Semester Projects (Semesterarbeiten)

If you are interested in one of the following topics, please contact the person listed with the respective topic.

If you don't find a compelling topic in our list, we are always keen on hearing about your ideas in the area of our research interests (see the Research section of our website). You are most welcome to discuss your interests with us.

Also, we have a list of finished student projects on our website.

List of Projects



Telemetry for a sounding rocket (SA)

mlerjen_aris.jpg
The ``Akademische Raumfahrt Initiative Schweiz (ARIS)'' aims at engaging students from Swiss educational institutions in hands-on space engineering challenges, specifically in the design, construction, and testing of a sounding rocket to compete in the 2018 Spaceport America Cup in New Mexico, USA.

The Spaceport America Cup is the world’s largest university rocket engineering competition. Student teams are challenged to design and build a rocket capable of launching a 4 kg payload to target altitudes of up to 30,000 ft and to recover the rocket completely post-flight.

In this multi-disciplinary project, you will contribute to the development of the telemetry system of the ARIS rocket. The goal is to implement, test, and verify a proof-of-concept solution for intra-rocket and rocket-to-ground communication. In a first step different options for antennas, communication modules, and protocols shall be analyzed and compared with telemetry systems of commercial rocket manufacturers. In particular, performance, implementation complexity, and cost, all in the context of the structure of the rocket and the specific environment (e.g. high velocity and acceleration), shall be taken into account. In a second step the favored solution shall be implemented, tested, and verified on an ARIS model rocket.

Type of project: 30% concept development, 70% avionics development, testing, and measurements
Prerequisites: Interest in embedded and wireless systems
Supervisor: Michael Lerjen
Professor: Helmut Bölcskei

References:
[1] Homepage of the ARIS project [Link to Document]



Implementation of an RF channel emulator (SA)

mlerjen_ChannelEmulator.jpg
In the development process of radio transceivers, it is important to test and optimize the performance of the implemented hardware, coding schemes, and signal processing algorithms under different channel conditions. This can be done, in principle, in over-the-air tests. However, such an approach exhibits several disadvantages, namely interference from other systems, various logistical problems, and legal restrictions in certain frequency bands. Furthermore, the results will depend on environmental conditions and are therefore often not reproducible.

A channel emulator, i.e., a piece of hardware that emulates a channel and applies it to a communication signal, is a tool built exactly to solve this problem. The key parameters of the channel (e.g., Doppler spread, delay spread, noise, ...) can be controlled through the configuration of the emulator [1]. Channel emulators often also have multiple inputs/outputs for the emulation of MIMO channels. MIMO channel emulation devices are available on the market [2], but are expensive and sometimes not flexible enough.

The aim of this project is to build a MIMO channel emulator that works in the digital baseband domain and can be controlled from an external PC. An FPGA platform including an RF transceiver and analog/digital (ADCs) and digital/analog converters (DACs) was already selected in a previous project and can be used as a starting point.

Type of project: 70% VHDL programming and simulation, 20% measurements, 10% theory
Prerequisites: Interest in RF hardware and wireless systems, Matlab, VHDL
Supervisor: Michael Lerjen
Professor: Helmut Bölcskei

References:
[1] S. Yang and Z. Can, “Design and implementation of multiple antenna channel emulator for LTE system,” Proc. 9th International Conference on Communications and Networking in China (CHINACOM), pp. 208–213, 2014 [Link to Document]

[2] Keysight Technologies, Inc., “Wireless device test sets & wireless solutions” [Link to Document]



Grounded language learning of visual-lexical color descriptions (SA)

pdmytro_example_color.png
Grounded language learning (GLL) is a technique for language acquisition that uses a multimodal set of inputs rather than just sets of words or symbols, e.g. it uses a combination of words and related sounds or visuals [1]. Due to the similarity of GLL with the way humans are exposed to language, studying GLL can potentially yield insights on how language is comprehended by humans.

