GAN总结【三】
文獻整理
? | 題目 | 主要內(nèi)容 |
? ? ? ? ? ? ? ? ? ? ? ? ? ? GAN綜述 | 【1】 「無中生有」計算機視覺探奇 (下)?? | 1. 1)超分辨率重建;2)圖像著色;3)看圖說話;4)人像復原;5)圖像自動生成 2.?生成對抗網(wǎng)絡博弈論中的零和博弈 3. 將GAN用深度卷積神經(jīng)網(wǎng)絡進行實現(xiàn)(稱作,DCGAN, Deep Convolutional GAN) 5.?基于生成式卷積網(wǎng)絡的最新工作STGConvNet 自動合成動態(tài)紋理,聲音 |
【2】LeCun:深度學習突破,對抗式網(wǎng)絡最值得期待 | 1.?????學習系統(tǒng)和因子圖(基于能量的模型)相結(jié)合:“結(jié)構(gòu)化預測”(structured prediction) 2.?????深度學習局限:依賴于監(jiān)督學習,人類標注。需要找到方法,訓練大型神經(jīng)網(wǎng)絡從沒有經(jīng)過標注的“原始”數(shù)據(jù)中,找出現(xiàn)實世界的規(guī)律。對抗訓練 3.?????生成對抗式網(wǎng)絡,以及現(xiàn)在被提出的一些變體,是深度學習領域過去10年我認為最有意思的idea。 4.?????它讓我們可以訓練一個鑒別器,作為一種非監(jiān)督的“密度估計”(density estimator);這個鑒別器必須要發(fā)展出一個好的數(shù)據(jù)內(nèi)部表征,鑒別器還可以被當成分類器中的一個特征提取器。 5.?????通過對抗性訓練建立的生成/預測模型; ????無監(jiān)督式的學習預測模型(例如視頻預測):使超大規(guī)模神經(jīng)網(wǎng)絡不需要通過明確的人工注釋數(shù)據(jù),而只通過觀看視頻,閱讀教材等,就能“學習世界如何運行”。 | |
【3】【Ian Goodfellow】生成對抗式網(wǎng)絡創(chuàng)始人Quora答疑 | 1.??????對抗網(wǎng)絡和對抗訓練的聯(lián)系和區(qū)別 ü??Christian Szegedy?發(fā)明了對抗訓練(adversarial training )這種算法包括訓練神經(jīng)網(wǎng)絡正確分類正常實例與「對抗實例(adversarial examples )」;《神經(jīng)網(wǎng)絡的有趣屬性》( Intriguing properties of neural networks)描述了對抗訓練。 ü? 我發(fā)明了生成式對抗網(wǎng)絡。生成式對抗網(wǎng)絡是成對的網(wǎng)絡,另一個是鑒別器網(wǎng)絡,這篇論文沒有使用術語「對抗訓練」。 ü??我找到了一種更快生成對抗實例的方法,這就在對抗訓練的每一步上讓制造一小批新的對抗實例變得實際可行,而不是在每個階段只能制造幾個實例?!秾箤嵗慕忉尯挽柟獭?(Explaining and Harnessing Adversarial Examples),我們首次給它命名為「對抗訓練」 ü??對抗訓練的最初指代:以對抗實例來訓練的術語; 后來其他人開始使用對抗訓練指代生成式對抗網(wǎng)絡 我們可以將生成式對抗網(wǎng)絡視作執(zhí)行對抗訓練,對抗訓練中的生成器網(wǎng)絡為鑒別器網(wǎng)絡制造對抗實例。 2.??圖像生成框架——GAN/VAE/PixelCNN/NICE GAN優(yōu)勢: ü??比其它模型產(chǎn)生了更好的樣本。 ü??能訓練任何一種生成器網(wǎng)絡;生成對抗式網(wǎng)絡能學習可以僅在與數(shù)據(jù)接近的細流形(thin manifold)上生成點。 ü? 不需要設計遵循任何種類的因式分解的模型, 任何生成器網(wǎng)絡和任何鑒別器都會有用。 3.??????與其他生成式模型比較 ü? 與 PixelRNN相比,生成一個樣本的運行時間更小。 ü? 與VAE相比,它沒有變化的下限。如果鑒別器網(wǎng)絡能完美適合,那么這個生成器網(wǎng)絡會完美地恢復訓練分布。換句話說,各種對抗式生成網(wǎng)絡會漸進一致(asymptotically consistent),而 VAE 有一定偏置。 ü? 與深度玻爾茲曼機相比,既沒有一個變化的下限,也沒有棘手的分區(qū)函數(shù)。它的樣本可以一次性生成,而不是通過反復應用馬爾可夫鏈運算器(Markov chain operator)。 ü? 與GSN 相比,它的樣本可以一次生成,而不是通過反復應用馬爾可夫鏈運算器。 ü? 與NICE 和 Real NVE 相比,在 latent code 的大小上沒有限制。 完善GAN:?解決GAN不收斂(non-convergence)的問題:我們面臨的基本問題是,所有的理論都認為 GAN?應該在納什均衡(Nash equilibrium)上有卓越的表現(xiàn),但梯度下降只有在凸函數(shù)的情況下才能保證實現(xiàn)納什均衡。當博弈雙方都由神經(jīng)網(wǎng)絡表示時,在沒有實際達到均衡的情況下,讓它們永遠保持對自己策略的調(diào)整是可能的。??? 我的興趣在于,設計可以在高維、非凸連續(xù)博弈中實現(xiàn)納什均衡(?Nash equilibria)的算法。 4.????? 深度無監(jiān)督學習的未來 ü? 懷疑:因為它會很難知道你要執(zhí)行什么樣的任務。 ü? 深度無監(jiān)督學習的未來將成為半監(jiān)督的學習: Takeru Miyato 等人的虛擬對抗訓練: Distributional Smoothing with Virtual Adversarial Training Virtual Adversarial Training for Semi-Supervised Text Classification 另外還有 Tim Salimans 的帶有特征匹配的GAN的半監(jiān)督學習: Improved Techniques for Training GANs 5.?????概率圖模型的未來:不是相互排斥的 神經(jīng)網(wǎng)絡的大多數(shù)應用可以看作是使用神經(jīng)網(wǎng)絡提供一些條件概率分布的圖模型。 很多新近的神經(jīng)網(wǎng)絡擁有簡單的圖結(jié)構(gòu)( GANs, VAEs 和 NICE都是二分圖( bipartite graph?),讓每個潛變量與每個觀察變量聯(lián)系起來;PixelRNNs/MADE/NADE 都是完整的圖,沒有潛變量)。還不是非常結(jié)構(gòu)化 6.????? 使用批量規(guī)范化(Batch Normalization)會不會削弱深度神經(jīng)網(wǎng)絡的性能 ü??表征能力并不會被影響,因為深度神經(jīng)網(wǎng)絡的規(guī)模和偏移量參數(shù)可以學習抵消規(guī)范化的影響,所以每一層都具有精確學會和以前一樣的功能集的能力。 ü??有效容量(effective capacity)更為復雜。由批量規(guī)范化(Batch Normalization)引入的噪聲具有一種正則化影響,但這可以通過優(yōu)化工作得到極大的改善。 7.????? 我喜歡?dropout,因為從單一模型構(gòu)建指數(shù)級大規(guī)模集合這種觀點太美妙了。 ü? Dropout基本上是用于正則化(regularization)。 它為神經(jīng)網(wǎng)絡引入噪聲以迫使神經(jīng)網(wǎng)絡學會更好的歸納方法以便應付噪聲(這種說法過于簡化了,Dropout 遠不止是在噪聲下的穩(wěn)健性)。 ü? 批規(guī)范化基本上是用于改善優(yōu)化(optimization)。 ??? 其有一個副作用:批規(guī)范化碰巧會向網(wǎng)絡中引入一些噪聲,所以它也可以在模型的正則化上做點貢獻。 ? 當你有一個大型數(shù)據(jù)集時,較好的優(yōu)化就很重要了,較好的正則化就沒有那么重要;所以在大型數(shù)據(jù)集上,批規(guī)范化更重要。你當然也可以同時使用 Dropout 和批規(guī)范化——我在我的 GAN 中這么做過:Improved Techniques for Training GANs 我也認為二分權(quán)重的技巧在近似預測集合方面表現(xiàn)得如此好。 8.????? 解釋為什么批規(guī)范化具有正則化效應(regularzing effect) ??? Batch 形式(batch norm)在某種意義上類似于 dropout ,它在訓練的每一步為每個隱藏單元乘上一個隨機值。在這種情況下,該隨機值是所有 minibatch 內(nèi)隱藏單元的標準差。因為不同實例在每一步驟是針對minibatch 所包含的東西隨機選擇出來的,標準差也是隨機浮動。 ??? Batch norm 也在每一步從隱藏單元減去了一個隨機值( minibatch 的均值)。 這兩種噪音的來源意味著每一層必須學會穩(wěn)健處理輸入的許多變量,就像 dropout 一樣。 9.????? 基于模型的優(yōu)化 將來(從現(xiàn)在到一個有限的時間范圍),我們將能夠使用優(yōu)化算法搜索模型的輸入,這種模型產(chǎn)生最優(yōu)化的輸 出。因為你不能獲得在真實世界中實際最優(yōu)的輸入。相反,你得到的是對抗實例,在模型世界里表現(xiàn)優(yōu)異而在現(xiàn)實世界中卻表現(xiàn)糟糕。 9. 生成式對抗網(wǎng)絡( GAN)未來 常常用于構(gòu)建世界模型的 GAN 現(xiàn)在用于強化學習/動作規(guī)劃,關于生成機器人運動視頻的論文「通過視頻預測的針對物理交互的無監(jiān)督式學習( Unsupervised Learning for Physical Interaction through Video Prediction )」 ? | |
