Reference no: EM133296810

In "Colorful Image Colorization", the authors adopt a "multinomial" or categorical loss function for color prediction. Which of the following is NOT a reason for this?

1. The baseline regression loss has a tendency of encouraging deep networks to predict the average color so that it never pays large penalties for drastically wrong color predictions. On the other hand, the multinomial loss pays the same penalty for any wrong color prediction so it is less biased towards predicting the average.

2. Multinomial losses require less memory because the output is discrete instead of continuous.

3. Multinomial outputs provide a measure of uncertainty, e.g. the deep network can express that an output might be red or green but not blue. With a regression loss, you only get a single prediction.