At a session today on “Deep Learning” at the ASCUE17 conference Steve Kenode lauded the merits of advanced machine learning, like Deep Mind‘s “learning” the best way to play Breakout

I did appreciate the schematic way he explained it for a lay audience (including me) with the way weights are assigned to relationships; I’m a bit bothered by the terminology of “hidden algorithms” but that’s not my thing to take on today.

Steve raved about the ability of Google Photos to automatically access your photos and do “amazing things” to organize them (“you don’t need to write captions”).

I like writing captions on my photos.

That’s not the point either.

I was intrigued by the demo and the site for Clarifai a photo and video API that provides (tagline) “Artificial Intelligence with a Vision”

Don’t get me wrong, it it’s mind boggling to see how just an analysis (hidden) of an image can identity or suggest often very accurate descriptors and recognize faces.

Like pizza, when it’s good it’s really good.

And since they offer a demo, I gave it a spin to see if I could give it a good challenge.

So I first sent it a photo I did this winter of some of my toys playing on a table with some snow. Here is the original

I uploaded an image to the Clarifai demo from a copy I have in my random desktop images folder.

What can Clarifai clarify?

.

Here are the “keywords” it suggested, along with a “probability”

competition	0.999
race	0.996
snow	0.985
vehicle	0.982
athlete	0.979
winter	0.979
action	0.978
championship	0.964
adult	0.963
hurry	0.957
people	0.957
water sports	0.954
festival	0.951
motion	0.942
fun	0.934
wear	0.911
fast	0.905
sports equipment	0.901
veil	0.896
exhilaration 0.893

I have to say maybe snow, winter, people, or vehicle are fairly good guesses.

There’s a lot that are fails.

Here is another one, a very large anchor chain, maybe this was taken in the Baltimore Inner Harbor:

Linked


Linked flickr photo by cogdogblog shared under a Creative Commons (BY) license

Let’s give it to Clarifai

no person	0.983
one	0.942
people	0.931
flame	0.901
food	0.859
adult	0.818
still life	0.813
industry	0.803
indoors	0.786
wood	0.780
iron	0.760
color	0.754
art	0.750
desktop	0.735
technology	0.733
invertebrate	0.703
hot	0.696
recreation	0.695
energy	0.693
chain 0.687

High confidence are both no person and people? and WTF desktop, hot, invertebrate?

The lowest rated is chain. Hah.

Of course the explanation is that with iteration and/or correction, it will improve.

I don’t deny that.

It’s just that the same thing can be amazing when it works and ridiculous when it fails.

And as “artificial” the subject cares not.

The technology is neat, and there is a free level to play with the API. Someday I may have that need.

But damnit, I am writing my own captions!


Featured Image: Public domain image from pixabay

If this kind of stuff has value, please support me by tossing a one time PayPal kibble or monthly on Patreon
Become a patron at Patreon!
Profile Picture for CogDog The Blog
An early 90s builder of web stuff and blogging Alan Levine barks at CogDogBlog.com on web storytelling (#ds106 #4life), photography, bending WordPress, and serendipity in the infinite internet river. He thinks it's weird to write about himself in the third person. And he is 100% into the Fediverse (or tells himself so) Tooting as @cogdog@cosocial.ca

Leave a Reply

Your email address will not be published. Required fields are marked *