methods related to the categories hierarchical, partitioning, fuzzy and spectral are explained in detail. Two representative clustering methods, viz. partitioning around medoids and kernel spectral clustering are used for the proposed method.

Seminar on 2019-06-17 16:00

A Very Short Introduction to Quantum Correlations

(Dr. Janis Nötzel, TUM ECE LTI)

Correlated sequences of bits held by distant parties are a basic ingredient to every communication task. We encounter them e.g. as secret keys, or as en- and decoding procedures. They coordinate the actions of distant parties and thus provide a basis for communication. In a thought experiment, it will be outlined how quantum technology could, hypothetically, empower us to solve coordination problems that cannot be solved without it.

Seminar on 2019-06-06 11:30

Song Learning of the Zebra Finch, a Bird Model of Vocal Learning 

(Professor Manfred Gahr, Department Behavioral Neurobiology, Max Planck Institute for Ornithology, Seewiesen, Germany)

There are periods in children’s’ lives in which they become biologically mature enough to gain certain skills that they could not have easily picked up prior to that maturation. For example, babies and toddlers are more flexible with regard to learning to understand and use language than older children. Similar, in songbird species such as the zebra finch song learning occurs in a critical developmental period. Zebra finches memorize a song during a critical period between 25 – 65 days (PHD) after hatching (Roper and Zann, 2006). These songs are strings of smaller motor units (so called syllables) that are produced in a stereotyped way in adulthood. The juveniles progressively form their song through a sensorimotor process of matching their own vocal output with the stored memory of the tutor song (Konishi, 1965). Since the brain circuit that controls vocal development is well known, zebra finches are a model for the study of neuronal mechanism of vocal learning. In this seminar, mechanisms of vocal learning are discussed.

Seminar on 2019-05-17 14:00

Joint State Sending and Communications: Theory and Vehicular Applications

(Professor Mari Kobayashi, Telecommunications Department at CentraleSupélec, Gif-sur-Yvette, France, on a sabbatical leave at TUM)

A communication setup is considered where transmitters wish to simultaneously sense network states and convey messages to intended receivers. The scenario is motivated by joint radar and vehicular communications where the radar and data applications share the same bandwidth. First, a theoretical framework is presented to characterize the fundamental limits of such a setup for memoryless channels with i.i.d. state sequences. Then, the author's recent work on joint radar and communication using OFDM and Orthogonal Time Frequency Space (OTFS) is presented. Although restricted to a simplified scenario with a single target, numerical examples are demonstrated that two modulations provide as accurate radar estimation as Frequency Modulated Continuous Waveform (FMCW), a typical automotive radar waveform, while providing a non-negligible communication rate for free.

Seminar on 2019-05-14 17:00

Deep Decoder: Concise Image Representations from Untrained Non-convolutional Networks

(Professor Reinhard Heckel, TUM ECE ML)

Deep neural networks, in particular convolutional neural networks, have become highly effective tools for compressing images and solving inverse problems including denoising, inpainting, and reconstruction from few and noisy measurements. This success can be attributed in part to their ability to represent and generate natural images well. Contrary to classical tools such as wavelets, image-generating deep neural networks have a large number of parameters typically a multiple of their output dimension and need to be trained on large datasets. In this paper, we propose an untrained simple image model, called the deep decoder, which is a deep neural network that can generate natural images from very few weight parameters. The deep decoder has a simple architecture with no convolutions and fewer weight parameters than the output dimensionality. This underparameterization enables the deep decoder to compress images into a concise set of network weights, which we show is on par with wavelet-based thresholding. Further, underparameterization provides a barrier to overfitting, allowing the deep decoder to have state-of-the-art performance for denoising. The deep decoder is simple in the sense that each layer has an identical structure that consists of only one upsampling unit, pixel-wise linear combination of channels, ReLU activation, and channelwise normalization. This simplicity makes the network amenable to theoretical analysis, and it sheds light on the aspects of neural networks that enable them to form effective signal representations.

Seminar on 2019-03-21 16:00

Learning to Rest: A Q-Learning Approach to Flying Base Station Trajectory Design with Landing Spots

(Harald Bayerlein, Eurecom, Sophia Antipolis, France)

We consider the problem of trajectory optimization for an autonomous UAV-mounted base station that provides communication services to ground users with the aid of landing spots (LSs). Recently, the concept of LSs was introduced to alleviate the problem of short mission durations arising from the limited on-board battery budget of the UAV, which severely limits network performance. In this work, using Q-learning, a model-free reinforcement learning (RL) technique, we train a neural network (NN) to make movement decisions for the UAV that maximize the data collected from the ground users while minimizing power consumption by exploiting the landing spots. We show that the system intelligently integrates landing spots into the trajectory to extend flying time without any explicit information about the environment.

