Scientific seminars given by experts working in industrial sectors working on topics such as biomedical imaging, cultural heritage, HPC, medical imaging.
Visual Autoregressive (VAR) models are emerging as a powerful paradigm for generative computer vision, offering scalability and quality comparable to Diffusion models. However, deploying these large-scale architectures on industrial platforms is hindered by their prohibitive memory footprint and significant computational constraints. This seminar explores the theoretical and practical challenges of compressing autoregressive models for constrained hardware, like consumer GPUs or even embedded platforms (e.g., NVIDIA Jetson). We will analyze the statistical properties of these networks, focusing on activation outliers and channel variance that render standard weight-activation quantization techniques ineffective. We will then combine advanced algebraic approaches—such as Singular Value Decomposition (SVD)—to structurally decouple high-rank outliers from weight matrices, enabling efficient low-bit representation without retraining. The talk aims to bridge the gap between mathematical optimization and real-world deployment, showing how generative AI can be enabled on the edge.
TBA...
Deep learning-based methods have revolutionized the field of imaging inverse problems, yielding state-of-the-art performance across various imaging domains. The best performing networks incorporate the imaging operator within the network architecture, typically in the form of deep unrolling. However, in large-scale problems, such as 3D imaging, most existing methods fail to incorporate the operator in the architecture due to the prohibitive amount of memory required by global forward operators, which hinder typical patching strategies. In this seminar, I will present a domain partitioning strategy and normal operator approximations that enable the training of end-to-end reconstruction models incorporating forward operators of arbitrarily large problems into their architecture. The proposed method achieves state-of-the-art performance on 3D X-ray cone-beam tomography and 3D multi-coil accelerated MRI, while requiring only a single GPU for both training and inference.
TBA...
TBA...