【4】Yoshua Bengio最新兩場講演:表征的深度監(jiān)督學習與深度生成模型 http://www.idiap.ch/workshop/dltm/ ? | 1.?????潛在收益: ü???利用無數(shù)的無標記數(shù)據(jù) ü???回答有關觀察變量的新問題 ü???正則化矩陣-遷移學習-領域適應性 ü???更簡單的優(yōu)化(分而治之) ü???聯(lián)合(結(jié)構(gòu)化的)輸出 2.???????潛在因素和無監(jiān)督表征學習——因果關系。隱藏變量幫助避免維度詛咒。 3.????自編碼的?manifold?與概率解釋 ü???依照歸納原則的降噪評分匹配 ü???能量函數(shù)梯度的評估 ü???通過馬爾科夫鏈取樣 ü???變分自編碼 2??參數(shù)的近似推斷 2??Helmholtz?機的繼任者 2??在對數(shù)似然上最大化變分下界 4.???????GAN:生成式對抗網(wǎng)絡 LAPGAN:生成式對抗網(wǎng)絡的拉普拉斯金字塔 卷積?GANs ALI:Adversarially Learned Inference(VAE & GAN) 5.??????神經(jīng)自回歸模型 ????依據(jù)條件,對觀察指導的模型聯(lián)合進行分解 ????邏輯自回歸 ?神經(jīng)版本 6.????循環(huán)神經(jīng)網(wǎng)絡RNN:一個?RNN?網(wǎng)絡能代表一個完全連接的直接生成式模型:每一個變量都能從之前全部的變量進行預測。 7.????Pixel RNNs ü???近似于?NADE?和?RNNs,但卻是?2-D圖像的 ü???驚人的銳利以及現(xiàn)實的生成 ü???準確得到紋理特征,但卻不需要全局結(jié)構(gòu) | |
【5】谷歌大腦團隊在線答疑,Hinton 壓縮神經(jīng)網(wǎng)絡進展 | 1.????深度學習最新領域 新技術(特別是生成模型)在增強人類創(chuàng)造性方面的潛力 所有關于無監(jiān)督學習和生成模型的近期工作【Bengio】 深度強化學習和針對學習策略的低樣本復雜度算法 2. 重要但尚未被充分研究 對訓練數(shù)據(jù)進行智能自動收集 元數(shù)據(jù)中的系統(tǒng)性問題 將神經(jīng)網(wǎng)絡視為對程序的參數(shù)化表示,而非視為參數(shù)化的函數(shù)逼近器 3.????ML算法所需學習的例子遠遠大于人類學習è數(shù)據(jù)利用率低;ML算法所需的數(shù)據(jù)量高度取決于它要完成的任務 | |
【6】2016 ScaledML會議演講合輯:谷歌Jeff Dean講解TensorFlow & IIya Sutskever :生成模型的近期進展 ? ? ? ? ? ? ? ? 【7】IIya Sutskever :生成模型的近期進展 ? ? ? | 1.? 什么是生成模型? l? 能學習你的數(shù)據(jù)分布 ü? 分配高概率給它 ü? 學習生成合理的結(jié)構(gòu) l? 探索數(shù)據(jù)的「真實」結(jié)構(gòu) 2.?????? 傳統(tǒng)的應用:好的生成模型一定會有以下功能: 結(jié)構(gòu)化預測(例如,輸出文本);更強大的預測; 檢測異常;基于模型的強化學習 3.?????? 推測可以加以應用的領域 非常好的特征學習;?在強化學習中探索;?逆向強化學習;?真正實用的對話;「理解這個世界」;?遷移學習 4.?????? 生成模型的三大類: ü??變化的自動編碼器(VAE) ü??生成對抗式網(wǎng)絡(GAN) ??? 一個生成器 G(z)和一個鑒別器 D(x) 鑒別器的目標是將真實的數(shù)據(jù)從生成器樣本分離出來 生成器嘗試混淆鑒別器 生成對抗式網(wǎng)絡常常會會產(chǎn)生最好的樣本 ü??自動回歸模型 5.?????? 早期有前景的結(jié)果 目前為止任一模型的最好的高分辨率圖像樣本: ü? 深度生成圖像模型使用一個對抗性網(wǎng)絡的拉普拉斯金字塔(Laplacian pyramid)。— Denton 等人 ü??DCGAN?— Radford 等人 6.????? 難以訓練 這個模型被定義在最小的極小化極大算法問題中 沒有損失函數(shù) 很難區(qū)分是否正在取得進展(沒有損失函數(shù),我在訓練時不知道模型訓練的進展如何?) 7.????? 改進 GAN 訓練的簡單想法 ü? GAN 無法學習是因為崩潰問題:(collapse problem) 生成器開始退化并且這個學習也卡主了 ü? 解決方法:鑒別器應該看到整個 mini batch ???如果所有的案例都是相同的,區(qū)別起來就很簡單 8.????? 帶有生成對抗式網(wǎng)絡的半監(jiān)督學習 ü? 鑒別器分辨訓練樣本的類別,也能將真實的樣本從假樣本中辨別出來。 ü? 具體方法的完成過程很重要,但是也需要技術,我不做解釋。 ü? 這個生成對抗式訓練算法也不同 ü? 使用GANs來改進判別模型的新方法 9.????? InfoGAN,Xi Chen,Rein Houthooft 解開的表征Disentangled representations 表征學習的圣杯 ü? 訓練一個GAN ü? 像這樣:它的變量的一個小子集是可從生成的樣本中來精確預測的 ü? 直接添加這個約束 10.? Exploration with generative models Rein Houthfootf,Xi Chen 問題: ü? 在強化學習中,我采取隨機的行動 ü? 有時這些行動做的不錯 ü? 然后我會在未來做更多這些行動 | |
【8】Hinton預言十年內(nèi)將研發(fā)出具有常識的計算機 08-09 新智元 | 1.????? 常識是關于人類世界如何運作的基本知識。它不是建立在規(guī)則上的,也不完全合乎邏輯。它是一套啟發(fā)式教學法,幾乎所有的人類都能迅速掌握。 ??研發(fā)出具有常識的計算機,具備了人類世界如何運作的知識,知道我們的慣例。 | |
【9】LeCun Quora 問答讀后:深度學習走向何方 | 1.? 從統(tǒng)計意義上講,是要估計人體運動所在空間的一個概率分布。這個空間太大,我們用貝葉斯方法和人的先驗知識,控制模型復雜程度。加上BIC這樣的準則,保證在有限數(shù)據(jù)集上訓練出結(jié)果來。 2.???????DBN的觀測模型上,觀測模型本質(zhì)上是要學習從系統(tǒng)內(nèi)部狀態(tài)到外部數(shù)據(jù)表示的一個映射關系。在Jordan的統(tǒng)計框架下用的最多的是混合高斯,混合高斯其實過分抽象了,表現(xiàn)不了數(shù)據(jù)樣本的細微分布。 3.? Lecun提到用深度學習和圖模型做結(jié)合,DL對狀態(tài)到數(shù)據(jù)的映射關系表現(xiàn)能力更強,用圖模型做reasoning。 4.?????? 對狀態(tài)到數(shù)據(jù)的映射關系表現(xiàn)能力更強, 但是模型參數(shù)太多,數(shù)據(jù)有限,學起來太難。模型參數(shù)太多è數(shù)據(jù)有限,學起來太難;搞條件分布,壓縮下供學習的概率空間;對我的生成數(shù)據(jù)來說,受計算能力限制,信息量總是有限的,全random肯定不行,建模時丟了太多信息。怎么把丟的東西找回來,還是得靠知識。 5.?????合成怎么做,真的靠噪聲驅(qū)動模型就可以?否。模型表示能力畢竟有限,此外一個非線性動態(tài)系統(tǒng),趨向于混沌,你就算模型全對,時間一長也沒法預期,加約束 | |
