Deep Learning and Generative AI Week 9 Graded Assignment Solutions

🧠 Deep Learning and Generative AI Week 9 Graded Assignment Solutions

Version: v1_t32025 —
Complete solutions covering Seq2Seq Models, Transformer Architecture (Multi-Head Attention, Positional Encoding), and Generative Model Theory.

This downloadable PDF bundle provides complete, detailed solutions for the ninth week of “Deep Learning and Generative AI”.
The assignment focuses on the mechanics and components of advanced sequence models and Attention, which form the backbone of modern LLMs.

Core Assignment Topics & Question Details (Approx. 10–12 Questions)

The solutions cover architecture design, computational steps, and sequence theory:

• Seq2Seq & Attention Fundamentals:
  • Analyze the function of the Encoder–Decoder architecture and the bottleneck issue addressed by Attention.
  • Determine the correct procedure for computing Attention Weights (e.g., dot product scaling) based on Query/Key matrices.
• Transformer Components:
  • Identify the purpose and necessity of Positional Encoding in the Transformer architecture.
  • Analyze the computational advantage of Multi-Head Attention over single-head attention.
  • Determine the input/output dimensions for the Feed Forward Network (FFN) layers within the Transformer block.
  • Analyze the role of the Masked Multi-Head Attention layer used specifically in the Decoder block.
• Training & Decoding:
  • Calculate the output shapes and dimensions for different layers (e.g., linear layers, softmax) within the Transformer.
  • Analyze decoding strategies (e.g., greedy vs. beam search) used for sequence generation.
• Generative Model Theory:
  • Analyze theoretical questions related to advanced generative models like Variational Autoencoders (VAEs) or Conditional GANs (CGANs), focusing on their loss functions or data conditioning techniques.

File Format:

• v1_t32025: Complete solution set.