In this paper we propose a novel modification of Contrastive Language-Image Pre-Training (CLIP) guidance for the task of unsupervised backlit image enhancement. Our work builds on the state-of-the-art CLIP-LIT approach, which learns a prompt pair by constraining the text-image similarity between a prompt (negative/positive sample) and a corresponding image (backlit image/well-lit image) in the CLIP embedding space. Learned prompts then guide an image enhancement network. Based on the CLIP-LIT framework, we propose two novel methods for CLIP guidance. First, we show that instead of tuning prompts in the space of text embeddings, it is possible to directly tune their embeddings in the latent space without any loss in quality. This accelerates training and potentially enables the use of additional encoders that do not have a text encoder. Second, we propose a novel approach that does not require any prompt tuning. Instead, based on CLIP embeddings of backlit and well-lit images from training data, we compute the residual vector in the embedding space as a simple difference between the mean embeddings of the well-lit and backlit images. This vector then guides the enhancement network during training, pushing a backlit image towards the space of well-lit images. This approach further dramatically reduces training time, stabilizes training and produces high quality enhanced images without artifacts, both in supervised and unsupervised training regimes. Additionally, we show that residual vectors can be interpreted, revealing biases in training data, and thereby enabling potential bias correction.

The reconstruction of images from their corresponding noisy Radon transform is a typical example of an ill-posed linear inverse problem as arising in the application of computerized tomography (CT). As the (naive) solution does not depend on the measured data continuously, regularization is needed to re-establish a continuous dependence. In this work, we investigate simple, but yet still provably convergent approaches to learning linear regularization methods from data. More specifically, we analyze two approaches: One generic linear regularization that learns how to manipulate the singular values of the linear operator in an extension of our previous work, and one tailored approach in the Fourier domain that is specific to CT-reconstruction. We prove that such approaches become convergent regularization methods as well as the fact that the reconstructions they provide are typically much smoother than the training data they were trained on. Finally, we compare the spectral as well as the Fourier-based approaches for CT-reconstruction numerically, discuss their advantages and disadvantages and investigate the effect of discretization errors at different resolutions.

We propose a novel framework for the regularised inversion of deep neural networks. The framework is based on the authors' recent work on training feed-forward neural networks without the differentiation of activation functions. The framework lifts the parameter space into a higher dimensional space by introducing auxiliary variables, and penalises these variables with tailored Bregman distances. We propose a family of variational regularisations based on these Bregman distances, present theoretical results and support their practical application with numerical examples. In particular, we present the first convergence result (to the best of our knowledge) for the regularised inversion of a single-layer perceptron that only assumes that the solution of the inverse problem is in the range of the regularisation operator, and that shows that the regularised inverse provably converges to the true inverse if measurement errors converge to zero.

We introduce a novel mathematical formulation for the training of feed-forward neural networks with (potentially non-smooth) proximal maps as activation functions. This formulation is based on Bregman distances and a key advantage is that its partial derivatives with respect to the network's parameters do not require the computation of derivatives of the network's activation functions. Instead of estimating the parameters with a combination of first-order optimisation method and back-propagation (as is the state-of-the-art), we propose the use of non-smooth first-order optimisation methods that exploit the specific structure of the novel formulation. We present several numerical results that demonstrate that these training approaches can be equally well or even better suited for the training of neural network-based classifiers and (denoising) autoencoders with sparse coding compared to more conventional training frameworks.

This research project investigates the application of deep learning to timbre transfer, where the timbre of a source audio can be converted to the timbre of a target audio with minimal loss in quality. The adopted approach combines Variational Autoencoders with Generative Adversarial Networks to construct meaningful representations of the source audio and produce realistic generations of the target audio and is applied to the Flickr 8k Audio dataset for transferring the vocal timbre between speakers and the URMP dataset for transferring the musical timbre between instruments. Furthermore, variations of the adopted approach are trained, and generalised performance is compared using the metrics SSIM (Structural Similarity Index) and FAD (Frech\'et Audio Distance). It was found that a many-to-many approach supersedes a one-to-one approach in terms of reconstructive capabilities, and that the adoption of a basic over a bottleneck residual block design is more suitable for enriching content information about a latent space. It was also found that the decision on whether cyclic loss takes on a variational autoencoder or vanilla autoencoder approach does not have a significant impact on reconstructive and adversarial translation aspects of the model.