【10】對抗樣本和對抗網(wǎng)絡 | 1.? 對抗 樣本是指將實際樣本略加擾動而構(gòu)造出的合成樣本,對該樣本,分類器非常容易將其類別判錯,這意味著光滑性假設(相似的樣本應該以很高的概率被判為同一類別)某種程度上被推翻了。Intriguing properties of neural networks, by Christian Szegedy at Google, et al,2014. 這篇論文應該是最早提出對抗樣本概念的。 2.? Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images, by Nguyen A, et al, CVPR 15 3.??kdnuggets上的一篇文章(Deep Learning’s Deep Flaws)’s Deep Flaws, by Zachary Chase Lipton指出,深度學習對于對抗樣本的脆弱性并不是深度學習所獨有的,事實上,這在很多機器學習模型中都普遍存在。 4.? Ian Goodfellow:Deep learning Adversarial Examples – Clarifying Misconceptions | |
【11】OpenAI 首批研究成果聚焦無監(jiān)督學習,生成模型如何高 效的理解世界 | 1.?????? OpenAI 的首批研究結(jié)果:在機器學習中提升或使用生成模型。 2.?????? 如何開發(fā)出能分析和理解現(xiàn)實世界大量數(shù)據(jù)的模型和算法? 用作生成模型的神經(jīng)網(wǎng)絡的參數(shù)數(shù)量明顯少于我們用于訓練的數(shù)據(jù)量,所以模型會被迫去發(fā)現(xiàn)和有效地內(nèi)化數(shù)據(jù)的精華以便生成它。 3.?????? 短期應用:圖像降噪、圖像修復,超分辨率、結(jié)構(gòu)化預測、強化學習中的探索。。。 長期來看:它們有自動化學習數(shù)據(jù)集自然特征的潛力,完全不管分類或維度或其它什么東西。 4.?????? 生成式模型目標:目找到網(wǎng)絡參數(shù)θ,使之能夠生成與真實數(shù)據(jù)分布高度匹配的分布。 5.?????? 打造生成模型的三個方法 ü? 生成對抗網(wǎng)絡(GAN:Generative Adversarial Networks) ü? 變化自編碼器(VAE: Variational Autoencoders)讓我們可以在概率圖形模型(probabilistic graphical models )的框架中對這一問題進行形式化——我們在數(shù)據(jù)的對數(shù)似然上最大化下限(lower bound)。 ü? 而 PixelRNN 這樣的自回歸模型(Autoregressive models)則通過給定的之前的像素(左側(cè)或上部)對每個單個像素的條件分布建模來訓練網(wǎng)絡。 6.????? OpenAI近期工作 ü? 改進 GAN。缺陷:方案之間振蕩,或生成器有崩潰的傾向。Tim Salimans、Ian Goodfellow、Wojciech Zaremba 及同事們引入了一些讓 GAN 訓練更穩(wěn)定的新技術。 ü? 為使用 GAN 的半監(jiān)督學習引入了一種方法,該方法涉及到能產(chǎn)生指示輸入的標簽的額外輸出的判別器。 ü? Improving VAE。Durk Kingma 和 Tim Salimans 為變分推理(variational inference)的準確度的提升引入了一種靈活的、在計算上可擴展的方法?!?span style="font-weight:700;">逆自回歸流(IAF: inverse autoregressive flow)」 ü??InfoGAN。Peter Chen 和同事們引入了 InfoGAN——一種可以學習圖像的解開的和可解釋的表征的 GAN 的擴展。無監(jiān)督的學習到好的、解開的表征(disentangled representations) ü??強化學習上的研究,也涉及到了一個生成模型組件:Rein Houthooft 及同事提出了 VIME,一種在生成模型上使用不確定性的實用探索方法。 ü? 生成對抗模仿學習(Generative AdversarialImitation Learning)。Jonathan Ho?及同事呈現(xiàn)了一種用于模仿學習(imitation learning)的新方法。 | |
【11】Generative Adversarial Networks(GAN)的現(xiàn)有工作 程序媛的日常02-29 | 1.? 梳理GAN一系列論文和論文之間的關系發(fā)展軌跡:GANèCGANèLAPGANèDCGANèGRANèVAEGAN 2.? Generative Models【VAE & GAN】: l? VAE將學習的目標變成去盡可能滿足某個預設的先驗分布的性質(zhì)。(在對數(shù)似然上最大化變分下界)這種需要“假設先驗分布”的方式仍然有局限。 l? GAN啟發(fā)自博弈論中的納什均衡, 學習過程就變成了一種生成模型(G)和判別模型(D)之間的競爭過程 3.? 原始GAN. Ian Goodfelow 最小的極小化極大算法問題 GAN?這種競爭的方式不再要求一個假設的數(shù)據(jù)分布,不 用 formulate p(x),而是直接進行 sampling,從而真正達到了理論上可以完全逼近真實數(shù)據(jù)。 【問題】不需要預先建模的方式的缺點就是在于它太過自由了,對于較大的圖片,較多的 pixel的情形,基于簡單 GAN 的方式就不太可控了。在 GAN中,每次學習參數(shù)的更新過程,被設為 D 更新 k 回,G 才更新 1 回,也是出于類似的考慮。 4.? Conditional Generative Adversarial Nets(CGAN) 為了解決 GAN 太過自由,給GAN加約束è條件GAN: 在 D?和 G 的建模中分別加入?conditional 變量 y 5.? 另一方面,為了改進 GAN 太自由的問題,還有一個想法就是不要讓 GAN 一次完成全部任務,而是一次生成一部分,分多次生成一張完整的圖片。(類似于DeepMind的工作DRAW思路:sequential VAE 的模型)。Facebook 等人提出的 LAPGAN[3] 則是采用了這樣的思想,在 GAN 基礎上做出了改進。 在實現(xiàn) sequential version 的方式上,LAPGAN采用了Laplacian Pyramid 的方式。這個方式主要的操作便是 downsample 和 upsample,而優(yōu)勢是每次只考慮樣本和生成圖像之間的殘差的學習效果,某種程度上和 Residual Network 的思想是一樣的。 LAPGAN 其實也是 LAPCGAN,都是?conditional?的;每一步的 GAN 都是independently trained 的。 6.? DCGAN:指出了許多對于 GAN 這種不穩(wěn)定學習方式重要的架構(gòu)設計和針對?CNN 這種網(wǎng)絡的特定經(jīng)驗 開源代碼現(xiàn)在被使用和借鑒的頻率最高,比 LAPGAN 更robust的工程經(jīng)驗分享: ü? Strided convolutional networks作為一個可以 fully differentiable 的 generator G,更加可控和穩(wěn)定。 ü??DCGAN 中則成功將 BN 用在了 G 和 D 上,避免collapse ü? interpolate space,看出圖像逐漸演變過程 7.? GAN 和 LSTM 結(jié)合,稱為?GRAN 改進GAN,可以采用sequential version,好處便是可以讓下一步的 model?利用上一步得到的結(jié)果,在之前的結(jié)果上做出修改,類似于一種?conditional?的方式:通過變成?sequential versions?來減弱?GAN?的自由性。 ü? 因為完美利用了 gradient of convolution 還是 convolution 的性質(zhì),這個改造后的GRAN 可以將每次的求導看做一次 decoding 過程,而每次的 convolution 等操作變成encoding 過程,也就可以因此對應到 DRAW 中的 decoder 和 encoder 部分。 ü? GAN 和DRAW 最大的不同之處: GAN 中在計算 loss 時是在?hidden space?中,而 DRAW 是在原始 input space 中。 generative models?的 evaluation: 可以讓兩組 GAN 互相“競爭”評價?;樵u委,互為選手。 VAEGAN 將GAN 中學出來的 feature 重新運用在 VAE 的?reconstruction objective?中,從而結(jié)合了 GAN 和 VAE 的優(yōu)點。 ü? 以前的reconstruction objective:element-wise distance Metrics,這種metrics其實對于很多hidden feature/space 的學習并不好 ü? idea 就是利用 GAN 中 Discriminator D,使其當做 learned similarity measure,來替代/彌補reconstruction objective 中的這種 similarity measure component。 | |
