The speaker discusses the need for improved benchmarks in machine learning as a way to promote further development of deep neural networks beyond the ImageNet dataset, which has already been extensively explored.