A Random Matrix Analysis and Optimization Framework to Large Dimensional Transfer Learning

This article proposes a first performance analysis and optimization of a simple transfer learning method, extending the standard least squares support vector machine. By means of a random matrix analysis, we prove that, for simultaneously large and numerous data, the correct classification rate of the learning task is asymptotically predictable and the hyperparameters in the problem are easily tuned so to maximize the output performance. Simulations confirm our findings. This preliminary work opens the path to a systematic exploration of transfer learning methods by means of large dimensional statistics.