? | ? | |
主干文章 | Goodfellow I, Pouget-Abadie J, Mirza M, et al.?Generative adversarial nets[C]//Advances in Neural Information Processing Systems. 2014: 2672-2680. | GAN |
Mirza M, Osindero S.?Conditional Generative Adversarial Nets[J]. Computer Science, 2014:2672-2680. | CGAN | |
Denton E L, Chintala S, Fergus R.?Deep Generative Image Models using a Laplacian Pyramid of Adversarial Networks[C]//Advances in neural information processing systems. 2015: 1486-1494. | LAPGAN | |
Radford A, Metz L, Chintala S.?Unsupervised representation learning with deep convolutional generative adversarial networks[J]. arXiv preprint arXiv:1511.06434, 2015. | DCGAN | |
Im D J, Kim C D, Jiang H, et al.?Generating images with recurrent adversarial networks[J]. arXiv preprint arXiv:1602.05110, 2016. | GRAN | |
Larsen A B L, S?nderby S K, Winther O.?Autoencoding beyond pixels using a learned similarity metric[J]. arXiv preprint arXiv:1512.09300, 2015. | ü??GAN + VAE ü? An autoencoder : leverages learned?representations?to better?measure similarities?in data space. ü? Use learned feature representations in the GAN discriminator as basis for the VAE reconstruction objective. | |
Wang X, Gupta A.?Generative Image Modeling using Style and Structure Adversarial Networks[J].?arXiv preprint arXiv:1603.05631, 2016. | ü? Structure-GAN + Style-GAN ü? Current generative frameworks:?end-to-end learning?and?generate images by?sampling?from?uniform noise distribution ü? Basic principle of?image formation:?(a)?Structure: the underlying 3D model; (b)?Style: the texture mapped onto structure. ü? Style and Structure Generative Adversarial Network (S2-GAN) ü? We now explore whether the?representationlearned by the?discriminator network?in our Style-GAN can be?transferredto tasks such as?scene classi | |
Chen X, Duan Y, Houthooft R, et al.?InfoGAN:?Interpretable Representation?Learning by Information Maximizing Generative Adversarial Nets[J].?arXiv preprint arXiv:1606.03657, 2016. | InfoGAN, an?information-theoretic?extension to the?Generative Adversarial Network?that is able to learn?disentangled representations?in a completely unsupervised manner. Maximizes the mutual information between a small subset of the latent variables and the observation. | |
Kurakin A,?Goodfellow?I, Bengio S.?Adversarial examples?in the physical world[J]. arXiv preprint arXiv:1607.02533, 2016. | ü? Adversarial example ü? Even in such?physical world scenarios, machine learning systems?are vulnerable to adversarial examples. ü? A large fraction of adversarial examples are classified incorrectly even when perceived through the camera | |
Salimans T, Goodfellow I, Zaremba W, et al. Improved Techniques for Training GANs[J].?arXiv preprint arXiv:1606.03498, 2016. | ü? A variety of?new architectural features?and?training procedures?to GANs framework. ü? Focus: GANs:?semi-supervised learning, and the?generation of images that humans find visually realistic. ü? we achieve state-of-the-art results in?semi-supervised classification?on MNIST, CIFAR-10 and SVHN ü? The generated images are of high quality as confirmed by a visual Turing test. ü? Learn recognizable features of ImageNet classes. | |
l??Odena A.?Semi-Supervised?Learning with Generative Adversarial Networks[J]. arXiv preprint arXiv:1606.01583, 2016. | We extend Generative Adversarial Networks(GANs) to the?semi-supervised?context by?forcing?the discriminator network to?output class labels. ??? We train a generative model G and a discriminator D on a dataset with inputs belongingto one of N classes. At training time, D is made to predict which of?N+1 classes the input belongs to, where an extra class is added to correspond to the outputs of G. | |
