Hacker News Reader

Comment by Mitt | Hacker News Reader

This whole thing is a difficult issue. Who is to decide what devices look similar enough, to deserve calling it a “copy”? In principle what we do here is: we implement a neural network for object recognition, and train it to recognize a certain device. Now we run this classifier on other devices and need to settle for a confidence value which says if a device is a copy or not. But who should decide what exactly this confidence value should be?

For example, what if we decide to say that a 0.921 value means a device is a copy, and we find out that Samsungs older devices get a 0.884 confidence?

All those values are very artificial.

View on HN

The law will always have to deal with continuous and imprecisely defined concepts. Speed limits, drinking age, gross vs non-gross misconduct, etc. Continuity alone is a weak argument against anything, sometimes it's so weak that it has its own fallacy: http://en.wikipedia.org/wiki/Continuum_fallacy

So, while I believe the intellectual "property" should be eliminated, it's not because of the continuity involved in enforcing it.

In order to train your Neural Net you will need sample data; which will consist of examples of what are and are not "similar" devices. Someone will have to decide what looks similar enough to the object device, therefore your Neural Net will be as objective as the sample data.

Don't think of Neural nets as a higher order intelligence - it's only as clever as the sample data used, but rather something that can make consistent decisions.

>This whole thing is a difficult issue. Who is to decide what devices look similar enough, to deserve calling it a “copy”?

Here's a guideline: if there is a document in your company saying ("let's copy this thing") --as in Samsung's case--, it's a copy.

Or if you created your design from leaked photos of prototypes ("as in this case") --it's a copy.

For the rest, how about a "jury" (i.e impartial to both parties common people)?

Even when we directly say “We’ll copy it!”, it could still come out as having the classifier outputting a small confidence value. So, in my opinion this should not count. There should be a confidence value of a neural network involved.

In the case of a real copy, the classifier should have a confidence of 1.0, or be enormously close to it.

Reply on news.ycombinator.com