Deep Neural Networks (DNNs) are dominating various tasks such as classification, object detection, speech recognition and natural language processing. Despite their undisputable success, understanding the decision-making process of DNNs and identifying key factors involved in the process remains an on-going challenge. Gaining further insights in the decision-making process of DNNs is especially crucial for safety-critical applications such as autonomous driving, where explainable artificial intelligence (XAI) aims to produce interpretations for machine learning based decision. Explainability methods in the field of computer vision (CV) can be divided into two categories, namely proxy (LIME, ...) and direct strategies (Guided Grad-CAM). As the name implies, proxy strategies rely on a proxy model that approximates the DNN of interest, where the decision-making of the DNN at hand is interpreted by querying the proxy model. These proxy approaches are considered as "black-box" approaches of explainability methods. Contrary, direct strategies represent "white-box" approaches of explainability methods requiring the networks parameters of the target DNN such as gradients and activations of respective layers. Utilizing these internal parameters, direct strategies are able to identify the key factors within the input that are crucial for the decision-making process of the target DNN.

In the scope of this seminar topic, state-of-the-art explainability methods are identified and surveyed. Additional to highlighting the principles of state-of-the-art techniques, advantages and disadvantages of the presented technique should be portrayed.

Lukas Frickenstein

Email: Lukas.Frickenstein@bmw.de

Graph neural networks (GNNs) have emerged as a technique for understanding relationships and dependencies in complex data structures. The target of this seminar is to investigate whether such networks are suitable in identifiying feature correlation within vision based neural networks, potentially aiding a pruning algorithm in making decisions on which filters have dependencies on one another.

Nael Fasfous
Department of Electrical and Computer Engineering
Chair of Integrated Systems

Phone: +49.89.289.23858
Building: N1 (Theresienstr. 90)
Room: N2116
Email: nael.fasfous@tum.de

Learning classifier systems are a paradigm of rule-based machine learning methods that combine the power of a discovery component like genetic algorithms with a learning component like Q-learning. XCS is a type of LCS based on accuracy of the rules instead of strength, in order to achieve a rule-set with accurage and general rules to cover the input space. 

Ruleset reduction algorithms are used to reduce the size of the final ruleset. This work surveys the different types of ruleset reduction algorithms.

anmol.surhonne@tum.de

n recent years, numerous approximation methods have been published to bridge the gap between increasing performance demands and staggering technology improvements. Many of these methods target specific parts of an application, e.g. arithmetic execution units. Complex systems containing many different operations therefore offer lots of opportunities to apply approximations in various places. This leads to large and diverse design spaces which need to be explored in a clever way to find globally optimal configurations in the resource-quality trade-off space. The goal of this Seminar Topic is to review and summarize the state of the art for design space exploration techniques dealing with Hardware/FPGA-based signal processing systems that employ approximate computing.

Simon Conrady

ARRI - Arnold & Richter Cine Technik GmbH & Co. Betriebs KG

SConrady@arri.de

The rise of Multicore processors in the recent decade induced also the necessity appropriate Memory Models. Especially with the ever-growing number of corses, it is not distinct by default whether a global memory hierarchy or a distributed memory structure is more beneficial. While former can potentially become a bottleneck if too many cores access it simultaneously, latter arises new challenges of data-to-task locality and a prescient data placement.

This seminarr topic is about a comparative analysis of both approaches. This includes amongst others their benefits and drawbacks, their suitable application fields and typical system architectures.

The task during the seminar consists of an exhausitve literature study on that topic in order to describe, analyze and compare the most important approaches. Individual adjustments of the the seminar topic according to the student's preferences is possible.

B.Sc. in Electrical Engineering, Computer Science or a similar topic

Oliver Lenke

o.lenke@tum.de

The Rooflie Model, presented by Williams et.al. in 2009, is an intuitive and visual way to illustrate the system's performance and to observe potential bottlenecks. Based on the proposal in 2009, various appraoches have been presented recently to apply the Roofline Model to a wide range of Computer architectures.