Seminar on 2019-02-08 15:00

Subarray Hybrid Architecture for mmWave Communications 

(Marcin Iwanow, Huawei Technologies, Munich, TUM ECE MSV)

A particular hardware architecture which is a possible candidate for the implementation of mmWave transceivers is proposed. The subarray hybrid beamforming architecture (SPI-HBF), which is the structure in question, exhibits pronounced hardware simplifications with respect to more complicated and widely considered fully interconnected hybrid beamforming architectures (FI-HBF) as proposed in literature. A framework is presented that helps to determine the degradation of performance induced by such a simplification. To this end, firstly a tight upper bound for the achievable rate of an SPI-HBF system is derived, then a novel precoding scheme which performs close to the derived bound is proposed. Lastly, a comparison of FI-HBF with SPI-HBF based on fair power modeling is presented.

Seminar on 2019-02-05 10:30

The Information Bottleneck Method for Discrete Information Processing

(Professor Gerhard Bauch, Technische University Hamburg)

The information bottleneck method proposed by Tishby et. al. is a classification tool which has been widely applied in the field of machine learning, particularly in the context of text classification. In this work, the concept is applied to message passing with coarse quantization in discrete signal processing and communications problems. State-of-the-art discrete signal processing is typically based on an optimum or close-to-optimum algorithm which is designed for real or complex-valued samples, at least for high resolution such as double precision. The optimum algorithm is then made implementable by approximations and reduced resolution, i.e., a coarser quantization with a limited number of bits per sample. Optimality of an algorithm is usually lost in such a quantized implementation, which often results in a significant performance degradation compared to a double precision implementation. In contrast, we pursue a fundamentally different approach. The key idea is to design each signal processing stage such that an appropriately defined relevant mutual information is preserved as much as possible. All operations degenerate to simple look-up tables and the passed messages are represented by a small number of bits. Performance very close to the optimum double precision implementation can be obtained with a small number of bits per sample and low computational complexity in many applications. The seminar lecture will focus on decoding of an LDPC code as a concise example which allows to explain the concept. This example is furthermore motivated by the fact that forward error control decoding is the by far computationally most expensive part of the baseband processing in a wireless communications system. Then the concept is extended to adjacent signal processing stages: to channel estimation and detection.

Seminar on 2018-12-06 16:00

From Iterative Algorithms to Deep Learning

(Professor Ami Wiesel, Hebrew University of Jerusalem, Israel)

Deep neural network and machine learning have recently revolutionized all fields of engineering. The most significant progress is probably in image processing and speech, but these trends are penetrating other applications as wireless communication, radar, and tomography. The switch from classical signal processing to modern deep learning involves fundamental changes. The applications mentioned above all have a well understood physical model that should be exploited in the design. This leads to a switch from model based design to data driven design, and, more importantly the shift to architecture based design. Another significant change is that standard machine learning methods are not suitable for parametric signal processing, and must be re-learned from scratch each time the parameters change. These challenges led researchers to consider hybrid approaches that exploit the benefits of both worlds: the network’s architectures are based on unfolding iterative signal processing algorithms, while deep learning allows more degrees of freedom and more expression power. Iterations are transformed into layers, and algorithms are succeeded by networks (see also the related concept of Recurrent Neural Networks). The resulting architectures allow a single training for multiple parametric models and achieve state of the art accuracy vs complexity tradeoffs. We will demonstrate these ideas in the context of MIMO decoding and robust parameter estimation.

Seminar on 2018-10-15 14:00

Time Modulated Arrays: an unconventional way to perform beamforming

(Professor Luis Castedo, University of A Coruña, Spain)

Time Modulated Arrays (TMA) is an unconventional way to perform beamfoming using switches instead or variable phase shifters and amplitude gains. TMAs modify the radiation pattern by periodically switching on and off the different excitations of an antenna array. In this talk, the main features of TMAs describing its main pros and cons will be discussed. In particular, the talk will explain how TMAs creates multiple beams at harmonic frequencies of the switching period and how they can be successfully used to exploit the wireless channel spatial diversity. The use of TMAs at transmission will be also considered and several methods to perform precoding with TMAs will be presented.

