Jan Lennart Haug

Title: Jan Lennart Haug and Logarithmic Mathematics for AI Matrix Multiplications.

Introduction

Jan Lennart Haug is an inventor based in Munich, Germany, whose patent work addresses efficient mathematical operations for artificial intelligence. His disclosed invention focuses on performing matrix multiplication through coordinated conversions between logarithmic and linear domains.

Latest Patents

Haug is associated with 1 patent: “Method and Apparatus for Logarithmic Mathematics in AI Matrix Multiplications.” The patent describes a system that performs multiplication in the logarithmic domain and accumulation in the linear domain using co-designed conversion circuits.

The linear-to-logarithmic circuit divides a mantissa into four non-uniform windows and applies affine mappings implemented with shift and add operations. The logarithmic-to-linear circuit uses four equal-width windows and corresponding affine mappings, also implemented with shift and add operations. The mappings are designed as analytic inverses, with the first circuit’s window boundaries derived by inverse mapping the second circuit’s boundaries. This co-design aims to provide round-trip exactness apart from quantization effects and reduce systematic bias during accumulation.

Career Highlights

The available information identifies Haug’s work with Tensordyne, Inc. His patent describes a hardware-oriented approach to mathematical processing in AI, using shift-and-add operations in its conversion circuits.

Collaborations

Haug’s listed coworkers include Jian Hui Huang and Gary S. Goldman. The provided information does not specify their roles in the patent or their particular contributions.

Conclusion

Jan Lennart Haug’s patent presents a co-designed method for logarithmic-domain multiplication and linear-domain accumulation, with the goal of supporting accurate AI matrix computations while limiting conversion-related bias.