NCA-GENM Question 194
Select 3You are tasked with building a U-Net model for a generative AI system that can create images from pure noise. Which of the following architectural choices and configurations are critical for achieving this functionality?
- A
Using skip connections to preserve spatial information across encoding and decoding stages
- B
Training the model with a pixel-wise loss function like Mean Squared Error (MSE)
- C
Replacing convolutional layers with recurrent layers to capture temporal dependencies
- D
Adding upsampling layers in the decoder to reconstruct high-resolution images
- E
Utilizing a fully-connected layer at the end of the decoder to produce the final image
Show answer and explanation
Correct answers: A, B, D
Explanation
The U-Net architecture is specifically designed for tasks that require spatial information preservation during encoding and decoding. Skip connections ensure that spatial features are retained, upsampling layers reconstruct the original resolution, and pixel-wise loss functions guide the model in generating accurate outputs. Recurrent layers and fully-connected layers are not suitable for the spatial image generation tasks the U-Net is built for, as they either focus on temporal dependencies or lose spatial structure.
- A. Correct.
Skip connections are a fundamental part of the U-Net architecture, allowing the model to preserve spatial details from the encoder and pass them to the decoder for accurate image reconstruction.
- B. Correct.
Pixel-wise loss functions like MSE are often used in generative tasks to measure the difference between the generated output and the ground truth, making them suitable for training the U-Net.
- C. Incorrect.
Recurrent layers are not needed in this case, as U-Net is designed for spatial, not temporal, data processing.
- D. Correct.
Upsampling layers in the decoder are essential for reconstructing the high-resolution image from the compressed feature representation produced by the encoder.
- E. Incorrect.
Fully-connected layers are not typically used at the end of a U-Net decoder, as they lose spatial information and are unsuitable for image generation tasks.