In this project, you will explore GLL by training a generative neural network on combinations of linguistic and visual information. Specifically, you will build a generative model based on the recently introduced Deep Recurrent Attentive Writer (DRAW) neural network architecture [2]. You will then train this network to predict lexical color descriptions from visual color representations and vice versa. Both descriminative [3] and generative [4, 6] approaches were used in the literature to deal with the problem of color description. The goal of this project is to investigate whether DRAW networks can improve on the autoencoder approach in [4]. As a training/testing dataset you will use a subset from an online survey [5, 6].

Type of project: 20% theory, 80% programming
Prerequisites: Strong programming skills (Python), background in machine learning and deep learning
Supervisor: Dmytro Perekrestenko
Professor: Helmut Bölcskei

References:
[1] A. Lazaridou et al., “Multimodal word meaning induction from minimal exposure to natural text,” Cognitive Science, 2016. [Link to Document]

[2] K. Gregor, I. Danihelka, A. Graves, D. Rezende, and D. Wierstra, “{DRAW}: A recurrent neural network for image generation,” Proceedings of the 32nd International Conference on Machine Learning, 2015. [Link to Document]

[3] W. Monroe, N. D. Goodman, and C. Potts, “Learning to generate compositional color descriptions,” arXiv, 2016. [Link to Document]

[4] D. Bhargava, G. Vega, and B. Sheffer, “Grounded learning of color semantics with autoencoders,” [Online source], 2017. [Link to Document]

[5] R. Munroe, “Color survey results,” [Blog], 2010. [Link to Document]

[6] B. McMahan and M. Stone, “A bayesian model of grounded color semantics,” Transactions of the Association for Computational Linguistics, 2015. [Link to Document]



Distribution-preserving lossy compression (SA/DA)

michaelt_dplc.png
Recent advances in extreme image compression [1] allow artifact-free image reconstruction even at very low bitrates. Motivated by these results, [2] formalizes the concept of distribution-preserving lossy compression (DPLC), which optimizes the compression rate-distortion tradeoff under the constraint of the reconstructed (decompressed) samples following the empirical distribution of the training data. Specifically, a DPLC system (almost) perfectly reconstructs the training data when enough bits are allocated to the compressed representation. When zero bits are assigned to the compressed representation it learns a (deep) generative model of the data, and for intermediate bitrates DPLC smoothly interpolates between matching the distribution of the training data and perfectly reconstructing the training samples (cf. the figure on the left; the numbers at the top correspond to different rates (in bits per pixel) and each row corresponds to a different decoder realization). The DPLC framework introduced in [2] was so far applied to images only. This project shall explore new applications.

Type of project: 20%-40% theory, 60%-80% programming, depending on the student's preference
Prerequisites: Programming, linear algebra, experience with deep learning software is a plus
Supervisor: Michael Tschannen
Professor: Helmut Bölcskei

References:
[1] E. Agustsson, M. Tschannen, F. Mentzer, R. Timofte, and L. Van Gool, "Generative adversarial networks for extreme learned image compression," arXiv:1804.02958, 2018. [Link to Document]

[2] M. Tschannen, E. Agustsson, and M. Lučić, "Deep generative models for distribution-preserving lossy compression," arXiv:1805.11057, 2018. [Link to Document]



Control policy learning for a flight simulator (SA)

mlerjen_simulator.jpg
Reinforcement learning (RL) [1] refers to the problem of finding suitable actions to take in a given situation in order to maximize a reward. RL algorithms that recently led to breakthrough results in practical machine learning tasks (such as learning to play Atari games [2] or the strategy board game Go [3]) typically employ a deep convolutional neural network [4] for feature extraction.

This project aims at developing an RL-based approach for learning control policies for a flight simulator. The scope of the project may encompass value function design, training data acquisition, as well as a first implementation in Python.

Type of project: 90% theory/literature research, 10% programming
Prerequisites: Python, machine learning basics is a plus
Supervisor: Michael Tschannen, Michael Lerjen
Professor: Helmut Bölcskei

References:
[1] R. S. Sutton and B. A. Barto, “Reinforcement learning: An introduction,” MIT press, 1998.