Springenberg J T. Unsupervised and Semi-supervised Learning with?Categorical Generative?Adversarial Networks[J]. arXiv preprint arXiv:1511.06390, 2015. | ü? A method for?learning a discriminative classifier?from?unlabeled?or?partially labeled data. ü? Our approach is based on an?Objective function?that?trades-off?mutual information?between?observed examples?and their?predicted categorical class distribution, against robustness of the classifier to an adversarial generative model. ü? Categorical generative adversarial networks (or CatGAN) – on synthetic data as well as on challenging image classification tasks UNSUPERVISED AND SEMI-SUPERVISED LEARNING OF IMAGE FEATURES | |
? | ? | |
? | ? | |
衍生論文 ? n??? 代表理論性文章 | n??Probst M.?Generative Adversarial Networks in Estimation of Distribution Algorithms for Combinatorial Optimization[J]. arXiv preprint arXiv:1509.09235, 2015. | We integrate a GAN into an EDA and evaluate the performance of this system when solving combinatorial optimization problems with a single objective. GAN-EDA?doe not yield competitive results – the GAN lacks the ability to quickly learn a good approximation of the probability distribution. |
n??Edwards H, Storkey A.?Censoring Representations?with an Adversary[J]. arXiv preprint arXiv:1511.05897, 2015. | ü? This adversary is trying to?predict the relevant sensitive variable?from the representation, and so?minimizing the performance of the adversary?ensures there is?little or no information in the representation about the sensitive variable. ü? We formulate the adversarial model as a?minimax problem, and optimize that minimax objective using a?stochastic gradient alternate min-max optimizer. ü? We demonstrate the ability to?provide discriminant free representations, showing statistically significant improvement across most cases. ü? The flexibility of this method: 2??Removing annotations?from images, from separate training examples of annotated and unannotated images, and?with no a priori knowledge?of the form of annotation provided to the model. | |
Goodfellow I J.?On distinguishability criteria for estimating generative models[J].?arXiv preprint arXiv:1412.6515, 2014. | ESTIMATING GENERATIVE MODELS Generative adversarial networks(GANs) are pairs of generator and discriminator networks, with the generator network learning to generate samples by attempting to fool the discriminator network into believing its samples are real data. We show a variant of NCE, with a dynamic generator network, is equivalent to maximum likelihood estimation. However,we show that recovering MLE for a learned generator requires departing from the distinguishability game. Specifically: | |
Mallat S.?Understanding deep convolutional networks[J]. Phil. Trans. R. Soc. A, 2016, 374(2065): 20150203. | Deep convolutional networks provide state-of-the-art classifications and regressions results overmany highdimensional problems. We review their architecture, which scatters data with a cascade of linear filter weights and nonlinearities. A mathematical framework is introduced to analyse their properties.Computations of invariants involve multiscale contractions with wavelets, the linearization of hierarchical symmetries and sparse separations.Applications are discussed. | |
Li Y, Swersky K, Zemel R. Generative moment matching networks[C] //International Conference on Machine Learning. 2015: 1718-1727. | GANs, whose training involves?a difficult?minimax optimization problem | |
