News for hackers. Stuff that matters.
Re:Implementing a Convolutional Neural Network from Scratch ... by amelius on Wednesday May 22, 2019 @02:52PM (#19984266)
> If you trained a network to detect dogs, you’d want it to be able to a detect a dog regardless of where it appears in the image. Imagine training a network that works well on a certain dog image, but then feeding it a slightly shifted version of the same image. The dog would not activate the same neurons, so the network would react completely differently!

But CNNs only deal with translations. What if the image of the dog is rotated?

Re:Implementing a Convolutional Neural Network from Scratch ... by vzhou842 on Wednesday May 22, 2019 @03:00PM (#19984340)
True! If dealing with stuff like rotations is a concern, you could augment the training set by applying small random transforms to it (like rotations, cropping, scaling, color adjustment, etc).
Re:Implementing a Convolutional Neural Network from Scratch ... by amelius on Wednesday May 22, 2019 @03:27PM (#19984588)
Ok, this makes me wonder how humans do it. Would a person who has never seen a rotated upside-down dog recognize it?
Re:Implementing a Convolutional Neural Network from Scratch ... by greiskul on Wednesday May 22, 2019 @03:34PM (#19984656)
Well, human children do train their vision in a lot of different ways while playing, and probably very early develop a mechanism for dealing with all kinds rotated objects.

For some objects though, even adults are not instantly good, reading text upside down is something that is initially very hard, but with some practice it can be done, I've met some teachers that can do it from years of tutoring people from other sides of desks.

Re:Implementing a Convolutional Neural Network from Scratch ... by panic on Wednesday May 22, 2019 @07:00PM (#19986510)
The way something looks can definitely change based on its rotation -- see http://thatchereffect.com for example.
Re:Implementing a Convolutional Neural Network from Scratch ... by mrhwick on Wednesday May 22, 2019 @03:38PM (#19984691)
For that level of effectiveness you would want to investigate using a capsule network.