Stochastic three-term conjugate gradient method with variance technique for non-convex learning

In the training process of machine learning, the minimization of the empirical risk loss function is often used to measure the difference between the model’s predicted value and the real value. Stochastic gradient descent is very popular for this type of optimization problem, but converges slowly in theoretical analysis. To solve this problem, there are already many algorithms with variance reduction techniques, such as SVRG, SAG, SAGA, etc. Some scholars apply the conjugate gradient method in traditional optimization to these algorithms, such as CGVR, SCGA, SCGN, etc., which can basically achieve linear convergence speed, but these conclusions often need to be established under some relatively strong assumptions. In traditional optimization, the conjugate gradient method often requires the use of line search techniques to achieve good experimental results. In a sense, line search embodies some properties of the conjugate methods. Taking inspiration from this, we apply the modified three-term conjugate gradient method and line search technique to machine learning. In our theoretical analysis, we obtain the same convergence rate as SCGA under weaker conditional assumptions. We also test the convergence of our algorithm using two non-convex machine learning models.