Seminar on 2018-10-11 13:00

Harmonic Radar: Detection and localization of electronic devices

(Florian Bischeltsrieder, DLR, TUM ECE MSV)

This talk gives a brief overview about the research activities at DLRs project “Harmony”. The goal of this project is the development of a harmonic radar in combination with classic radar imaging within a close-range measurement scenario. A harmonic radar excites and detects higher order frequency components (harmonics) of the transmit signal's frequency to detect targets that contain predominantly electronic components (e.g. diodes, transistors, etc.). The frequency separation between transmission and reception causes the suppression of non-electronic targets (e.g. background objects), so that only targets having a harmonic response function contribute to the received signal. This circumstance allows describing the measurement scene as a collection of point-like scatterers. Within this talk, special attention is given to the connection between classic radar imaging and the possibilities that arise from the harmonic radar measurement.

Seminar on 2018-07-23 12:00

Massive MIMO with Spatial Sigma-Delta ADCs

(Professor Lee Swindlehurst, University of California Irvine, TUM IAS, TUM ECE MSV)

There has recently been significant research exploring the limits of massive MIMO performance when implemented with one-bit ADCs on the uplink. Here we explore the possibility of using a different one-bit sampling architecture that is based on a spatial-analog of the well-known Sigma-Delta ADC for temporally oversampled systems. This so-called "Spatial Sigma Delta" implementation shapes the spatial spectrum of the quantization noise, pushing it towards higher spatial frequencies, and reducing its impact at low spatial frequencies (e.g., near broadside for a uniform linear array). This can provide some benefit when an oversampled array is used at the basestation, or when the signals from the users of interest are known to be confined to near broadside directions. We present a Bussgang analysis of both first- and second-order spatial Sigma-Delta architectures, and give some examples of its performance.

Seminar on 2018-07-12 14:00

Synthetic Molecular Communication: Fundamentals, Results, and Challenges

(Professor Robert Schober, Friedrich-Alexander-University Erlangen-Nürnberg)

In this talk, first a broad introduction to the field of synthetic molecular communication is given. Components of synthetic molecular communication networks, possible applications, and the evolution of the field will be reviewed. Thereafter, various synthetic molecular communication strategies such as gap junctions, molecular motors, and diffusion based molecular communication will be discussed. The focus is particularly on diffusion based synthetic molecular communication, identify the relevant basic laws of physics and discuss their implications for communication system. Subsequently, communication engineering design problems will be discussed, including detection, channel estimation, parameter estimation, and synchronization, and potential solutions will be offered. Furthermore, experimental results for an optical-to-chemical signal converter will be presented. Finally, some research challenges and open problems are discussed.

Seminar on 2018-06-18 16:00

Secrecy Capacity, Dirty Paper Coding, and Partial Decode-and-Forward -- Three System Models, Three Coding Schemes, One Mathematical Expression

(Dr. Christoph Hellings, TUM ECE MSV)

Achievable rates in information theoretic models of communication systems are related to mutual information expressions that can be expressed as differences of (differential) entropies. While the subtrahend only depends on the noise properties in case of a point-to-point transmission, it can depend on the properties of the input signals in systems with interference. This makes theoretical considerations more difficult (since maximum-entropy arguments can no longer be applied directly), and it usually lets the transceiver optimization become nonconvex (due to nonconcave rate expressions). However, in some cases, such entropy differences have special structures that make them tractable for numerical optimization and theoretical considerations. In this talk, we consider three examples where this is the case, namely the secrecy capacity in the MIMO wiretap channel, the dirty paper coding sum rate in the two-user MIMO broadcast channel, and the partial decode-and-forward rate in the MIMO relay channel. Interestingly, the rate expressions in these three very different scenarios can be reformulated in a way that the same mathematical structure is revealed in all of them. It thus becomes possible to transfer theoretical results and numerical algorithms from one scenario to the others, which is demonstrated in the talk by means of several examples.

