[rvense]

Do you ever think that it's a fundamental limitation of these systems that they aren't good at knowing what they don't know? Like they always give an answer, and their failure modes are so different to ours that that it can be hard for non-experts to interpret the outputs.

In some of the less harmless applications of computer vision and machine learning, sometimes it will have very severe consequences for real people that a computer says yes or no when it really doesn't have the information to say either or. Some people are afraid of what will happen to society when these systems become as accurate as humans - I am honestly more worried about what will happen if they don't.

[lopmotr]

Who's going around designing a system that will have severe consequences without mitigating the problem of misclassification? There are techniques available, such as autoencoders, sensor fusion, ensembles, using multiple images, training on "neither" examples, asking for human confirmation, etc. It might never be perfect but neither are humans. We see monsters in the dark, the virgin Mary on toast, a face on Mars, optical illusions, get our attention distracted (magic tricks), act maliciously, fall asleep, etc.

The weird failure modes thing already happened with lossy image compression. Characters in non-OCRed text go replaced with different ones by photocopiers, and people saw spaceships in space probe photos of the sun. We'll get used to the odd banana riding a motorbike and realize what's up.

[candiodari]

How is the failure mode different from a human’s? The human mind comes with an answer no matter the situation it’s presented with, and that may be a stupid answer.

There is thankfully no (known) input that makes the human mind fail. There are known inputs for some animals though (like chickens).

[sbierwagen]

https://twitter.com/dmimno/status/949302857651671040?lang=en

>Optimist: AI has achieved human-level performance!

>Realist: “AI” is a collection of brittle hacks that, under very specific circumstances, mimic the surface appearance of intelligence.

>Pessimist: AI has achieved human-level performance.

[Cybiote]

Though humans do often confabulate, they can also say, "I don't know", "what is that", "wait a minute", "there's something fishy about this", "huh?". Sometimes they can invent a new label or phrase on the spot that captures component attributes.

[candiodari]

That would be a response, wouldn't it? If you give an AI "a way out", which is essentially the same, it will take a way out when it thinks it the smartest thing to do.

The trick is to lower the punishment for taking the way out. It's not free, but saying a car is a dog gets you -1, where as seeing a car and saying I don't know, only gets you a -.1 punishment (or even a .1 reward, vs a 1 reward for a correct answer).

[yawnxyz]

Or conservatives just feign knowledge/expertise. It would be funny if someone made a joke version that would confidently misclassify everything

[datameta]

I'm curious about the chickens. Is this maybe in the same genre of how a shark becomes entranced if flipped upside down?

[candiodari]

It's similar, yes.

https://www.youtube.com/watch?v=8Yo2UkL-n_Q

Pretty inconvenient, with chickens apparently this sometimes happens by accident. They don't get up, ever. They lie there until they get attacked or just die.

[datameta]

Fascinating... That must be a "drop dead" self-defense mechanism triggered by what looks to it like its beak drawing a line in the dirt due to it being dragged away by a predator.