Model-Based Deep Learning

Deep neural networks provide unprecedented performance gains in many real-world problems in signal and image processing. Despite these gains, the future development and practical deployment of deep networks are hindered by their black-box nature, i.e., a lack of interpretability and the need for very large training sets.

In our lab, we are working on model-based deep learning, where the design of learning-based algorithms is based on prior domain knowledge. This approach allows to integrate models and other knowledge about the problem into both the architecture and training process of deep networks. This leads to efficient, high-performance and yet interpretable neural networks which can be employed in a variety of tasks in signal and image processing. Model-based networks require far fewer parameters than their black-box counterparts, generalize better, and can be trained from much less data. In some cases, our networks are trained on a single image, or only on the input itself so that effectively they are unsupervised.

We are developing theory and applications of model-based learning and applying this paradigm to solve a wide variety of tasks, from super-resolution in optical microscopy and ultrasound to speed-of-sound reconstruction using ultrasonography, to the development of modern wireless systems, to image deblurring, signal separation, and more.

vv
Model-Based Network via Unrolling

 

References