Which architecture is typically favored for tasks involving sequential data and memory of past elements?

• A. Long short-term memory (LSTM) networks
• B. Transformer architecture
• C. Generative adversarial networks (GANs)
• D. Model assessment

Answer: (A)

Maintaining memory across time steps in sequential data is essential when past elements influence the present prediction. Long short-term memory networks achieve this with a memory cell and gating mechanisms that control information flow. The forget gate decides what to discard from the memory, the input gate determines what new information to add, and the output gate decides what part of the memory to output at the current step. This structure helps preserve relevant information over long sequences and mitigates vanishing gradients, making them particularly well-suited for tasks where context from far in the past matters. While transformers use attention to relate all positions in a sequence and have become very effective, the explicit, step-by-step memory retention of LSTMs aligns directly with maintaining past information as data streams in, which is why they are typically favored for sequential data with memory of past elements. GANs don’t model temporal memory; they’re for generative adversarial tasks. Model assessment isn’t an architecture.