Bouarfa Mahi
I'm a researcher in machine learning and neural network optimization currently working on structured knowledge accumulation and entropy-based learning frameworks. My work focuses on developing biologically-inspired alternatives to traditional gradient-based optimization methods, particularly through the lens of information theory and entropy reduction.
My research centers on the Structured Knowledge Accumulation (SKA) framework, which reinterprets neural learning as a continuous-time process governed by entropy minimization principles. I've developed novel concepts including the Tensor Net function and characteristic time properties of neural learning, establishing connections between computational learning and physical systems that evolve according to natural laws. This work demonstrates how layer-wise entropy reduction can lead to autonomous, hierarchical knowledge accumulation without relying on traditional backpropagation.
I'm particularly interested in forward-only neural learning, mechanistic interpretability, and efficient transformers. My approach bridges information theory with artificial intelligence, offering promising applications in resource-constrained environments and parallel computing systems. The framework I've developed provides a scalable, biologically plausible alternative to gradient-based learning while maintaining mathematical rigor through variational principles and the Euler-Lagrange equation.
My research centers on the Structured Knowledge Accumulation (SKA) framework, which reinterprets neural learning as a continuous-time process governed by entropy minimization principles. I've developed novel concepts including the Tensor Net function and characteristic time properties of neural learning, establishing connections between computational learning and physical systems that evolve according to natural laws. This work demonstrates how layer-wise entropy reduction can lead to autonomous, hierarchical knowledge accumulation without relying on traditional backpropagation.
I'm particularly interested in forward-only neural learning, mechanistic interpretability, and efficient transformers. My approach bridges information theory with artificial intelligence, offering promising applications in resource-constrained environments and parallel computing systems. The framework I've developed provides a scalable, biologically plausible alternative to gradient-based learning while maintaining mathematical rigor through variational principles and the Euler-Lagrange equation.