Convex and Non-Convex Approaches for Statistical Inference with Class-Conditional Noisy Labels