Huawei HCIP-AI-EI Developer (H13-321) Certification Sample Questions

02. At time step t in an LSTM layer, the forget gate outputs values close to 0 in every dimension, while the input gate outputs values close to 1.
What happens to the cell state at this step?
a) It keeps its previous contents and adds the new candidate values on top
b) It is largely replaced by the new candidate values computed at this step
c) It passes through unchanged, and only the hidden state is overwritten
d) It is reset to zero, and the new candidate values are discarded

03. In a scaled dot-product attention layer, the query and key vectors have dimension d_k = 64, and their components are independent with mean 0 and variance 1.
What is the approximate standard deviation of an unscaled score q·k, and what does dividing by √d_k accomplish?
a) About 8; dividing by √d_k restores unit variance, so softmax avoids saturating
b) About 8; dividing by √d_k makes every row of attention weights sum to one
c) About 64; dividing by √d_k lowers the variance to 8, which enlarges the softmax gradients
d) About 1; dividing by √d_k rescales the value vectors to unit length

04. A Transformer block has d_model = 512 and a position-wise feed-forward network with an inner dimension d_ff = 2,048.
Ignoring biases, how many weights does this feed-forward network contain, and how is it applied?
a) About 1.05 million, shared across all positions
b) About 2.1 million, applied identically at each position
c) About 2.1 million, with a separate copy for each position
d) About 4.2 million, applied identically at each position

05. A layer maps a 256-channel feature map to 256 channels with a 3×3 convolution. A bottleneck block replaces it with a 1×1 convolution down to 64 channels, a 3×3 convolution from 64 to 64 channels, and a 1×1 convolution back up to 256 channels.
Ignoring biases, how many weights does the bottleneck block contain?
a) 36,864
b) 147,456
c) 589,824
d) 69,632

06. During training, a Transformer decoder receives the whole target sentence at once through teacher forcing, and its self-attention sublayer applies a triangular look-ahead mask.
What does this mask accomplish?
a) It stops each position from seeing later target tokens
b) It hides the padding positions so that sentences of different lengths can share one batch
c) It randomly hides a share of the target tokens so the decoder learns to reconstruct them
d) It blocks the decoder from the encoder output until the whole target sentence has been read

07. A team doubles the input length of a self-attention layer from 2,048 to 4,096 tokens while keeping the model width, head count, and batch size unchanged.
By roughly what factor does the memory needed to hold the attention score matrices grow?
a) About 2 times
b) About 8 times
c) About 4 times
d) About 16 times

08. A CTC model emits one symbol per frame, where - is the blank symbol. Decoding first merges consecutive repeated symbols and then removes blanks.
Which frame-level output decodes to the word hello?
a) h h e - l l l o
b) h - e e l - l o
c) h e - l l o o -
d) - h e l l - o o

09. An engineer implementing the Canny edge detector inspects intermediate output. After Gaussian smoothing and gradient computation, the edges in the gradient-magnitude image are several pixels wide.
Which subsequent Canny step reduces these edges to one-pixel-wide ridges?
a) Hysteresis thresholding with a high and a low threshold
b) A second pass of Gaussian smoothing with a larger sigma
c) Applying the Laplacian to locate second-derivative zero crossings
d) Non-maximum suppression along the gradient direction

10. In a FastSpeech-style model, the encoder outputs one hidden vector for each of four phonemes. The duration predictor assigns them 3, 2, 5 and 2 mel frames respectively.
What does the length regulator pass to the decoder?
a) 12 vectors, made by repeating each phoneme vector for its predicted frame count
b) 12 vectors, made by the decoder generating frames one at a time until a stop token
c) 4 vectors, which the decoder expands internally through encoder-decoder attention
d) 48 vectors, made by pairing every phoneme vector with every output frame position