數學軟體應⽤ 457 Autoencoder encoder decoder z latent vector x m >> k k維 x m維 於是, z 可以取代 x (或者說 z 是 x 的 一個 presentation) Autoencoder 是輸入什 麼, 就輸出什麼的函數。 感覺很奇怪, 原來是中間我 們會⽤⼀層比較⼩的 維 神經元。 這⼀層的輸出就是我們準 備當特徵向 的。 k
數學軟體應⽤ 467 LeCun 認為 GAN 是深度學習最有潛⼒的 model There are many interesting recent development in deep learning… The most important one, in my opinion, is adversarial training (also called GAN for Generative Adversarial Networks). “ —Yan LeCun (楊⽴昆), 2016 ”
數學軟體應⽤ 509 更好的是⽤ Triplet Loss F. Schroff, D. Kalenichenko, J. Philbin (Google). FaceNet: A Unified Embedding for Face Recognition and Clustering. arXiv preprint arXiv:1503.03832. CNN ̂ y 1 ̂ y 2 ̂ y n CNN ̂ y 1 ̂ y 2 ̂ y n Positive Sample Negative Sample
數學軟體應⽤ 515 Self-Supervised Learning We believe that self-supervised learning is one of the most promising ways to build such background knowledge and approximate a form of common sense in AI systems. “ —Yan LeCun (楊⽴昆)/Ishan Misra, 2021 ” Self-supervised Learning: The Dark Matter of Intelligence https://ai.facebook.com/blog/self-supervised- learning-the-dark-matter-of-intelligence/