A new research paper titled "Algebraic Machine Learning: Learning as Computing an Algebraic Decomposition of a Task" by Fernando Martin-Maroto, Nabil Abderrahaman, David Mendez, and Gonzalo G. de Polavieja introduces a novel approach to machine learning based on Abstract Algebra. This framework offers an alternative to traditional methods rooted in statistics and optimization, providing a new perspective on the foundations of learning.
A New Foundation for Machine Learning
This research proposes a new framework where the task's goal and the data are encoded as axioms of an algebra. A model is then constructed where only these axioms and their logical consequences are valid. Unlike traditional models, this approach allows for a precise analysis of learning by breaking down tasks into algebraic atoms through subdirect decomposition.
Generalization Through Algebraic Decomposition
While the initial model does not generalize, selecting specific subsets of algebraic atoms leads to a model that does generalize. This approach was validated on well-known datasets, including:
- MNIST
- FashionMNIST
- CIFAR-10
- Medical Images
The results show performance comparable to optimized multilayer perceptrons, highlighting the effectiveness of this new learning principle.
Applications Beyond Data-Driven Tasks
This algebraic method is not limited to traditional data-driven tasks. It extends to formal problems, such as identifying Hamiltonian cycles from their specifications without using search algorithms, opening new possibilities for problem-solving in computational theory.
Key Features of Algebraic Machine Learning
- Direct Learning from Training Data: No validation dataset required.
- Scalability Through Model Additivity: Efficiently scales while maintaining performance.
- Asymptotic Convergence: Converges to the underlying rule in the data.
Read the Full Paper
We invite researchers and practitioners in machine learning and artificial intelligence to read the full paper and explore this new approach.
Read the paper here
Stay connected with eProsima for more updates on research and developments in machine learning.
MORE INFORMATION ABOUT ALMA:
For any questions please contact
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 952091.