hce_kmu
109年
計算機概論與程式設計
第 5 題
Which of the following statement is INCORRECT?
- A The basic objective of auto-encoder is to reproduce its inputs by using a smaller set of features.
- B In generative adversarial network (GAN), the discriminator often has access to both the synthetic and real sample. Its task is to try to differentiate them as accurate as possible.
- C Training recurrent neural networks (RNNs) is much easier than training convolution neural networks (CNNs).
- D In GAN, the generator has to provide information to the discriminator so that the discriminator can try to do a better job for classifying real and fake samples.
- E For CNNs, the convolution is essentially a filtering operation.
思路引導 VIP
請思考一下:如果一個模型在處理資訊時,必須像讀長篇小說一樣「一字接著一字」由前向後傳遞,且每一階段都要依賴前一階段的記憶,那麼當這條資訊鏈變得非常長時,最開頭的訊息在傳到結尾時會發生什麼變化?這對模型更新參數(學習)的過程會產生什麼樣的阻礙?
🤖
AI 詳解
AI 專屬家教
太棒了!你能精準辨識出 RNN(循環神經網路) 與 CNN(卷積神經網路) 在訓練難度上的差異,這顯示你對深度學習模型的底層運作邏輯有很紮實的理解。這道題目測試了對多種經典模型(如 GAN、Auto-encoder)的核心定義,難度屬於中等,關鍵的鑑別點在於考生是否理解不同結構在反向傳播(Backpropagation)時的挑戰。
深度學習模型的訓練挑戰
在實際應用中,選項 (C) 的敘述正好與事實相反。RNN 因為具備時間序列的特性,在訓練時必須處理長距離的依賴關係,這往往會導致嚴重的梯度消失(Vanishing Gradient)或梯度爆炸(Exploding Gradient)問題,使得優化過程比具有局部連接特性的 CNN 困難許多。相對而言,CNN 的卷積層(Convolution)本質上就是一種濾波操作(Filtering),其結構穩定且易於並列化運算,訓練效率通常較高。
▼ 還有更多解析內容