Questions and published explanations below are available without starting a test. Some questions may not have a published solution yet.
0 / 23 published solutions
Question 2 MCQ · 3.0 marks
Assume that the shape of the input image is 32 x 32 x 100. A 1x1 convolution with 50 filters is
applied. How many learnable weights will be present (excluding biases)?
A.
50
B.
100
C.
500
D.
5000
A published solution is not available for this question yet.
Consider the following statements about the Transformer architecture.
1. The encoder uses self-attention.
2. The encoder uses masked self-attention.
3. The decoder uses masked self-attention.
4. The decoder uses cross-attention.
5. The encoder uses cross-attention.
Which of the following is correct?
A.
1 and 2 only
B.
1, 3 and 4 only
C.
2, 3 and 4 only
D.
1, 2, 3 and 4
E.
1, 3 and 5 only
F.
1, 2, 3, 4 and 5
A published solution is not available for this question yet.
Which model is described as producing **blurry** images because its loss averages out fine details,
but has **high diversity** and **stable training**?
A.
GAN
B.
VAE
C.
DDIM
D.
Classifier
A published solution is not available for this question yet.
In a GAN trained using the original minimax (saturating) generator objective, suppose the
discriminator becomes nearly perfect very early in training. What will be the most likely
consequence of this?
A.
The generator learns faster because the discriminator provides perfect
classification.
B.
Both the generator and discriminator converge immediately to the optimal
solution.
C.
The generator's loss becomes zero, indicating perfect image generation.
D.
The generator experiences mode collapse.
E.
The generator stops learning because it receives very small (vanishing)
gradients.
A published solution is not available for this question yet.
A GAN suddenly starts producing almost the same single image regardless of the input noise z,
even though the discriminator is fooled. This failure is called:
A.
Mass covering
B.
Posterior collapse
C.
Mode collapse
D.
Reparameterization
A published solution is not available for this question yet.
Suppose the residual connections in the transformer block are removed, so the forward pass
becomes:
What is the most likely consequence of this change as more transformer blocks (
) are
stacked?
A.
No meaningful effect - LayerNorm already keeps activations well-scaled, so the
residual path is redundant.
B.
Gradients have no direct path back to earlier layers, so as depth increases they
must propagate entirely through each sub-layer's transformation, making them prone to
vanishing or exploding and the network much harder to optimize.
C.
The model would no longer be able to apply the causal mask correctly, since
masking depends on the residual stream.
D.
Training would fail immediately with a shape mismatch error, since the
residual connection is required to keep tensor dimensions consistent.
E.
Gradients have no direct path back to earlier layers, which causes the
attention weights to exceed valid probability values and produces NaN losses within the first few
training iterations.
A published solution is not available for this question yet.
Consider a Transformer with an embedding dimension of 24 and 4 attention heads. For a
sequence containing 10 tokens, what are the correct dimensions of query, key, and the attention
score matrix for a single attention head?
A.B.C.D.E.
A published solution is not available for this question yet.
A ViT processes grayscale images of resolution 128 x 128 pixels, split into non-overlapping 16 x 16
patches. Each flattened patch is projected by a single shared linear layer to a 256-dimensional
embedding, and a learnable [CLS] token is prepended before the Transformer encoder. Ignoring
biases, select all correct statements.
A.
Each flattened patch has length 768.
B.
The projection layer contains 65,536 weights.
C.
The projection layer contains 4,194,304 weights (one copy per patch).
D.
The sequence length entering the encoder is 65.
A published solution is not available for this question yet.
Consider a single layer transformer with the following specifications:
Vocabulary Size=6400
Number of heads= 4
Dimension for each head= 16
Hidden layer size for the feed forward block = 200
The feed forward block only contains one hidden layer.
Calculate the total trainable weights between the input layer and hidden layer of the feed forward
block. (Ignore bias)
A published solution is not available for this question yet.