Seminar on 2018-06-05 17:00

Holography of Wi-Fi radiation

(Professor Friedemann Reinhard, TUM WSI)

From a physicist's perspective, wireless radiation is just light, to be precise: coherent electromagnetic radiation. It is virtually the same as the beam of a laser, except that its wavelength is much longer. We have developed a way to visualize this radiation, providing a view of the world as it would look like for eyes sensitive to wireless radiation. Our scheme is based on holography, a technique to record three-dimensional pictures by a phase-coherent recording of radiation in a two-dimensional plane. We have adapted it to work with wireless radiation, and recorded holograms of building interiors illuminated by the omnipresent stray field of wireless devices. In the resulting three-dimensional images we can see both emitters and absorbing objects. Our scheme does not require any knowledge of the data transmitted and works with arbitrary signals, including encrypted communication. This result has several implications: it could provide a way to track wireless emitters in buildings, it could provide a new way for through-wall imaging of building infrastructure like water and power lines. It finally touches upon security concerns: wireless signals transmit a 3D view of their surrounding, even when encrypted.

Seminar on 2018-05-17 13:00

Achieving Centralized Interference Reduction with Decentralized Precoding in Broadcast Channels

(Dr. Paul de Kerret, Eurecom, Sophia Antipolis, France)

With the current trends towards dense networks of flexible, moving, and low cost transmitters, decentralized/partially centralized transmission from multiple transmitters becomes a cost-efficient alternative to the so-called Cloud-RAN architecture. Yet, interference management in such a decentralized setting is faced with new challenges as the channel state information is then likely to be not only imperfect estimated, but also imperfectly shared between the cooperating transmitters. In this talk, we show how it is possible (and necessary) to use novel robust joint transmission schemes. Surprisingly, it is possible in some configurations to achieve the same degrees-of-freedom (which measures the interference reduction) as in the Cloud-Ran setting where the transmitters perfectly cooperate. Considering the particular case of two single-antenna transmitters, it is possible to achieve the same degrees-of-freedom independent of whether a channel estimate is present at one TX, at the other TX, or at both, and regardless of the channel pathloss topology. This is in contrast with the usual design guidelines in practical networks which require the channel information to be available at both transmitters.

Seminar on 2018-04-27 11:30 (room N1135)

On Deep Learning-Based Communication Over the Air

(Professor Stephan ten Brink, University of Stuttgart)

An over-the-air communications system is demonstrated which is solely based on deep neural networks and has so far been only validated by computer simulations for block-based transmissions. Transmitter and receiver can be jointly trained end-to-end for an arbitrary differentiable end-to-end performance metric, e.g., block error rate (BLER). It is demonstrated that it is possible to build and train such a system using off-the-shelf software-defined radios (SDRs) and open-source deep learning software libraries. A comparison of the BLER performance of the “learned” system against that of a practical baseline shows competitive performance. We identify several practical challenges of training such a system over-the-air, in particular the missing channel gradient, and propose a learning procedure that circumvents this issue.

Seminar on 2018-03-28 14:00

Title: Performance Analysis of Non-Orthogonal Multiple Access Systems with Residual Hardware Impairments

(Afsaneh Gharouni, M.Sc., Friedrich-Alexander-University Erlangen-Nürnberg)

Non-Orthogonal Multiple Access (NOMA) is a promising technology which is considered as a candidate for 5G, however, the performance of NOMA systems has not yet been studied considering the effects of hardware non-idealities. This talk discusses the impact of transceiver hardware impairments (HWIs) on the performance of a single-cell two-user downlink NOMA system which performs successive interference cancellation (SIC) at the receivers. It can be seen from the analysis that HWI-aware design of a NOMA system, in which HWI parameters are taken into the account in determining the optimum SIC order, shows robustness in presence of hardware impairments.

Seminar on 2018-02-07 16:00

Operation and Control of Device-to-Device Communication in Cellular Networks

(Markus Klügel, TUM ECE LKN)

Direct communication among user equipments, called Device-to-Device (D2D) communication, is considered to be integral part of future cellular networks. The introduction of such direct links into the otherwise strictly hierarchical cellular system creates a paradigm shift towards more ubiquitous networking. However, it also creates challenges, because D2D links need to be coordinated remotely as not to degrade system performance. In this presentation, core results on the impact and management strategies of D2D communication are presented. The impact of spatial device distributions and higher layer applications on the expected number of D2D links is assessed. The gain of D2D communication, with and beyond utility maximization, is investigated and a method for interference management is proposed.

