Just started to explore deep learning and chose Juice as the starting framework, since I want to stick with Rust. Since I'm pretty new to the domain, it might be just my mistake.

I was looking at the NegativeLogLikelihood::compute_output() and I think there is a bug. Instead of

        for &label_value in native_labels {
            let probability_value = native_probabilities[label_value as usize];
            writable_loss.push(-probability_value);
        }

it should be

        let mut offset = 0;
        for &label_value in native_labels {
            let probability_value = native_probabilities[offset + label_value as usize];
            writable_loss.push(-probability_value);
            offset += self.num_classes;
        }

Otherwise we're comparing all labels in the batch with the first output from the batch?

Interesting that I've tried changing it and there were no noticeable effect on MNIST example...