Seminar of the School of Electrical and Computer Engineering – 21/05/2025
Departmental seminar
Detecting multiple targets in heavy-tailed, correlated clutter is a challenge for both classical detectors and standard neural nets. This research introduces a novel dual-head diffusion model that learns to denoise range-Doppler maps and, in parallel, pinpoint target locations-yielding more reliable multi-target detection in complex radar scenarios
by Ari Granevich
This work presents enhanced frameworks for peak-to-average power ratio (PAPR) reduction and waveform design, tailored for Single-Input-Single-Output (SISO) systems with single-carrier (SC) waveforms. The system leverage convolutional autoencoder (CAE) architectures, where an end-to-end learning-based autoencoder (AE) models the communication system through an encoder and decoder, with the latent representation passing through a physical communication channel
We propose a joint learning scheme utilizing projected gradient descent to optimize spectral mask behavior and transmitted signal detection under the influence of a non-linear high-power amplifier (HPA). The novel waveform design technique employs a PAPR reduction, ensuring throughput losslessness as no side information is required at the decoder
Performance is evaluated in terms of bit error rate (BER), PAPR, and spectral response, and is benchmarked against classical PAPR reduction. The proposed system demonstrates competitive performance across all optimization criteria. Furthermore, gradual loss-based multi-objective optimization empirically shows that a single trained model effectively addresses PAPR reduction, spectrum design, and signal detection tasks across a wide range of SNR levels
by Rom Hirsch
