Untrained Neural Network

Untrained neural networks (UNNs) leverage the inherent structure of neural network architectures to solve various problems without requiring traditional supervised training. Current research focuses on applying UNNs to inverse problems, such as image reconstruction and signal processing, often employing architectures like U-Nets and recurrent networks, and exploring techniques like pruning and latent-space disentanglement to enhance performance and efficiency. This approach offers significant advantages in scenarios with limited or unavailable labeled data, leading to improved solutions in diverse fields ranging from medical imaging to wireless communications.