Hacker News Reader

Comment by ImprobableTruth | Hacker News Reader

Both can introduce wrong information or cause artifacts. "It only uses the available information" is a red herring.

Then this is a problem inherent with digital photography. Any image you see is interpolated from Bayer data. There are multiple algorithms with different results, the colour needs to be adjusted, and the camera distortion needs to be reversed. All of this happens before you see anything on your screen.

A zooming or scaling algorithm that only uses available image data is deterministic, it should get the same result every time.

One that relies on a neural network trained on other images is not, it might produce a subtly different image each time it's run and introduce artefacts from other images.

What? Not only is this a bizarre argument (why would deterministically creating wrong information be okay?), it's not even correct. There's nothing that fundamentally prevents an image scaling NN from being deterministic and I doubt you think that deterministically adding someone's face would be fine, right?

‘Wrong’ information, if it’s expected, understood, and predictable, is still usable. Especially if we know it’s not going to introduce completely alien artefacts into an image like Ryan Reynold’s face.

All of the standard zoom algorithms fall into that category. Whether bilinear, nearest neighbour, or similar, they’re all going to produce similar outcomes with which, importantly, a teenaged witness on the stand will be familiar. Same with the way TVs upscale content.

It’s not like this is new evidence being introduced in its zoomed-in form. It’s an iPad and footage or images being used in the same way prosecutors and defence attorneys have used them for years when cross-examining witnesses.

The reason it’s okay to be deterministic and wrong in engineering is because it’s a lot easier to explain the conditions under which the wrong result will be produced and how you can identify those situations in the specific photo. The reason I think a simple resize that deterministically adds misinformation is better than a neural network is because I can calculate the bounds under which the algorithm produces wrong information and label the image. And I can do that without running the algorithm.

What stops a neural network zoom to be just labelled as "zoom"? I think its a very pertinent question from the judge.

Asking a question is fine, accepting the defence's claim without further substantiation and the rest of the follow up is less so.

And how do you resolve conflicting technological claims from the defense and prosecution? By expert witness. Which is what the judge did. So what’s the problem?

That's not what happened. The judge placed the burden of proving that zooming didn't do that on the prosecution, and only afforded them a few minutes to find an expert witness.

In the end the iPad was not used.

Untill iPad is open source, you can't trully know what they did to the image

Neural network is deterministic, in the same sense that bilinear interpolation is deterministic.

Reply on news.ycombinator.com