隱變量深度生成模型 (Deep latent generative model) 大多需要選擇簡單、 可解析 (tractable) 機率分佈,來做為隱變量先驗假設,但從最近研究發現, 選擇不同的先驗分佈可能會影響生成模型能力;本研究提出一種從資料中 學習隱變量先驗分佈的方法,無須人為給定選擇特定的先驗分佈,並用於 先驗正規化自編碼器 (Prior Regularied Autoencoder),引入一個編碼生成網路 (Code generator) 來學習先驗分佈,可以更好的捕捉資料特性,最後提出一個 訓練框架,聯合訓練生成模型與先驗分佈;從實驗上來看,本研究提出方法 能有效提高生成樣本品質,並且於表徵學習任務、文字至影像的轉換任務上 都有良好的表現。
Most deep latent factor models choose simple priors for simplicity, tractability or not knowing what prior to use. Recent studies show that the choice of the prior may have a profound effect on the expressiveness of the model, especially when its generative network has limited capacity. In this paper, we propose to learn a proper prior from data for Prior Regularied Autoencoder. We introduce the notion of code generators to transform manually selected simple priors into ones that can better characterize the data distribution. Experimental results show that the pro- posed model can generate better image quality and learn better disentangled rep- resentations than AAEs in both supervised and unsupervised settings. Lastly, we present its ability to do cross-domain translation in a text-to-image synthesis task.