Gauthier J.?Conditional generative adversarial nets for convolutional face generation[J]. Class Project for Stanford CS231N: Convolutional Neural Networks for Visual Recognition, Winter semester, 2014, 2014. | CGAN | |
Yeh R, Chen C, Lim T Y, et al.?Semantic Image Inpainting with Perceptual and Contextual Losses[J]. arXiv preprint arXiv:1607.07539, 2016. | Raymond Yeh 和 Chen Chen 等人的論文「Semantic Image Inpaintingwith Perceptual and Contextual Losses」中的方法,此論文于 2016年 7月 26日 在 arXiv 上發(fā)表。這篇論文演示了如何通過一個?DCGAN用深度學習進行圖像修復。 | |
Koo S.?Automatic Colorization with Deep Convolutional Generative Adversarial Networks[J]. | DCGAN,自動著色 | |
Kwak H, Zhang B T.?Generating Images Part by Part with Composite Generative Adversarial Networks[J]. arXiv preprint arXiv:1607.05387, 2016. | RNN/LSTM + GAN + DRAW | |
Shu R.?Stochastic?Video Prediction?with Deep Conditional Generative Models[J]. | Frame-to-frame stochasticity?remains a big challenge for?video prediction. The use of feed-forward and recurrent networks for video prediction often leads to?averaging of future states. This effect can be attributed to the networks’ limited ability to model?stochasticity. We propose the use of?conditional variational autoencoders (CVAE)?to model frame-to-frame transitions. 【使用DCGAN來做viedo prediction?】 | |
Grosse K, Papernot N, Manoharan P, et al.?Adversarial Perturbations Against Deep Neural Networks for Malware Classification[J].?arXiv preprint arXiv:1606.04435, 2016. | In this paper, we show how to?construct highly-effective adversarial samplecrafting attacks for neural networks used as malware classifiers. The application domain of malware classification introduces additional constraints in the adversarial sample crafting problem when compared to the computer vision domain: | |
l??Odena A.?Semi-Supervised?Learning with Generative Adversarial Networks[J]. arXiv preprint arXiv:1606.01583, 2016. | We extend Generative Adversarial Networks(GANs) to the?semi-supervised?context by?forcing?the discriminator network to?output class labels. ??? We train a generative model G and a discriminator D on a dataset with inputs belongingto one of N classes. At training time, D is made to predict which of?N+1 classes the input belongs to, where an extra class is added to correspond to the outputs of G. | |
Springenberg J T. Unsupervised and Semi-supervised Learning with?Categorical Generative?Adversarial Networks[J]. arXiv preprint arXiv:1511.06390, 2015. | ü? A method for?learning a discriminative classifier?from?unlabeled?or?partially labeled data. ü? Our approach is based on an?Objective function?that?trades-off?mutual information?between?observed examples?and their?predicted categorical class distribution, against robustness of the classifier to an adversarial generative model. ü? Categorical generative adversarial networks (or CatGAN) – on synthetic data as well as on challenging image classification tasks UNSUPERVISED AND SEMI-SUPERVISED LEARNING OF IMAGE FEATURES | |
l??Theis L, Oord A, Bethge M.?A note on the?evaluation of generative models[J]. arXiv preprint arXiv:1511.01844, 2015. | ü? Probabilistic generative models can be used for?compression, denoising, inpainting, texture synthesis, semi-supervised learning, unsupervised feature learning, and other tasks. ü? A lot of heterogeneity exists reviews mostly known but often underappreciated properties relating to the evaluation and interpretation of generative models with a focus on image models. ü? 3 Criteria—average log-likelihood,?Parzen window estimates, and?visual fidelity of samples ü? Extrapolation from one criterion to another is?not warranted?, need to be evaluated directly?with respect to the application(s) they were intended for. ü??Avoid?Parzen window estimates should | |
