PICS VIII: 10% LARGER THAN PICS VII

[Trivial]

It needs to be trained on something other than dogs.  It's all dogsdogsdogs.



Gif unrelated.

[Reginald Ret]

It needs to be trained on something other than dogs.  It's all dogsdogsdogs.

Maybe dogs were only a small part of its training but the way we have made dogs into all sizes and shapes has made the dog-pattern too broad, combine that with our ability to recognize human faces and the role of dogs in our history and society (dogs are people too!)

The most you could say is that it hasn't seen enough human faces.

[Rococo Modem Basilisk]

It needs to be trained on something other than dogs.  It's all dogsdogsdogs.

Maybe dogs were only a small part of its training but the way we have made dogs into all sizes and shapes has made the dog-pattern too broad, combine that with our ability to recognize human faces and the role of dogs in our history and society (dogs are people too!)

The most you could say is that it hasn't seen enough human faces.

I think the training set was made at the height of the doge meme. It's clear that the model has connected dog faces and cat faces with faces of other small mammals with wet noses, because occasionally the "dog" faces will look more like squirrels. Note that the model also frequently hallucinates birds.

The prevalence of the doge meme was a statistical sampling error; while the average animal face is cat-like, several individual pictures of shiba inu were copied millions of times, and should not have been counted.

[Mesozoic Mister Nigel]

It needs to be trained on something other than dogs.  It's all dogsdogsdogs.

Maybe dogs were only a small part of its training but the way we have made dogs into all sizes and shapes has made the dog-pattern too broad, combine that with our ability to recognize human faces and the role of dogs in our history and society (dogs are people too!)

The most you could say is that it hasn't seen enough human faces.

It's dogs and eyes. So many eyes.

This is interesting to me because human visual processing also devotes a disproportionate amount of real estate to eyes.

[Reginald Ret]

It needs to be trained on something other than dogs.  It's all dogsdogsdogs.

Maybe dogs were only a small part of its training but the way we have made dogs into all sizes and shapes has made the dog-pattern too broad, combine that with our ability to recognize human faces and the role of dogs in our history and society (dogs are people too!)

The most you could say is that it hasn't seen enough human faces.

It's dogs and eyes. So many eyes.

This is interesting to me because human visual processing also devotes a disproportionate amount of real estate to eyes.

I think that we see the eyes and not the other things because we are so focused on eyes.

[Mesozoic Mister Nigel]

It needs to be trained on something other than dogs.  It's all dogsdogsdogs.

Maybe dogs were only a small part of its training but the way we have made dogs into all sizes and shapes has made the dog-pattern too broad, combine that with our ability to recognize human faces and the role of dogs in our history and society (dogs are people too!)

The most you could say is that it hasn't seen enough human faces.

It's dogs and eyes. So many eyes.

This is interesting to me because human visual processing also devotes a disproportionate amount of real estate to eyes.

I think that we see the eyes and not the other things because we are so focused on eyes.

That starts to get into a chicken and egg argument; did we evolve to devote more cortical region to processing eyes and faces because we pay more attention to them, or do we pay more attention to eyes and faces because we have more cortical region dedicated to processing those features?

[Mesozoic Mister Nigel]

I guess I could unpack my original statement a little more: I think it's interesting because it is learning about what to pay attention to based on what visual features are most important to us, as gauged by images that we post on the internet. Therefore, simply by observing the images that we produce, it is doing a fairly good job of mimicking our cortical allocation. This is interesting to me partly because of a phenomenon which I am pretty sure has another name but which I have been calling "persistence" since I was a kid; the idea that that which is is that which is most likely to be. It isn't all that complicated, I'm just expressing it poorly.

[Reginald Ret]

I guess I could unpack my original statement a little more: I think it's interesting because it is learning about what to pay attention to based on what visual features are most important to us, as gauged by images that we post on the internet. Therefore, simply by observing the images that we produce, it is doing a fairly good job of mimicking our cortical allocation. This is interesting to me partly because of a phenomenon which I am pretty sure has another name but which I have been calling "persistence" since I was a kid; the idea that that which is is that which is most likely to be. It isn't all that complicated, I'm just expressing it poorly.

Ah, I understand now and I completely agree.
I was communicating at least as badly so thanks for the clarification.

[hooplala]

I guess I could unpack my original statement a little more: I think it's interesting because it is learning about what to pay attention to based on what visual features are most important to us, as gauged by images that we post on the internet. Therefore, simply by observing the images that we produce, it is doing a fairly good job of mimicking our cortical allocation. This is interesting to me partly because of a phenomenon which I am pretty sure has another name but which I have been calling "persistence" since I was a kid; the idea that that which is is that which is most likely to be. It isn't all that complicated, I'm just expressing it poorly.

That's really fascinating.  It makes me wonder what implications this might have for a theoretical AI mind.

[Doktor Howl]

[Rococo Modem Basilisk]

It needs to be trained on something other than dogs.  It's all dogsdogsdogs.

Maybe dogs were only a small part of its training but the way we have made dogs into all sizes and shapes has made the dog-pattern too broad, combine that with our ability to recognize human faces and the role of dogs in our history and society (dogs are people too!)

The most you could say is that it hasn't seen enough human faces.

It's dogs and eyes. So many eyes.

This is interesting to me because human visual processing also devotes a disproportionate amount of real estate to eyes.

I think that we see the eyes and not the other things because we are so focused on eyes.

As the people who wrote the thing in the first place noted in their paper, in addition to dog faces and eyes, the google imagenet model seems to have an obsession with tropical birds, pagodas, waterfalls, and gothic cathedral latticework. In fact, the architectural features are a little bit more prevalent; these images are all the result of a starting image that's just randomly colored pixels:




This started out as a picture of the sky:


But, people tend to take pictures of people and animals. And, correctly recognizing organic shapes associated with animals, deepdream proceeds to overfit the definition and make animal faces look more like its conception of an average animal face (which appears to be equal parts cat, dog, and human) and make animal orifices look like its conception of the average animal orifice (an eye).

[Cain]

I eagerly await the Lovecraftian horror that arises from when this program gets onto instagram and turns everything into someone's dinner.

[Doktor Howl]

I eagerly await the Lovecraftian horror that arises from when this program gets onto instagram and turns everything into someone's dinner.

Cram already did goatse, so it's over.

http://imgur.com/dVXFPUp

[Cain]

I almost clicked on that at work.

Then I remembered someone else would have to click on it to be able to pull me up on it, and I went ahead and clicked anyway

[Junkenstein]

I almost clicked on that at work.

Then I remembered someone else would have to click on it to be able to pull me up on it, and I went ahead and clicked anyway

Not only that, they'd have to try and describe what they were pulling you up for.