This seminar puts the Roofline Model in the context of Near-Memory Computing and Hardware Accelerators. It differentiates thereby between fixed-function accelerators and software-programmable cores and considers also the system architecture und the memory hierarchy.

The task during the seminar consists of an exhausitve literature study on that topic in order to describe, analyze and compare the most important approaches. Individual adjustments of the the seminar topic according to the student's preferences is possible.

B.Sc. in Electrical Engineering, Computer Science or a similar topic

Oliver Lenke

o.lenke@tum.de

The Rooflie Model, presented by Williams et.al. in 2009, is an intuitive and visual way to illustrate the system's performance and to observe potential bottlenecks. Based on the proposal in 2009, various appraoches have been presented recently to apply the Roofline Model to a wide range of Computer architectures.

This seminar puts the Roofline Model in the context of Manycore Architectures. It differentiates thereby between Homogeneous and Heterogeneous Architectures, as well as between traditional Multicores and Tile-based MPSoCs.

The task during the seminar consists of an exhausitve literature study on that topic in order to describe, analyze and compare the most important approaches. Individual adjustments of the the seminar topic according to the student's preferences is possible.

B.Sc. in Electrical Engineering, Computer Science or a similar topic

Oliver Lenke

o.lenke@tum.de

Convolutional neural networks (CNNs) present many computation and memory challenges for embedded devices. An effective way to reduce the complexity of a CNN is by binarizing the weights and activations of the network. This drastically decreases the size of the CNN and results in XNOR-popcount operations replacing the standard Multiply-Accumulate operations, typically performed billions of times over the course of a single inference. However, binary neural networks (BNNs) are difficult to train and can lead to a severe degradation in accuracy. To effectively train BNNs, special methods are typically applied to help gradients flow during backpropagation through the non-differentiable functions employed in the forward pass. Other researchers have completely re-invented the training approach to accomodate the complexity of BNN training.

Manoj Vemparala

Email: Manoj-Rohit.Vemparala@bmw.de 

Convolutional neural network (CNN) compression has become a standard approach of optimization before real-world deployment. Particularly in embedded scenarios, a high-accuracy CNN is usually heavily overparametrized for the target device. The most common techniques to compress a pretrained neural network are quantization and pruning. However, both techniques can also be applied at training-time, saving GPU hours and leading to one-shot training and optimization of CNNs.

Manoj Vemparala

Email: Manoj-Rohit.Vemparala@bmw.de 

Going beyond the efficient inference libraries provided by NVIDIA, researchers have started developing customized kernels, which better exploit their applied neural network compression techniques. In this seminar, these computation kernels will be studied and analyzed for the benefits they could offer to binary and sparse neural network execution.

Nael Fasfous
Department of Electrical and Computer Engineering
Chair of Integrated Systems

Phone: +49.89.289.23858
Building: N1 (Theresienstr. 90)
Room: N2116
Email: nael.fasfous@tum.de

Random number generation (RNG) is an essential part of many applications, be it for encryption, testing, or other purposes. However, randomness is difficult to accomplish in digital circuits which are supposedly deterministic.

The goal of this seminar topic is to understand and explain the challenges of RNG in hardware, and to compare different true random number generation (TRNG) and pseudo random number generation (PRNG) methods that are suitable for implementation on FPGA.

Managing power is fundamental for multicore processor performance. For this, accurate information is necessary about where and when power is consumed. A multitude of neural network modeling methodologies have been proposed in the last yeras. In this seminar, you will make a literature review into the newest power modeling techniques proposed in the last 3 years.

Mark Sagi
Department of Electrical and Computer Engineering
Chair of Integrated Systems

Email: mark.sagi@tum.de

Camera sensors are an important input to Advanced Driver Assistance Systems (ADAS) and Autonomous Drive (AD) of cars. A challenge for the camera sensors are the very high dynamic ranges of the input signal and the variation of the illumination of the environment. The candidate should work on understanding principles of high dynamic range (HDR) image capturing, different pixel technologies for HDR sensing, exposure control for HDR images, relations to LED flicker mitigation, algorithms to create HDR images from the captured input data and algorithms to compress the high dynamic range images to display the images to a human driver or vision processing system.

