Boffins harnessed the brain power of mice to build AI models that can't be fooled

In a bizarre experiment, researchers recorded the brain activity of mice staring at images and used the data to help make computer vision models more robust against adversarial attacks. Convolutional neural networks (CNNs) used for object recognition in images are all susceptible to adversarial examples. These inputs have been …

[Andy Non]

Douglas Adams was right all along then...

Mice are not, as is commonly assumed on Earth, small white squeaking animals who spend a lot of time being experimented on.

In fact, they are the protrusions into our dimension of hyper-intellegent pan-dimensional beings. These beings are in fact responsible for the creation of the Earth.

Two mice (Frankie and Benjy) escaped from Earth before the premature termination of its programme. They had belonged to an Earthling known as Trillian. They were rather keen to remove Arthur Dent's brain to reveal the ultimate question, which they had devoted a lot and time and money to finding.

The whole business with the cheese and the squeaking is just a front.

Douglas Adams

[A.P. Veening]

Re: Douglas Adams was right all along then...

Damn you for killing a potentially nice comment topic with an unbeatable first post ;)

[Morrie Wyatt]

Re: Douglas Adams was right all along then...

An unbeatable post?

Maybe not.

Not if you are pondering what I am pondering.

"One is a genius, the other's insane." NARF!

[TRT]

Re: Douglas Adams was right all along then...

An electronic brain. A simple one should suffice. Just program it to say "What?", "I don't understand " and "Where's the tea?"... Who'd notice the difference?

[Graybyrd]

Re: Douglas Adams was right all along then...

Which definitively explains why mouse brain pattern analysis at long last provides the final corrective element for American TSA terrorist screening software as used in airport security scans to, at long last, properly identify human female mammaries as breasts and not as explosives-packed 1965 VW Beetle boots. Critics had suggested that the elongated lifting handle at the boot's lower edge was a significant clue but TSA technical experts assumed the scans were of elderly subjects.

[Blackjack]

Re: Douglas Adams was right all along then...

Sadly, they forgot rhe rule about always carrying towels and were last seen "hidding" in some nasty alien stomach.

[steelpillow]

Cheese bias

So instead of racial, gender and ageist bias in our AI systems they are going to end up with cheese and cat-phobia biases.

Painting a triangular nose and whiskers on the road will cause a self-driving car to swerve to avoid, instead running down a pedestrian eating a cheese sandwich.

[Alister]

Re: Cheese bias

You're just pandering to outdated stereotypes. Modern, forward thinking mice have put all that cheese and cats business behind them long ago.

[Chris G]

Re: Cheese bias

Indeed, modern mice have digital showing a preference for vital cable insulation over cheese, up to nine times out of ten.

I am curious as to the effects of exposure to different categories of images to mice.

For example hours of mouse porn or millions of cat pictures, something like that could bring down the internet.

[TRT]

Re: Cheese bias

Cable insulation? Well a fear of Cat6 could be quite handy really.

[CrazyOldCatMan]

Re: Cheese bias

Modern, forward thinking mice have put all that cheese and cats business behind them long ago

The ones that survived anyway..

(Sounds of distant feline cheering when they realise that the mice won't be hiding any more)

[Inventor of the Marmite Laser]

+++++OUT OF CHEESE ERROR+++++

[Siberian Hamster]

Why on earth would anyone down vote your post? I really am intrigued if this was just an error or someone really found your post offensive..

[Evil Scot]

Cast a hex on them.

[Michael Wojcik]

What's the errno value? EDAM?

(If the mouse escapes errno is set to EEEK.)

[Il'Geller]

What does the brain and neuroscience have to do with AI? AI is based on language understanding. Indeed, why waste time and effort to understand the cause? When the consequences can be easily analyzed? Especially since over the last couple of thousand years all trying to understand how the brain works has not led to any success.

[A.P. Veening]

I'm more concerned over India-Pakistan, TBH.

That is just a very limited part of the field, useful for automated translation, but still very limited.

Indeed, why waste time and effort to understand the cause?

Ever heard of pure science? And of its stepdaughter serendipity?

[Il'Geller]

Money, only money! If science does not make money, it’s nothing: your pure science is a kind of Go or poker, suitable only for spending time. Brain analysis is for pastime.

[CrazyOldCatMan]

That is just a very limited part of the field, useful for automated translation, but still very limited

Especially as the mammalian brain will quite happily misclassify stuff that it sees as something else. Especially chair legs.

[Alister]

AI is based on language understanding

No. Machine Learning might be, but true AI has got to be based at least in part on neuroscience.

