Sheep learn to recognise celebrity faces from different angles | Hacker News Reader

Sheep learn to recognise celebrity faces from different angles

newscientist.com

42 pointsrbanffy24 comments

I was worried this was another #metoo incident where sheep were identifying the celebrities who sexually assaulted them from a line-up.

Move over deep learning, hello sheep learning.

deepsheep learning (or dipshit learning)

SheepMind

I must say, these past couple days on HN have been exceedingly sheepish. Yesterday I believe there was a story about [Google] Sheep View.

Its just another field of neuro-science, but- so fluffy.

Did you hear about the new BaaS from AWS, Lamb-da? I hear bleat learning practitioners fleeced by Google Cloud are flocking to it

Word from the field this summer is the practitioners are advancing rapidly through sheer will. They literally shed the fluff and get down to business. With skin in the game, they jump hurdles in the field overnight!

Sheeple Magazine is a great training corpus for this.

How do they decide when the sheep gets a reward? To me it looked like it was quite ambiguous sometimes what the sheep was doing, it kind of waddled to the left/right and once it was at the right choice it got a reward... Anyone knows how the reward was determined? Was there a person observing?

It would be more convincing if the sheep had to commit to a choice by selecting a left/right door it needs to enter based on the information presented.

I find celebrity worshipping a bit... mêh.

(Sheep? Is this political science research?)

>For the first time, the study reveals that sheep can mentally take a two-dimensional image of a face and rejig it into a three-dimensional representation

Visual cortex can automatically adjust and learn some tasks, but some others require intentional mental effort, at least initially.

I haven't read the paper yet. But what I want to know how much intentional processing the sheep do to achieve this. As far as I know, best way to detect mental effort is the delay before action.

In humans mental rotation of picture is not automatic after the angle grows past some threshold. Then humans must intentionally rotate the image in their mind.

Huh, I'm more surprised that most other animals can't do this...

i don't think there is study saying other animals can't do this. There is also much to be said about successful construction of the study like the one in the article. As for example dogs studies show, dogs are more interested in other dog faces than human faces. And dogs are the animals who have the most contact with us. Even shelter dogs show less interest in human faces then the dogs who normally live with people. One can expect that other animals who have even less contact with people would have even less interest in human faces. They should have did this sheep experiment using photos of celebrity sheep - i think the results would be off the charts :)

Anyway, studies on dog recognition of human faces (and dogs are very good at it, recognizing emotions, etc..) show that dogs pay much attention to eyes. Another clue here is that deep learning networks trained for human and animal faces contain a lot of eye and snout kernels in the internal layers - it seems that image processing by our brain as well as by DNNs is mostly edge based - edge means high spatial frequency and higher frequency means more informative signal - and snouts and eyes are high density edge clusters compare to the rest of face. That suggests that eyes is a key element in face recognition. The sheep looking at left or right profile photo of a face still see one of the eyes which were shown on the straight mugshot photo of the same face, and thus i think the sheep has enough info to recognize the face. Somebody running a human face recognition DNN can perform the same test using the DNN instead of a sheep and see which neuron paths would get activated. My bet is on eye kernels.

The article says:

> Humans and monkeys are the only other creatures known to do this from two-dimensional images.

I have to assume they have at least tried it with other animals?

Publication bias could be a factor.

This open the door to a new age of solving crimes against sheeps. Show the survivor sheeps in an attacked herd a photo of each domestic dog living in the area and see what happens.

This seems a candidate for the ig Nobel prize..

Now if only sheep could recognize each other from different angles:

https://m.youtube.com/watch?v=91ClBEM6JAU

Finally we can replace those kludgy ANNs with these nice sheeps!

This is the very definition of animal cruelty!

You signed up to HN just to say that?

Slow night in St Petersburg.

Sheep-based facial login, anyone?

"Jack Gyllenhaal"

Reply on news.ycombinator.com