Seminar on 2018-01-16 14:00

On the Statistical Properties of Constant Envelope Quantizers

(Hela Jedda, TUM ECE MSV)

In massive multiple-input multiple-output (MIMO) wireless systems, where the number of power amplifiers (PAs) scales up with the number of transmit antennas, constant envelope (CE) signaling is beneficial in terms of power efficiency. With CE input signals, the PAs can be operated at their saturation regions and hence use the available power efficiently. Moreover, the input signals of the PAs are the radio-frequency signals that are at an early stage converted from the digital to the analog domain by the digital-to-analog converters (DACs). The power consumption of the DACs grows exponentially with their resolution To further enhance the hardware power efficiency it is desirable to decrease the DACs' resolution. Thus, we combine the idea of CE signaling at the PA input and the coarse quantization of the DACs to end up with coarsely quantized CE signals. In this talk, we study the impact of CE quantizers on the statistical properties of complex-valued circularly-symmetric Gaussian distributed input signals.

Seminar on 2017-12-19 16:00

Untersuchung zu einem abbildenden harmonischen Radar mit digitaler Strahlformung

(Florian Bischeltsrieder, DLR, TUM MSV)

Ein harmonisches Radar verursacht und detektiert die an Messzielen mit nichtlinearen Strom-Spannungs-Kennlinien entstehenden, harmonischen Frequenzanteile des am Objekt einstrahlenden Sendesignals. Im Vortrag werden theoretische und experimentelle Untersuchungen zur räumlich aufgelösten Abbildung einer harmonischen Messszene im Nahbereich vorgestellt.

Seminar on 2017-12-19 14:00

Data-Selective Adaptive Filtering

(Professor Paulo S. R. Diniz, Universidade Federal do Rio de Janeiro, Brazil)

The current trend of acquiring data pervasively calls for some data-selection strategy, particularly in the case a subset of the data does not bring enough innovation. As a byproduct, in addition to reducing power consumption and some computation, the discarding of data results in more accurate parameter estimation. In many practical situations, it is possible to verify if the acquired set of data qualifies to improve the related statistical inference or if it consists of an outlier or a non-innovative entry. In this presentation, we discuss some extensions of the existing adaptive iterative algorithms enabling data selection which also addresses the censorship of outliers measured through unexpected high estimation errors. Simulation results show the effectiveness of the proposed algorithms for selecting the innovative data without sacrificing the estimation accuracy, while reducing the computational cost.

Seminar on 2017-12-12 13:30

Time Series Prediction using Long Short-Term Memory Networks

(Cansu Sancaktar, TUM ECE)

Recurrent Neural Nets (RNN) are powerful tools to model sequential data. However, basic RNN cells struggle to capture long-term dependencies due to the vanishing gradients problem. Long Short-Term Memory Networks (LSTM) are a special kind of RNN that address this problem by introducing gates to the net architecture. After an introduction to machine learning concepts, the presentation will cover the network topology and forward pass of LSTMs as well as examples of time series prediction using LSTMs.

Seminar on 2017-12-05 10:30

Monotonic and Sequential Fractional Programming for Performance Optimization in 5G and Beyond

(Professor Eduard A. Jorswieck, Technical University of Dresden)

Resource allocation in interference networks is a timeless and important challenge. In future generations of mobile networks, heterogeneous and conflicting performance criteria are introduced leading to multi-objective programming problems. When efficiency is optimized often fractional programming - a well established technique - can be applied. However, in interference networks, the fractional objective functions are in general not suitable for fractional programming. In my talk, the combination of fractional programming with monotonic optimization (to achieve global optimality with high complexity) and with sequential convex programming (to achieve local optimality with lower complexity) are proposed. Their applications are illustrated with distributed and centralized power control in 5G and beyond.

Seminar on 2017-12-01 15:00

Digital Predistortion in Practice

(Professor Thomas Magesacher, Electrical and Information Technology, Lund University, Sweden)

Digital predistortion is an important element of radio-frequency transmitters driving power amplifiers at high efficiency. This talk gives a tutorial introduction to the underlying nonlinear signal processing techniques and associated system architecture considerations. Typical questions arising during the design of a practical system, such as choice of learning architecture, selection of behavioral models, or bandwidth requirements, are discussed. An outlook of challenges arising in emerging communication systems using multiple and wider bands and many antennas is presented.

Seminar on 2017-11-23 16:00

On Deformation Stability of Deep Convolutional Neural Networks (DCNNs)

(Michael Koller, TUM ECE MSV and Johannes Großmann, TUM ECE LTI)

We are interested in extracting features from a finite energy input function/signal via a deep convolutional neural network (DCNN). The features should be stable with respect to deformations of the input. Intuitively, this means that features extracted from two similar input functions should also be similar. We investigate what conditions a DCNN needs to fulfill such that deformation stability can be guaranteed. Our work is based on a theory of DCNNs that has been in development since 2011.

