Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[tonnot98]

-snip, wrong thread-

[Starver]

(My homegrown NN experiments always have been so simple. That's the point, right? Once you start passing anything more than a simple value (and boolean is better, to be converted to a trained weighted value to be added/subtracted with all the other inputs to the next neural node) along a link, you're multiplying the necessary complexities of link indexing and thresholding beyond exponentiality. Which might give you a good pseudo-neural behaviour not available otherwise, but makes the undulating landscape of 'fitness' to training so much more expansive that you're going to have to run and exclude far more evolving configurations than may be practical, in your goal of finding a significant local maxima to the training data...)

(Look, it seems clear that I'm not at all up-to-date on the current state of the art NN research, obviously, but this is like Object Orientated programming all over again, in that I see why some people think it's the answer to everything, but my head worries about the overheads it inevitably produces. And if this lot have rediscovered an equivalent of a good, tight low-level functional (if pseudo-multi-threaded) programming scheme then I'm with them, but (if I've not got the wrong end of the wrong stick) it's not a new idea.)

[Reelya]

Oh, I wasn't serious about the hidden agenda. I just don't see what's exciting you, yet. Or the novelty aspect, if it's the weighted input thing, because that's par for the course in NNs, isn't it?

I thought the new part was the bit about "All areas of the neocortex assign their input to a 'location' in the external world." Which reminds me of a stateful machine, perhaps formed in an abstraction of logical feedback loops. Designing that in (with preconceptions) is trivial, when surely the Holy Grail is to find it cropping up as an emergent property, during training, upon deep analysis of a system not actually primed (perhaps prejudicially) with that as an intention.

Or I might just be reading it/you wrong. That's not unlikely.

Nah, expecting some "magic" emergent property from a single network in real-time is unrealistic. The real brain has specific regions which evolved over 100s of millions of years to do very specific tasks.  e.g. if the brain is damaged in specific regions, then you get oddly specific losses of faculties, like losing you sense of what "up" means. e.g. think about newborn foals. They're walking within minutes of being born. If they really had a "blank slate" NN-brain it wouldn't be possible for them to pretty much instantly be up and walking and finding just the right food source, their mother's udder. So while in one sense they're "learning" how to do those things, it's only because they're learning at the conscious level how to access pre-programmed facilities that are hard-wired at the sub-conscious level.

So, while there's local plasticity for mastering tasks, real brains have an overall plan/structure, so it's very unrealistic to expect that we can make a single undifferentiated network that can just bootstrap itself to meaningful human-scale intelligence, without someone (possible another machine) actually coding a precise plan for how that's going to work. That's sort of been the holy grail, but it's not very realistic, from what we know about how real brains are designed.

[bloop_bleep]

All the descriptions of neural networks I read online (and the limited implementations I created myself) have Boolean outputs, which are produced by weighting and summing all the inputs, feeding the sum through a non-linear sigmoid function (such as the logistic function), and then comparing the output of the function to a threshold. Not sure that is all that novel of an idea.....

[Reelya]

All the descriptions of neural networks I read online (and the limited implementations I created myself) have Boolean outputs, which are produced by weighting and summing all the inputs, feeding the sum through a non-linear sigmoid function (such as the logistic function), and then comparing the output of the function to a threshold. Not sure that is all that novel of an idea.....

Well they're trying to simulate how real brain neurons work. Clearly, real neurons don't work exactly like the reference artificial ones, but to explain the difference you do in fact need to know exactly how real brain neurons work, which is a non-trivial question to be posed as idle chat on a forum, given that you'd probably get a Nobel Prize if you knew that.

A big part of it would be that back-propagation is completely fabricated tech which has zero to do with how real neurons learn. It also doesn't cover creation of new synapses which is clearly very important. Biological neural networks don't have set in stone structures, which are then "trained", they have ever-changing structures, and the changes in structure themselves are part of the directed learning process. So what we need are neural networks with dynamic connections that can grow new cells or prune connections as needed, in real-time.

So sure, the artificial neurons they're working with might well sum inputs x weights, and have a threshold etc. But the difference is in how the connections themselves change in response to stimuli.

[Jopax]

Mind reading killer robots when?

It's a paid article so I only got the first bit, but from that, an AI is capable of somewhat reconstructing an image a person is seeing by reading stuff from an fMRI scan.

[Parsely]

That reminds me, my coworker claimed that the 7th Fleet collisions happening in the Pacific Ocean are caused by Chinese directed energy weapons that are either: altering current in electronic systems on American warships OR mind controlling the sailors.

[nenjin]

Shameless attempt to get at purchasing habits so they can advertise more bullshit to you. Who the fuck  honestly needs to be told their milk is a week old.

[scourge728]

Someone who can't read? IDEK

[Parsely]

nenjin's frothing rage isn't confined by mere thread boundaries.

[wierd]

In recent news, (to re-rail the topic, even if this is just run of the mill, and not "Amazing, Such tech! Much wow!" like usually gets posted here) Google has announced that they are going to enable Google Lens on all android devices running google photos, and also on iOS.

https://www.theverge.com/2018/3/6/17086688/google-lens-android-photos-launch-roll-out

I can hardly wait for the false positives about sheep in fields that are totally devoid of sheep.
https://slashdot.org/story/18/03/05/1720258/do-neural-nets-dream-of-electric-sheep

[PTTG??]

Frackin' toasters!

[Egan_BW]

"-raising concerns about self-driving cars"

[wierd]

Those crazy dutch would like to remind us that for a mere 400,000$, you too can own your own flying car.
https://techcrunch.com/2017/02/13/pal-v-begins-pre-sales-of-its-flying-car-starting-at-400000/

(Old article, posted because item shown on nightly news)

[Egan_BW]

Now, ask yourself when you'd ever actually need a car that can turn into a weak helicopter if you find enough space to take off.