Designing appropriate loss functions is critical to the success of supervised learning models. However, most conventional losses are fixed and manually designed, making them suboptimal for diverse and dynamic learning scenarios. In this work, we propose an Adaptive Meta-Loss Network (Adaptive-MLN) that learns to generate task-agnostic loss functions tailored to evolving classification problems. Unlike traditional methods that rely on static objectives, Adaptive-MLN treats the loss function itself as a trainable component, parameterized by a shallow neural network. To enable flexible, gradient-free optimization, we introduce a hybrid evolutionary approach that combines Genetic Algorithms (GA) for global exploration and Evolution Strategies (ES) for local refinement. This co-evolutionary process dynamically adjusts the loss landscape, improving model generalization without relying on analytic gradients or handcrafted heuristics. Experimental evaluations on synthetic tasks and the CIFAR-10 and MNIST datasets demonstrate that our approach consistently outperforms standard losses such as Cross-Entropy and Mean Squared Error in terms of accuracy, convergence, and adaptability.
Yunita et al. (Thu,) studied this question.