2 PhD students “Beyond Kalman Neural Network-controlled Smart Digital Filters” (# of pos: 2)

We are looking to hire two PhD students under projects 5 and 6 of the 'Beyond Kalman Neural Network-controlled Smart Digital Filters' work package (WP2) in the RAISE (Robust AI for safe (radar) signal processing) project. WP2 will investigate and develop efficient low-complexity machine learning (artificial neural networks and deep learning) techniques for smart digital signal processing. In particular, Project 5 'Neural network-controlled smart digital filters for audio processing' will concentrate on portable sound-based applications (e.g. context-aware noise cancellation algorithms in headphones, hearing aids, etc.). On the other hand, Project 6 'Smart digital filters for radio broadcast interference rejection in electrical vehicles' will aim at developing low-complexity hybrid model-based and data-driven processing solutions for the automotive radio communication systems. Both projects will be carried out with a strong accent on implementing the state-of-the-art algorithms in hardware.

The RAISE is situated on the edge between academia (Eindhoven University of Technology) and industry (NXP Semiconductors), and in total will involve 3 post-docs, 6 PhD and 2 PDEng students.

TU/e is opening 2 PhD vacancies in the field of Robust AI for automotive signal processing ICs:

Background:

Future cars will incorporate ever more artificial intelligence (AI) to support driver safety, navigation, and comfort. Much of this AI will be embedded in integrated circuits (ICs) that serve, for example, to sense and perceive the car's environment and to support wireless communication and radio reception.

To help drive this development TU/e has teamed up with NXP, the globally leading Automotive IC design company, in a joint research program. In this program PhD and PDEng students will be co-supervised by top AI experts from TU/e and top automotive IC-design experts from NXP, and students will also be hosted by NXP on a part-time basis. As a result, the positions in the program offer a unique combination of high-quality scientific and industrial experience.

The following positions in this program are still available:

1. PhD position on smart digital filters for audio processing

Modern headphones, earbuds and hearing aids already have active noise cancelation. However, the user experience must be further improved. Whether a component of an audio signal contains relevant information or just irrelevant noise depends on the context. E.g., while most environmental noises must be suppressed, certain alarms or anomalous sounds might have to be passed on to the user. Or, a hearing aid should help its user to focus on a certain conversation in a setting with many simultaneous speakers (the cocktail party effect). This might be done by adaptive spectral/spatial digital filtering of multi-channel audio signals. Again, the required filter response depends on the context. In addition, user feedback might be used for adaptive selection of the operation mode (listening to music, following a conversation, etc.) and personalization.

This PhD position seeks to develop neural network-based (controllers for) multi-channel adaptive filters that accomplish these objectives at low latency and complexity levels consistent with ultra-low-power implementation as required e.g. for hearing aids. To train these networks an existing data base of labeled audio signals will be exploited and further extended. We envisage hybrid AI solutions in which neural networks efficiently predict auto-labeled target filter coefficients from extracted audio signal features (e.g. Mel-spectrograms). Validation of the developed approach will moreover require innovative neural model compression solutions that are tailored to the hardware specifications, facilitating hardware implementation and convincing real-time demos.

2. PhD position on smart digital filters for radio broadcast interference rejection in electrical vehicles

Interference is a common problem in wireless communications and is becoming more severe as the wireless spectrum becomes more crowded. Conventional interference rejection methods focus on radiated interferences, which originate from dedicated wireless transmitters. Recently, due to the fast development of electrical vehicles, a new vehicle-internal type of interference has emerged. For example, it is observed that DC/DC converters in electrical-cars introduce significant interference to radio receivers from the low-frequency (around 1 MHz) Amplitude Modulation (AM) signal-band all the way up to the Digital Audio Broadcasting (DAB) band (around 200 MHz). This new type of interference differs from the conventional interference in many respects. For example, it varies widely in time, frequency, and among different car models, and no statistical or deterministic interference models are available yet.

As such, this PhD position seeks to go all the way from characterizing and modeling these statistically complex interferences using generative AI models (e.g. normalizing flows, VAEs or score-based models) to developing and validating AI based techniques to suppress it. Also here algorithmic complexity is an essential constraint, and complexity levels should be consistent with low-power real-time implementation.