[Il'Geller]

1. If you are financed?

2. What for?

[TRT]

I'll be sure to tell the 842 colleagues I've worked with over the last 30 years that you think they've been unsuccessful and engaged merely in the indulgence of a 'pastime'.

Oh, I forgot to include over 1,000 guest speakers and visiting lecturers.

[Il'Geller]

Yes, it is. For the last 75 years only n-gram parsing technology has been used, which led to a purely mechanical parsing of texts, followed by a purely mechanical search for words (not information). Now it is possible to apply AI-parsing instead, which provides meaningful patterns and helps to find meaningful information.

"Meaningful" means the ability to represent a human's mental sphere externally, without invasive study of his brain mechanics. That is, there is a AI-distinction between medical aspects and the study of cognitive abilities. Mouse are not needed anymore, as well as any animals.

[TRT]

Computer modelling does *such* a brilliant job of predicting election outcomes, that I think you might be right*

*not.

I've never read such a load of old rot in my life, save from a certain 'Professor McNutty' who writes in daily to the local newspaper.

[amanfromMars 1]

Well I never ........ whoever would have a'thunk it.

So .... mice are brain leading humans into programming machines properly ‽ .

That surely sort of fcuks up Darwin's evolutionary theories right royally ...... to name but one global situation in need of re-evaluation?

The thing you have to both look out for and be prepared for is machines reprogramming humans ..... which you cannot deny is a prime current default in present states of existence.

[Julz]

Hum, “While our results indeed show the benefit of adopting more brain-like representation in visual processing, it is however unclear which aspects of neural representation make it work. We think that it is the most important question and we need to understand the principle behind it".

I think they perhaps should do some research to understand how what they are making actually works.

Mines the one with the solution to the qualification problem in the pocket...

[Pascal Monett]

"they were “head-fixed” and put on a treadmill"

So, Clockwork Orange-style methods are being used to train "AI".

That could never backfire on us, right ?

[Anonymous Coward]

Sounds good to me

Given some of the morons and PHB's I've had to deal with, I'm sure a few AI taught "brains" placed into British industry would be a huge improvement!

[Dadz]

Unscientific, immoral and macabre

Convolutional neural networks are primitive statistical models - the neural network of the mouse is a few orders of magnitude of greater complexity and there is no serious attempt to duplicate it or model it, only to use it as an input. Using the notorious ImageNet database of 5100 stolen images, they downsampled them to 64x36 pixel grayscale - then through a ~200 um window into the head, they used a TPSLM to observe the neuronal activity (light flashes) in GFP-M genetically modified neurons in the visual cortex - but they have no way to measure or model the connectivity of the neurons, each of which connects to thousands of other neurons. This is like attempting to determine which TV shows are popular based on the output of a webcam pointed at a block of apartment buildings, observing the timing of light switches in each apartment. The equipment they are using have been previously used to determine the structure of the mouse visual cortex, but they are not trying to replicate the structure of the mouse brain here, they are just doing black box engineering.

It is immoral to use animal brains for engineering purposes (I felt severe distaste with the K9000 cyberdog gun in the Fallout: New Vegas game - which has a dog brain to select targets) This study is completely depraved, careless, contemptuous of the value of life. These sadists propose to further expand their operation.

Apparently the animal ethics committee of Baylor has not the least regard for life and will rubber-stamp any animal study. Animal ethics committees are supposed to weigh each case and determine if the death (to whom?) of animals has a significant positive benefit to humanity - the vast majority of animal experiments at universities are unscientific as they do not make an attempt to ensure that the mechanisms studied are the same in people as in animals - extensive genetic testing and careful simulation is needed, and that is infeasible; when you have thousands of grad students all required to produce paperwork to earn grant money, you can't do things at a high standard. There are 100,000 cages containing 300,000 animals at Baylor.

[Anonymous Coward]

Re: Unscientific, immoral and macabre

You're talking shite...

[TRT]

Re: Unscientific, immoral and macabre

I'll agree with you about the some of the ethics of this, but the techniques used are fine. Well established now, perfectly able to determine cell connectivity and circuit formation albeit at an order removed from direct electrophysiological observation which is unable to non-disruptively record from the number of neurones required for complex circuit analysis, but which CAN and HAS validated the optical technique.

It's also highly unlikely that this set of protocols was solely used for this purpose. The principals of replace, reduce, refine and re-use will come into play. Re-using existing data and refining techniques to get more diverse science from a single set of experiments allows someone studying biological circuits using this method to collaborate with researchers from a completely different field, computation and AI.