课堂 Quiz 英文自测|S21-S52
Quiz 2 准备的 32 道非官方英文判断题,覆盖 L04–L07,附答案、理由与课件页码。
手机上宽表格可左右滑动;自测答案默认收起,点击题目下方按钮查看。
非官方自拟练习,考试安排与范围以讲师通知为准。 先用英文作答,再逐题展开解析。
范围说明
Quiz 2 按课表在 L08(11-05)课上,约 20 道英文判断题,占 3%。范围按惯例推断,以课堂公告为准:Quiz 1 在 L04 课末进行、考 L1+L2+L3,照此类推 Quiz 2 覆盖 L04–L07。
复习资料 · 英文自测 1 · 英文自测 2 · 英文自测 3
英文自测
For each statement, write True or False and give a one-line reason. S21–S28 cover L04, S29–S36 cover L05, S37–S44 cover L06, S45–S52 cover L07. These are self-written practice items, not an official Quiz 2.
S21 · Logistic boundary
Because logistic regression applies the non-linear sigmoid function, its decision boundary on the original input features can be non-linear.
S21 · 查看答案与解析
False. sigmoid 单调,p > 0.5 当且仅当 wx > 0,边界 wx = 0 仍是直线,课件原话 Can only model linear relationships(L04 p.7、p.15)。
S22 · Output activation
In an MLP, every layer, including the output layer, must apply a non-linear activation function.
S22 · 查看答案与解析
False. 课件原话是 the output layer may not have a non-linear function,回归输出层可以不加激活(L04 p.18、p.22)。
S23 · Parameter count
A fully connected network with layer widths 4 → 3 → 2 → 3 → 1 and a bias on every non-input neuron has 36 trainable parameters.
S23 · 查看答案与解析
True. (4+1)×3 + (3+1)×2 + (2+1)×3 + (3+1)×1 = 15 + 8 + 9 + 4 = 36;不计偏置是 27,对应课件 3i+24 / 3i+15 在 i = 4 时的值(L04 p.13)。
S24 · XOR with a linear model
For the four XOR points, the best least-squares fit of y = w₁x₁ + w₂x₂ + w₃ outputs 0.5 for every input.
S24 · 查看答案与解析
True. 最小二乘解 w = (0, 0, 0.5),四点都输出 0.5,MSE 0.25,与课件打印的 error 0.25 一致(L04 p.14)。
S25 · Stuck loss
If a binary classifier outputs 0.5 for every training example, its average cross-entropy loss is about 0.693.
S25 · 查看答案与解析
True. 每个样本 CE = −ln 0.5 = ln 2 ≈ 0.6931,课件打印值 0.6931471805599453(L04 p.15)。
S26 · Multi-class output
For multi-class classification, the lecture pairs a sigmoid output layer with the MSE loss.
S26 · 查看答案与解析
False. 多分类配 softmax + CE;sigmoid 配 CE 是二分类,MSE/MAE 用于回归(L04 p.22)。
S27 · Paths in BP
In the L04 example network (2 inputs, two hidden layers of 2 sigmoid units, 1 output), the gradient of the first-layer weight w₁ involves only one chain-rule path.
S27 · 查看答案与解析
False. h₁ 同时连向 s₃ 和 s₄,dℓ/dw₁ 是两条路径之和(L04 p.31)。
S28 · BP versus GD
Backpropagation computes the gradients efficiently by reusing intermediate values; the weights are then updated by gradient descent.
S28 · 查看答案与解析
True. BP 是梯度下降第 3 步的求导方法,反向复用中间变量;第 4 步的更新由梯度下降完成(L04 p.36)。
S29 · Batch size
Larger batches give a more accurate gradient estimate, so they always produce the best final model.
S29 · 查看答案与解析
False. 课件原话 more accurate estimate of the gradient, but not the optimal,小 batch 还有正则化效果(L05 p.11)。
S30 · Momentum arithmetic
Minimising L(θ) = θ² with momentum (v ← αv − εg, θ ← θ + v), θ₀ = 1, v₀ = 0, ε = 0.1 and α = 0.9, gives θ = 0.46 after two steps.
S30 · 查看答案与解析
True. 第 1 步 v = −0.2、θ = 0.8;第 2 步 v = 0.9×(−0.2) − 0.1×1.6 = −0.34、θ = 0.46。普通梯度下降两步是 0.64(L05 p.12 的更新式,数值为注解自拟例)。
S31 · Sigmoid hidden layers
Using sigmoid in the hidden layers always causes gradient vanishing.
S31 · 查看答案与解析
False. 原话 is likely to result in gradient vanishing,把 likely 换成 always 就过头了(L05 p.15)。
S32 · ReLU
The derivative of ReLU is either 0 or 1; this mitigates gradient vanishing but can lead to dead neurons.
S32 · 查看答案与解析
True. 导数为 1 时梯度原样传回;s < 0 时输出和导数都是 0,出现 Dead neurons(L05 p.16)。
S33 · Sparsity
L2 regularization tends to drive many weights to exactly zero, giving a sparse W.
S33 · 查看答案与解析
False. 课件说的是 L1 regularization tends to learn sparse W;L2 的拉力 2αw 随 w 变小而变弱,只把权重压小(L05 p.21)。
S34 · L2 gradient
With the regularized loss α‖W‖² + ℓ(W), the gradient with respect to w₁ is 2αw₁ + dℓ/dw₁.
S34 · 查看答案与解析
True. α‖W‖² 中只有 αw₁² 和 w₁ 有关,求导得 2αw₁;常见错误是漏掉系数 2(L05 p.19)。
S35 · Early stopping
Early stopping halts training when the training error reaches its minimum.