Kurakin A,?Goodfellow?I, Bengio S.?Adversarial examples?in the physical world[J]. arXiv preprint arXiv:1607.02533, 2016. | ü? Adversarial example ü? Even in such?physical world scenarios, machine learning systems?are vulnerable to adversarial examples. ü? A large fraction of adversarial examples are classified incorrectly even when perceived through the camera | |
Harrigan C.?Deep Reinforcement Learning with Regularized Convolutional Neural Fitted Q Iteration[J]. differences, 14: 1. | ü? We review the deep reinforcement learning setting, in which an agent receiving high-dimensional input from an environment learns a control policy without supervision using multilayer neural networks. ü??Regularized Convolutional Neural Fitted Q Iteration?(RCNFQ) ü? Deep Q Network algorithm (Mnih et al) and?dropout regularization?to?improve?generalization performance. | |
l??Miyato T, Maeda S, Koyama M, et al.?Distributional smoothing?with?virtual adversarial training[J]. stat, 2015, 1050: 25. ? semi-supervised learning | ü? Propose local distributional smoothness (LDS), a new notion of?smoothnessfor?statistical model?that can be used as?a regularization term?to?promote the smoothness of the model distribution. ü? VAT resembles adversarial training, but?it?determines the?adversarial direction?from the?model distribution?alone?without using the label information, making it applicable to?semi-supervised learning. | |
Arild N?kland. Improving Back-propagation by Adding an Adversarial Gradient | ü? A common flaw in several machine learning; Small perturbations?added to the input data lead to consistent misclassification of data samples.(對抗樣本???) ü? Adversarial training has a?regularizing effect?also in networks with logistic, hyperbolic tangent and rectified linear units. ü? A simple extension to the back-propagation: adds an?adversarial gradientto the training. ü? The ”adversarial back-propagation” method increases the resistance to adversarial examples and boosts the classification performance. | |
? | ? | |
應用 | 如何在 TensorFlow 中用深度學習修復圖像? ? | 1.? 通過一個 DCGAN 用深度學習圖像修復。 2.??相關:論文「Semantic Image Inpainting with Perceptual and Contextual Losses」中的方法。 3.? TensorFlow實現(xiàn) |
Yeh R, Chen C, Lim T Y, et al.?Semantic Image Inpainting with Perceptual and Contextual Losses[J]. arXiv preprint arXiv:1607.07539, 2016. | Raymond Yeh 和 Chen Chen 等人的論文「Semantic Image Inpaintingwith Perceptual and Contextual Losses」中的方法,此論文于 2016年 7月 26日 在 arXiv 上發(fā)表。這篇論文演示了如何通過一個?DCGAN用深度學習進行圖像修復。 | |
Koo S.?Automatic Colorization with Deep Convolutional Generative Adversarial Networks[J]. | DCGAN,自動著色 | |
Cate H, Dalvi F,?Hussain Z. DeepFace: Face Generation using Deep Learning[J]. | 人臉生成 | |
Sauer C, Kaplan R, Lin A.?Neural Fill: Content Aware Image Fill with Generative Adversarial Neural Networks[J]. | 圖像補全 | |
l??Creswell A, Bharath A A.?Adversarial Training?For Sketch Retrieval[J].?arXiv preprint arXiv:1607.02748, 2016. | ü? Generative Adversarial Networks (GAN) can learn excellent?representations?for?unlabelled data?which have been applied to?image generation?and?scene classification. ü??Apply to visual search:?show that representations learned by GANs can be applied to visual search. ü? Introduce a novel?GAN architecture with design features?that makes it suitable for?sketch?understanding. | |
Mansimov?E, Parisotto E, Ba J L, et al.?Generating images from captions with attention[J]. arXiv preprint arXiv:1511.02793, 2015. | Motivated by?generative models, we introduce a model that?generates images?from natural?language descriptions. LANGUAGE MODEL:?THE BIDIRECTIONAL ATTENTION RNN IMAGE MODEL: THE?CONDITIONAL DRAW NETWORK LAPGAN Conv-Deconv VAE Fully-Conn VAE alignDRAW | |
l??Reed S, Akata Z, Yan X, et al.?Generative adversarial text to image synthesis[J]. arXiv preprint arXiv:1605.05396, 2016. ? DCGAN | ü? Automatic synthesis of realistic images from text ü? Meanwhile, deep convolutional generative adversarial networks (GANs) have begun to generate highly compelling?images of?specific categories ü? we develop a novel?deep architecture?and?GAN formulation?to effectively bridge these advances in?text and image modeling,?translating visual concepts from characters to pixels. | |
