hce_kmu
109年
計算機概論與程式設計
第 3 題
Which of the following statement is INCORRECT?
- A In designing a machine learning system, signal processing methods can be a very valuable data pre-processing tool.
- B When the training and testing errors of a neuron network is becoming larger and larger, we should terminate the training process to avoid the overfitting problem.
- C The key of human intelligence comes from the connection weights between the neuron.
- D Generally speaking, one of the difficulties encountered by most of the optimization methods is in finding global minimum.
- E Compared to batch learning, the sequential learning method can more accurately estimate the gradient of the error function.
思路引導 VIP
想像你正帶領一群人下山，若想找出整體下降最快的方向，是「先詢問隊伍中所有人的意見後取平均」比較準確，還是「隨便抓一個人的意見就當作全體前進的方向」比較能代表整體的真實路徑呢？
🤖
AI 詳解
AI 專屬家教
太棒了！你能精準辨識出選項 (E) 的敘述邏輯錯誤，展現了對機器學習基礎觀念的紮實掌握。這題的鑑別度在於區分「學習效率」與「梯度估算準確性」，許多學習者容易混淆兩者的優劣。
批次學習與循序學習的梯度差異
在優化過程中，批次學習 (Batch Learning) 會計算整個數據集的平均損失函數梯度，這在數學上是最能反映真實損失函數下降方向的。而循序學習 (Sequential/Online Learning) 或隨機梯度下降，每次僅根據單一樣本來更新權重。雖然循序學習計算速度快，且因其「噪聲 (Noise)」特性有助於跳出局部最小值 (Local Minimum)，但若論及對「整體誤差函數梯度」的估計，它必然遜於利用全體樣本計算出的精確方向。
▼ 還有更多解析內容