Which architecture uses attention mechanisms to enable parallel processing of input features?

• A. CNN
• B. RNN
• C. Transformer
• D. GAN

Answer: (C)

Attention that considers every position in the input at once enables processing all features in parallel. In a transformer, self-attention lets each position attend to all other positions, computing those relationships in a single pass. This means the model can evaluate and combine information across the entire sequence simultaneously, which makes training and inference faster on modern hardware and allows it to capture long-range dependencies effectively.

Other architectures don’t provide that same parallel, global view. Convolutional networks operate on local neighborhoods and, while parallelizable, rely on stacking many layers to widen their context. Recurrent networks process one step at a time, so parallelization across the sequence is limited. Generative adversarial networks describe a different training setup focused on GAN objectives, not specifically on how input features are parallel-processed via attention.