CoLoRA: Efficient Adaptation of Pretrained Convolutional Networks through Separable Kernel Updates

Fine-tuning pretrained neural networks is essential for adapting general-purpose models to specialized tasks, but updating all model parameters can be computationally expensive and prone to overfitting when limited training data are available. This talk presents *CoLoRA*, a parameter-efficient fine-tuning method designed specifically for convolutional neural networks. Inspired by Low-Rank Adaptation (LoRA), CoLoRA freezes the pretrained backbone and represents each convolutional-kernel update using lightweight pointwise and depthwise components. These updates can be merged into the original convolutional kernels after training, preserving the model size and computational complexity during inference.

We will introduce the mathematical formulation of CoLoRA, derive its parameter savings relative to full fine-tuning, and discuss how its placement across different convolutional blocks affects accuracy and training cost. Experimental results on OCTMNISTv2 and other medical-image benchmarks demonstrate the method’s accuracy, stability, and generalizability. Comparisons with transfer learning, adapters, BitFit, and convolutional LoRA variants illustrate the trade-offs among predictive performance, trainable parameters, training time, and deployment complexity. The talk will conclude with limitations and opportunities for extending CoLoRA to segmentation, volumetric imaging, and other convolutional architectures.

Bio.

Mariano Rivera is researcher at the Department of Computer Science at the Center for Research en Mathematics (CIMAT) since 1997. Member level III of the National System of Researchers (SNII), Mexico and of the Mexican Council of the Society of Mathematics and Member by Invitation of the Mexican Academy of Computing.  He has completed academic stays in the Department of Radiology at the University of Pennsylvania (Postdoctoral Researcher, 2001-2002), in the Department of Mathematics at Florida State University (Visiting Professor, Department of Mathematics, 2008-2009), and at the Center National Supercomputing Program of the Potosino Institute for Scientific and Technological Research (Academic Coordinator, 2017-2018).  His research lines focus on the topics: machine learning, image processing, computer vision, and numerical optimization. With applications in optical metrology, remote sensing data analysis and medical image analysis.

NOTA. Todos los seminarios se graban y están a tu disposición en el Canal de YouTube del Departamento, consúltalos: 

https://www.youtube.com/channel/UCwZp9bGRHC1FT5lof7kuLJg/playlists

Lugar: Auditorio Depto. Ciencias de la Computación y en línea Join Zoom Meeting https://us02web.zoom.us/j/81488375706?pwd=CqGfohJxsaK3NzbqCTcG6wmuhWVgXk.1

Fecha: 07-08-2026

Hora: 12:00 pm