Seminar on 2017-11-14 13:30

Channel Estimation and Hybrid Precoding/Combining for Frequency Selective Multiuser mmWave Systems

(Dr. José P. González Coma, University of A Coruña, Spain)

The problem of designing the precoders and combiners in a multiuser hybrid frequency selective millimeter wave (mmWave) system is a challenging problem. In this talk we explore two methods to design the hybrid analog-digital filters when only an estimate of the channel is available at both the base station (BS) and the users. The first one is a projected gradient method that factorizes the all-digital solution, while the second approach is based on an alternating minimization (AM) iterative strategy. The channel is estimated using a Simultaneous Weighted-Orthogonal Matching Pursuit (SW-OMP) algorithm that takes into account the hardware limitations of the hybrid architecture and leverages the sparse structure of mmWave channels. The proposed methods are evaluated in terms of achievable sum rate, outperforming the state of the art solutions.

Seminar on 2017-07-11 16:00

Downlink training sequence design for multi-user massive MIMO systems

(Dr. Samer Bazzi, Huawei Technologies, Munich)

In this work, we design multi-user downlink training sequences that leverage user spatial or spatio-temporal correlations. To keep signaling overheads acceptable, we consider the case where the base station (BS) selects sequences from a predefined codebook and conveys their indices to the users. To that purpose, we first solve an optimization problem where sequences do not necessarily belong to the codebook, then map the obtained sequence solution to sequences from the codebook using a novel “range matching pursuit” algorithm. Numerical results show the effectiveness of the proposed algorithm and additionally show that in the presence of strong correlations, the channels can be accurately estimated with training durations that can be much smaller than the number of BS antennas.

Seminar on 2017-07-04 10:00

Transient Stability Analysis by Reachable Set Computation

(Professor Matthias Althoff, TUM IN)

I present a formal technique for verifying the stability of transient responses of power systems. The procedure uses reachability analysis to compute the complete set of possible transient responses starting from a set of initial states, subject to a dynamics specified by differential-algebraic equations. The method is constructive and fully automatic, two properties that are often hard to achieve with direct Lyapunov methods when the differential-algebraic equations are not simplified. Reachability analysis is computationally expensive, but this work presents new techniques that make it possible to verify the stability of a transient response of IEEE benchmark power system networks.

Seminar on 2017-05-31 11:00-13:00

On Mixed-ADC Massive MIMO

(Professor Lee Swindlehurst, University of California Irvine, TUM IAS, TUM ECE MSV)

Implementing MIMO systems with cheap and low-power one-bit analog-to-digital converters (ADCs) is a promising approach that can alleviate the power consumption burden at the base station. However, this approach suffers from substantial channel estimation error caused by the coarse quantization which results in spectral efficiency (SE) loss. In this talk, we investigate the effectiveness of "mixed-ADC" architectures to overcome this issue. In particular, we study to what extent one can combat the SE loss by exploiting just N ≪ M pairs of high-resolution ADCs, where M denotes the number of BS antennas. Analytical results are presented that show the impact of the mixed-ADCs on channel estimation accuracy and in turn on SE for maximum ratio combining and zero-forcing detection in the low and high signal-to-noise ratio regime.

Subspace Matryoshkas for mmWave Communications

(Professor Wolfgang Utschick, TUM ECE MSV)

In mmWave communication systems so called hybrid architectures have been proposed, which decompose the conventional fully digital precoder or combiner into a digital and analog part. In this presentation, it will be demonstrated that the previously proposed Linear Successive Allocation (LISA) method perfectly fulfills the requirements of a hybrid architecture. A new low-complexity version of LISA based on the geometric nature of mmWave channel is presented. Numerical results demonstrate the superiority of the proposed solution over state-of-the-art methods.

Seminar on 2017-04-25 16:30-18:00

Discrete Optimization for Power Systems

(Niklas Winter, TUM ECE MSV)

Power systems are large networks with many system constraints and partly discrete control variables. For an economic and reliable operation of power systems, highly complex non-convex optimization problems need to be solved. Furthermore, security constraints increase the dimensionality of the problem by a huge factor. This presentation shall give an overview of the arising problems, as well as, introduce solution methods, e.g. the Generalized Benders Decomposition.