Dr. Ing Stephan Herrmann Architect for Vision SoC and IP
NXP Semiconductors Germany, Munich Phone: +49 (0) 170 678 60 94
stephan.herrmann@nxp.com

Advances in the deep learning architectures for computer vision applications have led to new neural architectures such as vision transformers. These differentiate themselves from typical convolutional neural network-based implementations by decoupling the process of feature aggregation and transformation. Excellent performance is achieved through self-attention and self-supervision.

In this seminar topic, these new ML for vision techniques will be investigated.

Nael Fasfous
Department of Electrical and Computer Engineering
Chair of Integrated Systems

Phone: +49.89.289.23858
Building: N1 (Theresienstr. 90)
Room: N2116
Email: nael.fasfous@tum.de

Multi-objective Optimization (MOO) is well-known for trade-off analysis between objectives in many real-world problems including embedded systems design, for example [1]. The MOO results in the formation of  Pareto Optimal points that allows the decision-maker to select the points based on his desired trade-off in an application. One typical example for MOO is a Genetic Algorithm based on NSGA selection [2]. However, NSGA algorithms often lead to the exploration and optimization of the entire design space in each objective dimension. This is not necessary for many applications and a significant computational effort is wasting for regions outside the threshold values in the decision maker's mind. 

This seminar aims to summarize different multi-objective optimization approaches which form Pareto Optimal solutions based on the preference given by the designer (e.g. [3]). Besides, comparisons between each method and the benefits and drawbacks of these methods in real-world applications also need to be investigated.

[1] Manu Manuel, Arne Kreddig, Simon Conrady, Nguyen Anh Vu Doan, Walter Stechele: Model-Based Design Space Exploration for Approximate Image Processing on FPGA. 2020 IEEE Nordic Circuits and Systems Conference (NorCAS), 2020.

[2] K. Deb, S. Agrawal, A. Pratap and T. Meyarivan, "A Fast Elitist Nondominated Sorting Genetic Algorithm for Multi-objective Optimization: NSGA-II", Parallel Problem Solving from Nature PPSN VI ser. Lecture Notes in Computer Science, pp. 849-858, 2000.

[3] Kalyanmoy Deb and J. Sundar. Reference point based multi-objective optimization using evolutionary algorithms. In Proceedings of the 8th Annual Conference on Genetic and Evolutionary Computation, GECCO ‘06, 635–642. New York, NY, USA, 2006.

manu.manuel@tum.de

Camera sensors are an important input to Advanced Driver Assistance Systems (ADAS) and Autonomous Drive (AD) of cars. A challenge for the camera sensors are the very high dynamic ranges of the input signal and the variation of the illumination of the environment. The candidate should work on understanding principles of high dynamic range (HDR) image capturing, different pixel technologies for HDR sensing, exposure control for HDR images, relations to LED flicker mitigation, algorithms to create HDR images from the captured input data and algorithms to compress the high dynamic range images to display the images to a human driver or vision processing system.

Dr. Ing Stephan Herrmann Architect for Vision SoC and IP
NXP Semiconductors Germany, Munich Phone: +49 (0) 170 678 60 94
stephan.herrmann@nxp.com

This seminar topic should result in a survey on state of the art scheduling methodes of tasks in NoCs and TSN connected ECU like in a car.

In contrast to the continued rise of various deep neural network ML variants, we favor the properties that come with explicit knowledge LCS (Learning Classifier System) -type ML techniques in our IPF project: interpretable and explainable rules which also can be pre-initialized with best-practice prior knowledge content.

This Seminar should provide an overview of existing, explainable ML techniques and of methodes to make black-box techniques (NNs) explainable.