A new easy-to-use and fast-to-setup tool developed by an international team of researchers, including from the CMCC, allows scientists to leverage artificial neural networks, deep learning and machine learning methods. Although developed to be applied to the problem of wildfire management, Kit4DL can improve research in many areas, including climate science, by enabling easy deployment of machine learning architecture.

As data storage and data collection capabilities continue to grow exponentially, scientists and professionals from a variety of different fields are increasingly turning to artificial neural networks, deep learning and machine learning for answering complex questions and to gain deeper insight. However, not all researchers have the skill or time to focus on the data preparation, extracting features and setting up multiple experiments for the training and validation required to run these kinds of models.

“The Kit4DL toolkit enables researchers to avoid having to write boilerplate code and models every time they want to use machine learning,” says CMCC researcher Shahbaz Alvi who was part of the international team that developed Kit4DL, a new toolkit that helps speed up and simplify the experimentation process of machine- and deep-learning.

“In a variety of research fields, including climate science, researchers want to leverage machine learning; however, not all researchers are experts in the underlying subtleties of how machine learning actually works,” says Alvi. “An analogy I like to use is that of a car: we can all drive cars, but we don’t necessarily have the knowledge or means to build them.”

Although there are various tools that make using machine learning easier, Kit4DL takes this one step further so that users only need to implement a few core methods outlined in configuration files. This simplifies and speeds up development time compared to traditional approaches and also facilitates code reusability by allowing researchers to leverage the same codebase across multiple experiments, reducing redundancy and streamlining the experimentation process.

Fire Danger Index images of the Gargano region in Italy were generated using machine learning with the Kit4DL toolkit. The images show how the inference changes for consecutive days. Kit4DL was used for fast prototyping Deep Learning Models for wildfire danger prediction. Source: Integrated Technological and Information Platform for Wildfire Management (SILVANUS).

The Kit4DL toolkit was developed in the context of the CMCC’s Integrated Technological and Information Platform for Wildfire Management (SILVANUS) project where researchers felt the need for a tool that would allow them to cross-check results effectively. Funding for the project came from the Polish National Centre for Research and Development under the LIDER XI program and from the European Union’s Horizon 2020 Research and Innovation program, SILVANUS Project.

“We wanted to simplify the process of using machine learning models and get consistent results when we moved from one experiment in machine learning to another,” explains Alvi. “Kit4DL allows you to create a modular structure automatically which then allows researchers and professionals in different fields to use machine learning effectively and easily in their work.”

For more information:

Jakub Walczak, Marco Mancini, Shahbaz Alvi, Kit4DL: Towards fast prototyping and experimentation in machine learning and deep learning, SoftwareX, Volume 26, 2024, 101707, ISSN 2352-7110, https://doi.org/10.1016/j.softx.2024.101707