Question 1 of 25

Why do CNNs work better than plain MLPs on images?

Question 2 of 25

What does parameter sharing in a conv layer mean?

Question 3 of 25

Pooling (e.g. max pool 2×2) mainly helps by:

Question 4 of 25

An RNN is designed for:

Question 5 of 25

Why do plain RNNs struggle with long sequences?

Question 6 of 25

In sentiment analysis word-by-word, the RNN hidden state at step *t* summarizes:

Question 7 of 25

What problem does LSTM mainly address vs a vanilla RNN?

Question 8 of 25

GRU compared to LSTM:

Question 9 of 25

The LSTM forget gate decides:

Question 10 of 25

What is a word embedding?

Question 11 of 25

Why are embeddings important for GenAI?

Question 12 of 25

Pre-trained embeddings (e.g. GloVe, Word2Vec) vs training from scratch on 500 reviews:

Question 13 of 25

A 3×3 conv with stride 1 on a 32×32 map (no padding) produces:

Question 14 of 25

In sentiment analysis, averaging word embeddings for a review is:

Question 15 of 25

For product review sentiment, why might LSTM beat a bag-of-words model?

Question 16 of 25

Stride 2 convolution (vs stride 1) typically:

Question 17 of 25

Why share the same filter weights across all image locations in a conv layer?

Question 18 of 25

In an RNN, hₜ (hidden state at time t) mainly summarizes:

Question 19 of 25

The LSTM input gate decides:

Question 20 of 25

GRU compared to LSTM often:

Question 21 of 25

Words with similar meaning (e.g. king and queen) in a good embedding space:

Question 22 of 25

Backprop through time (BPTT) means:

Question 23 of 25

Padding (e.g. same padding) around a conv layer is used to:

Question 24 of 25

Word2Vec / GloVe pre-trained embeddings help downstream tasks because:

Question 25 of 25

For very long documents, plain RNNs struggle mainly because: