Backpropagation in Transformers for Pedestrian Detection

This paper provides a detailed, manual derivation of backpropagation for transformer-based architectures, specifically focusing on layers relevant to next-token prediction and including LoRA layers for parameter-efficient fine-tuning. The authors emphasize the importance of understanding the backward pass for a deeper intuition of how each operation affects the final output, which is crucial for debugging and optimization. The paper's focus on pedestrian detection, while not explicitly stated in the abstract, is implied by the title. The provided PyTorch implementation is a valuable resource.