S35 · 查看答案与解析
False. 应在验证误差最低处停;训练误差会一直下降,那时已经过拟合(L05 p.22)。
S36 · Universal approximation
Since a single hidden layer can represent any function, deep networks are never needed in practice.
S36 · 查看答案与解析
False. 同页紧接两条限制:the layer may be infeasibly large;may fail to learn and generalize correctly(L05 p.25)。
S37 · Output size
Convolving a 7×7 input with a 3×3 kernel, stride 2 and no padding, gives a 3×3 feature map.
S37 · 查看答案与解析
True. (7 − 3)/2 + 1 = 3(L06 p.12 的公式,stride 2 为注解自拟变式)。
S38 · Padding
To keep a 128×128 input at 128×128 after an 11×11 convolution with stride 1, you pad 5 pixels on each side.
S38 · 查看答案与解析
True. stride 1 时总共补 k − 1 = 10,每边 5,(128 + 10 − 11)/1 + 1 = 128(L06 p.18)。
S39 · Parameters and image size
A convolutional layer with 10 kernels of size 5×5 on a 32×32×3 input has 760 parameters including biases; on a 64×64×3 input the same layer would have twice as many.
S39 · 查看答案与解析
False. 前半句对:5×5×3×10 + 10 = 760。后半句错:参数量 k×k×C_in×C_out 和输入长宽无关,64×64×3 输入仍是 760(L06 p.15)。
S40 · Multiple input channels
In a standard convolutional layer, each kernel processes only one input channel, so an RGB input needs a separate kernel per colour channel.
S40 · 查看答案与解析
False. 每个核的深度等于输入通道数,在各通道上相乘后全部相加,得到一张输出图(L06 p.15,课堂练习 Q7)。
S41 · Max-pooling
A 2×2 max-pooling layer with stride 2 halves the height and width, keeps the number of channels, and has no learnable parameters.
S41 · 查看答案与解析
True. 例:128×128×32 经 2×2、stride 2 得 64×64×32;pooling 对每个通道单独做,没有可学习参数(L06 p.20)。
S42 · VGG
Three stacked 3×3 convolutions have the same 7×7 receptive field as one 7×7 convolution but fewer weights (27 versus 49 per input–output channel pair).
S42 · 查看答案与解析
True. 感受野 1 + 3×(3 − 1) = 7;权重 3×9 = 27 对 7×7 = 49,课件原话 Use 3 x 3 kernels, less parameters(L06 p.25)。
S43 · ResNet
ResNet’s skip connections add the block input to the block output, which lets gradients flow back to earlier layers.
S43 · 查看答案与解析
True. 残差块输出 = F(x) + x,课件原话 Use skip-connections, gradients can be back-prop(L06 p.26)。
S44 · Deeper layers
Every deeper layer of a CNN is guaranteed to learn more semantic features than the layer before it.
S44 · 查看答案与解析
False. 原话 Deeper layers tend to learn more semantic meaningful patterns,tend to 不是 guaranteed(L06 p.27)。
S45 · Fixed input size
MLPs and CNNs assume that each input data point has a fixed number of dimensions.
S45 · 查看答案与解析
True. 课件红字 Each input data point has a fix(ed) number of dimensions,这正是要引入序列模型的原因(L07 p.8)。
S46 · Sentiment classification
Sentiment classification of a sentence is a one-to-many task.
S46 · 查看答案与解析
False. 一串词输入、一个标签输出,课件归为 Many to One;One to Many 的例子是 Image Captioning(L07 p.11)。
S47 · RNN parameters
In a vanilla RNN, the number of parameters grows with the length of the input sequence, because each time step has its own weight matrices.
S47 · 查看答案与解析
False. W_hh、W_hx、W_yh 在所有时间步共用,参数量和序列长度无关(L07 p.13)。
S48 · LSTM gates
In an LSTM, the forget gate, the input gate and the output gate all use the sigmoid function, so their values lie between 0 and 1.
S48 · 查看答案与解析
True. f_t、i_t、o_t 都是 σ(…),取值在 0 到 1 之间;只有候选 g_t 用 tanh(L07 p.15)。
S49 · LSTM and gradients
Thanks to the cell state, LSTM completely eliminates the gradient vanishing and exploding problems.
S49 · 查看答案与解析
False. 课件在 LSTM 一节仍把 Gradient exploding and vanishing problems 列为缺点,cell state 只是缓解(L07 p.18)。
S50 · Context vector
In attention with an RNN, the context vector is a weighted sum of all encoder features, and the weights come from a softmax so they sum to 1.
S50 · 查看答案与解析
True. c_i = Σ_j α_ij h_j,α_ij 由 e_ij 做 softmax 得到,和为 1(L07 p.21)。
S51 · Scaling
In scaled dot-product attention, the scores QKᵀ are divided by d_k, the dimension of the keys.
S51 · 查看答案与解析
False. 公式是 softmax(QKᵀ/√d_k)V,除的是 √d_k 不是 d_k(L07 p.23)。
S52 · Sequence to sequence
A sequence-to-sequence model can be viewed as many-to-one followed by one-to-many, so the output sequence may differ in length from the input.
S52 · 查看答案与解析
True. 课件把 seq2seq 标为 Many to One + One to Many,英文问句译成西班牙文,长度可以不同(L07 p.16)。
这组材料已读完,可以回课程页选择配套资料或练习:COMP5541 课程页。
个人整理的学习笔记,不是官方材料;数字与结论以课件和讲师为准。