[unimpressive]

> When has the exitement for discovery got so casual?

When we realized that the number of species on the planet is large enough to give up on any hope of ever cataloging them all.

I wonder if a machine learning algorithm could spot new species from flickr photos...

[masterzora]

> I wonder if a machine learning algorithm could spot new species from flickr photos...

You have to be careful with those things. My alma mater has a tale of a student 7 or 8 years ago who built a neural net to detect lizards in photos. He fed the system a bunch of pictures to teach it what lizards look like. When it came time to test he got remarkable results: it was 100% effective in positively identifying lizards. Then he fed it a picture of empty ground and the net happily confirmed that this picture contained a lizard as well! Turns out, if all of your training data is positive the computer just learns "everything is a lizard".

Of course, real-life researchers would never dare make such a mistake but this tale always amuses me so I couldn't resist :)

[relix]

So he really didn't understand neural networks?

[masterzora]

I wouldn't say that. I've always seen it like a physicist dropping a negative: he understands and when he sees the result he immediately knows what he did wrong but a little careless thought at the time led to a silly mistake.

[tripzilch]

David Icke would get better accuracy than that.

My university background is in Machine Learning. Sounds like this person did not understand ML. Was he new at this? You should learn about false positives, false negatives and plotting ROC curves from the first experiments with simple artificial random point cloud datasets (toy problems) before you even bother stepping up to image recognition.

When a ML researcher gets a 100% accuracy rate, they go look for the bug in their program, not think "oh awesome, it must be very good then"--it's a bit of an embarassing mistake if it turns out it's because your training set only contains examples for one of your two classes ("has a lizard" vs "does not have a lizard").

[masterzora]

> Was he new at this?

I mentioned he was a student for a reason. What do you think? This was a project for a class and even still he immediately recognised the error when he tested it.

Similar to as I mentioned in reply to a child comment, you kind of sound like someone chastising a physicist for dropping a negative somewhere in their calculations. It doesn't betray a fundamental lack of understanding such that they should be lectured to go back to first grade and learn how to do arithmetic; it just demonstrates that at some point they were acting somewhat careless in what they were doing.

And, yes, it is a bit of an embarrassing mistake. That's what makes it such a fun story!

[bagelicious]

I doubt machine vision is good enough to even classify known species.

[FaceKicker]

There is recent work along these lines for birds specifically, e.g., [1]. It's not anywhere near human accuracy and I think it's only a couple hundred species but it does kind of work.

[1] http://www1.icsi.berkeley.edu/~farrell/birdlets/iccv11-camer...

[colincsl]

I don't think vision/ML techniques will be a viable option for quite a while. These algorithms have a hard enough time correctly classifying dozens of classes, let alone 8.7 million species of animals. This is especially true given that many of these species only have a small number of images.