Jianwen Xie, Song-Chun Zhu,Synthesizing Dynamic Textures?and?Sounds?by?Spatial-Temporal Generative ConvNet | Dynamic textures are spatial-temporal processes that. ü??Modeling and synthesizing?dynamic textures?using a generative version of the convolution neural network (ConvNet or CNN) that consists of multiple layers of spatial-temporal filters to capture the spatial-temporal patterns in the dynamic textures. | |
? | ? | |
相關文章 | Gu J, Wang Z, Kuen J, et al.?Recent Advances in Convolutional Neural Networks[J]. arXiv preprint arXiv:1512.07108, 2015. | we provide?a broad survey of the recent advances in convolutional neural networks. Besides, we also introduce some?applications?of convolutional neural networks in computer vision. |
Zhenwen Dai, Andreas Damianou VARIATIONAL AUTO-ENCODED DEEP GAUSSIAN PROCESSES | ü? We develop a scalable deep non-parametric generative model by augmenting deep Gaussian processes with a recognition model. Inference is performed in a novel scalable variational framework where the variational posterior distributions are reparametrized through a multilayer perceptron. ü? We derive a new formulation of the variational lower bound that allows us to distribute most of the computation in a way that enables to handle datasets of the size of mainstream deep learning tasks. | |
Collapsed Variational Inference for Sum-Product Networks Han | ü? Sum-Product Networks (SPNs) are probabilistic inference machines that admit exact inference in linear time in the size of the network. ü? We propose a novel deterministic collapsed variational inference algorithm for SPNs that is computationally efficient, easy to implement and at the same time allows us to incorporate prior information into the optimization formulation. | |
? | ? |
?
?
Facebook 實驗室FAIR上一系列論文:
n??Denton et al. “Deep Generative Image Models using aLaplacian Pyramid of Adversarial Networks” (NIPS 2015)
n??Radford et al. “Unsupervised Representation Learning withDeep Convolutional Generative Adversarial Networks” (ICLR 2015)
n??Mathieu et al. “Deep multi-scale video prediction beyondmean square error”
最后一篇就是用對抗式訓練進行視頻預測的。
Generative Adversarial Networks(GAN)的現(xiàn)有工作
1.《GenerativeAdversarial Nets》(OpenAI)
2.《ConditionalGenerative Adversarial Nets》
3.《Deep GenerativeImage Models using a Laplacian Pyramid of Adversarial Networks》(FAIR)
4.《Unsupervised RepresentationLearning with Deep Convolutional Generative
Adversarial Networks》(FAIR)
5.《Autoencoding beyondpixels using a learned similarity metric》
6.《GeneratingImages with Recurrent Adversarial Networks》
?
谷歌研究室:
谷歌大腦(Google Brain)團隊介紹
關注長期人工智能研究的研究團隊;擁有很多計算機系統(tǒng)和機器學習研究專家;專注純粹的機器學習研究,以及機器人、語言理解、醫(yī)療等新興機器學習應用領域背景中的研究。
我們通過以下幾種方式傳播我們的研究成果:
l??發(fā)表我們的成果,詳情查閱:http://research.google.com/pubs/BrainTeam.html
l??以開源項目的形式發(fā)布了我們的核心機器學習研究系統(tǒng)?TensorFlow
l??發(fā)布我們在 TensorFlow?里面實現(xiàn)的研究模型
l??與谷歌的產(chǎn)品團隊合作,將我們的研究變成真正的產(chǎn)品
目標
?????? Buildartificial intelligence algorithms and system that learn from experience.
構(gòu)建能從經(jīng)驗中學習的人工智能算法和系統(tǒng),并使用這些算法和系統(tǒng)解決困難的問題以造福人類。
DeepMind
?
量子AI
?
OpenAI
人工智能的目標,保證人工智能確實對人類有益
l??深度監(jiān)督學習
視覺,演講,翻譯,語言,廣告,機器人
l??深度監(jiān)督學習
獲得大量輸入—輸出例子
訓練一個非常大的深度神經(jīng)網(wǎng)絡
卷積或者帶有注意力模型的序列到序列(seq2seq with attention)
生成模型:對許多即將出現(xiàn)的模型非常關鍵
什么是生成模型?
l??能學習你的數(shù)據(jù)分布
分配高概率給它
學習生成合理的結(jié)構(gòu)
l??探索數(shù)據(jù)的「真實」結(jié)構(gòu)
創(chuàng)業(yè)公司
有很多創(chuàng)業(yè)公司已經(jīng)以一種令人敬仰的方式成功地應用了深度學習:
ü??Indico 和 DCGANs (與 FAIR 合作)
ü? Quest Visual 和 Word Lens
ü? Nervana 和他們的 CUDA 核
ü? Clarifai 在 2013 年贏得 ImageNet 競賽
總結(jié)
- 上一篇: 如何进行高效JavaScript单元测试
- 下一篇: NoSQL生态系统