[2] V. Mnih et al., “Human-level control through deep reinforcement learning,” Nature, vol. 518 (7540), pp. 529–533, 2015. [Link to Document]

[3] D. Silver et al., “Mastering the game of Go with deep neural networks and tree search,” Nature, vol. 529 (7587), pp. 484–489, 2016. [Link to Document]

[4] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521 (7553), pp. 436–444, 2015. [Link to Document]



Analyzing cost functions for generative adversarial networks (SA/DA)

michaelt_gans_pic.jpg
Generative adversarial networks (GANs) [1] have led to remarkable results in machine learning, in particular in image generation tasks [2]. The GAN-learning problem is a two-player game between the so-called generator, who learns how to generate samples resembling the training data, and the discriminator, who learns how to discriminate between real and fake data points. Both players aim at minimizing their own cost function until a Nash-equilibrium is established.

The goal of this project is to analyze–-mathematically and possibly experimentally–-different cost functions in the context of GAN-learning for image generation tasks.

Type of project: 80%-90% theory, 10-20% simulation
Prerequisites: Analysis, linear algebra, probability theory, Python
Supervisor: Michael Tschannen
Professor: Helmut Bölcskei

References:
[1] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” Proc. of Neural Information Processing Systems (NIPS), pp. 2672–2680, 2014. [Link to Document]

[2] A. Radford, L. Metz, and S. Chintala “Unsupervised representation learning with deep convolutional generative adversarial networks,” arXiv:1511.06434, 2015. [Link to Document]



Reconstructing a neural network from its output (SA)

vlacicv_uniqueness.png
Deep neural networks have become a staple method applied to a wide range of machine learning tasks such as optical character recognition [1], image classification [2], and speech recognition [3]. A neural network effectively implements a function through concatenations of affine maps and nonlinearities. The number of concatenations together with the dimensions of the affine maps and the coefficients describing these maps is referred to as the architecture of the network. The nonlinearity typically acts componentwise and is fixed, whereas the coefficients of the affine maps need to be learned based on training data consisting of example input-output pairs. A natural question that arises is that of uniqueness: Can the architecture of a neural network be uniquely identified from the output function to be realized?

This uniqueness problem has not enjoyed much popularity, but it did receive serious treatment in the seminal work by Fefferman [4]. Specifically, it was shown in [4] that neural networks based on the tanh nonlinearity are, indeed, uniquely determined by their output function, under certain genericity assumptions, and up to obvious isomorphisms of networks. In addition to the obvious curtailment of practical relevance by restricting the nonlinearity to be precisely the function tanh, we have observed another undesirable issue. Specifically, the genericity assumptions, while indeed mild for a truly generic network, are restrictive enough to imply full connectedness of successive layers. This is an assumption which is not satisfied by a majority of neural networks used in practice.

The goal of this project is to generalize the results in [4] by considering nonlinearities other than tanh, as well as loosening the genericity assumptions, yielding results that apply more closely to neural networks used in practice.

Type of project: 80% theory/literature research, 20% own derivations and proofs
Prerequisites: complex analysis, basic familiarity with neural networks is beneficial
Supervisor: Verner Vlačić
Professor: Helmut Bölcskei

References:
[1] Y. LeCun, L. D. Jackel, L. Bottou, et al., “Comparison of learning algorithms for handwritten digit recognition,” International Conference on Artificial Neural Networks, pp. 53–60, 1995. [Link to Document]

[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Advances in Neural Information Processing Systems 25, Curran Associates, Inc., 2012, pp. 1097–1105. [Link to Document]

[3] G. Hinton, L. Deng, D. Yu, et al., “Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups,” IEEE Signal Process. Mag., vol. 29, no. 6, pp. 82–97, 2012. doi: 10.1109/MSP.2012.2205597. [Link to Document]

[4] C. Fefferman, “Reconstructing a neural net from its output,” Revista Matematica Iberoamericana, vol. 10, no. 3, pp. 507 555, 1994. [Link to Document]