Partial Label Supervision for Agnostic Generative Noisy Label Learning
Stats
"Deep neural network (DNN) has achieved remarkable success in computer vision [14,24], natural language processing (NLP) [10,63] and medical image analysis [29,45]."
"Noisy label learning has been tackled with both discriminative and generative approaches."
"Extensive experiments on computer vision and natural language processing (NLP) benchmarks demonstrate that our generative modelling achieves state-of-the-art results while significantly reducing the computation cost."
Quotes
"Despite the simplicity and efficiency of discriminative methods, generative models offer a more principled way of disentangling clean and noisy labels."
"Our proposal faces three challenges: the removal of Z makes p(X|Y) under-constrained because Z cannot 'anchor' the image generation process; as shown in Fig. 1-(a), Z and Y are not independent given the observation of X, so the removal of Z implies that X needs to be constrained by an informative Y; and how to design a single-stage training for a model that is agnostic to different causal processes (i.e., p(Y|X) or p(X|Y))."