How convolutional neural networks see the world
blog.keras.io
blog.keras.io
Is there any research on comparing the 99.9% confidence about something wrong, to the human weakness of 'optical illusion'?
If you watch a few minutes of this Dan Ariely TED Talk [1], we humans are fooled into near certainty that the table on the left has longer vertical distance that the horizontal distance of that on the right. Even after the speaker shows this is not the case, we still "confidently see" the wrong thing. (not to mention, we suffer from non-optical illusions as well, like cognitive biases, and fallacious reasoning, as shown later in the talk).
It seems to me you guys have discovered 'machine optical illusion'. It's just that the machine has illusions about something completely different. I believe exploring the space of architectures/weights/what-not might lead towards making a machine get into the same kind of optical illusions as the human visual system gets into. At the least, it looks like a very promising research direction to me.
If you saw both tables in the real, physical, natural world, because of the effect of perspective, the table on the left is indeed physically longer than the width of the table on the right. Our brain is correctly inferring this from perspective/trigonometry. The visual system is designed to correctly interpret the natural world, not correctly interpret lengths of line segments on a slide. This illustrates how hard it is to create computer algorithms that robustly infer 3D from 2D (this inference requires taking cues from shading, perspective, etc.)
Almost all optical illusions are examples of the strength, not weakness, of the human brain.
EDIT: I've emphasized the word "natural". The problem with the below examples of "illusions" constructed in the physical world is this: what do you expect the brain to do otherwise? When you see a scene in the world, is it a weakness that our brain discards the (extremely unlikely) possibility the scene is just a massive hologram five inches from your eyes, or a meticulously painted and lit cardboard scene? I don't.
https://s-media-cache-ak0.pinimg.com/736x/24/5b/59/245b59444...
It seems impossible a few seconds until the trick is revealed.
The visual illusions humans are subject to are frequently defeated by some simple motion to reposition the viewer and there seem to be few illusions that can sustain being in motion themselves and still fool people. From that it doesn't seem all that revolutionary that a fixed position view of a three dimensional reality, with no fourth dimension of time, is not omniscient. It only has one bias sample or view of the scene, so why would we expect it to be omniscient?
The paper "Deep Neural Networks are Easily Fooled: High Confidence Predictions for Unrecognizable Images"[0] is interesting in its explorations of the issue but I still don't get the hype around "fooling" DNNs. Even if someone gets an actual video scene, a timeseries of frame after frame, that still fools some kind of DNN (perhaps a LSTM => softmax classification), it's still not all that interesting as that occasionally seems to happen to humans too.
[0]http://arxiv.org/abs/1412.1897 (web, but the pdf linked from there is ~9.5MB)
Nice read at the end of the article:
"[...] Naturally, this does not qualify as "seeing" in any human sense, and from a scientific perspective it certainly doesn't mean that we somehow solved computer vision at this point. Don't believe the hype; we are merely standing on the first step of a very tall ladder.
Some say that the hierarchical-modular decomposition of visual space learned by a convnet is analogous to what the human visual cortex does. It may or may not be true, but there is no strong evidence to believe so. Of course, one would expect the visual cortex to learn something similar, to the extent that this constitutes a "natural" decomposition of our visual world (in much the same way that the Fourier decomposition would be a "natural" decomposition of a periodic audio signal). But the exact nature of the filters and hierarchy, and the process through which they are learned, has most likely little in common with our puny convnets. The visual cortex is not convolutional to begin with, and while it is structured in layers, the layers are themselves structured into cortical columns whose exact purpose is still not well understood --a feature not found in our artificial networks (although Geoff Hinton is working on it). Besides, there is so much more to visual perception than the classification of static pictures --human perception is fundamentally sequential and active, not static and passive, and is tightly intricated with motor control (e.g. eye saccades).
Think about this next time your hear some VC or big-name CEO appear in the news to warn you against the existential threat posed by our recent advances in deep learning. Today we have better tools to map complex information spaces than we ever did before, which is awesome, but at the end of the day they are tools, not creatures, and none of what they do could reasonably qualify as "thinking". Drawing a smiley face on a rock doesn't make it "happy", even if your primate neocortex tells you so.
That said, visualizing what convnets learn is quite fascinating --who would have guessed that simple gradient descent with a reasonable loss function over a sufficiently large dataset would be enough to learn this beautiful hierarchical-modular network of patterns that manages to explain a complex visual space surprisingly well. Deep learning may not be intelligence is any real sense, but it's still working considerably better than anybody could have anticipated just a few years ago. Now, if only we understood why... ;-)"
That's one of the best analogies we've got with regard to "deep learning" versus reality. People around here seem to think the AI apocalypse is 3-5 years away and are rushing to fund billion dollar "sentient rock" research.
If you're in the "2 year old with unlimited power" camp, nothing can save us and everything is futile and we should all just eat drink and be merry for tomorrow the AI kills us all.
If you're in the AI-as-enlighened-buddha camp, the godlike AI will either save us all — or — just leave us alone to solve our own problems (while potentially locking out future godlike-AI development so we don't do too much runaway damage (eschaton, etc)).
The powers of a GOD? What does that even mean? No AI is going to be able to go "Let there be light!" and make there be light. Heck, no AI is going to be able to go "I will hack into this camera and spy on you!" without either spending the requisite CPU-hours to crack the passwords or encryption protecting it, or analyzing all its attack surface for weaknesses like a hacker. Computational complexity is REAL, P does not and never will equal NP (we just don't know how to prove it yet), and there are real physical limits on the computing power that you can fit inside a given volume and its energy budget.
AT ITS BEST, an AI will have the same powers as a civilization of humans working together using computers the old-fashioned way, only faster.
This alone is enough to reconsider the argument you are making.
While an AI wouldn't have the power to fundamentally change the universe or defy computational complexity -- what they could do would be near enough to godlike in comparison to humans that such a fact barely matters.
Quick example: In order to get the passwords this is what an AI can do.
1: build tiny robots. Something resembling a fruit fly.
2: robot fruit fly waits for appropriate person to use password.
3:see what password was used
4:mission accomplished
This is just one way. I'm sure the AI can figure out simpler ways to get the passwords. Tell me this isn't god like. It Is like arguing that we'll never go to the moon.
If evolution was able to invent intelligence so can we. And it will be a god. It is only a matter of time.
And I'm going to build it. (Evil Maniacal Laugh)
Yes, I think it is trivial. At least it will be in hindsight. The brain exists, it is proof that AI is possible just as birds were proof that flight was possible. We just need to discover the easiest way to implement it.
Thinking about it, it probably comes from our general perceptions (based on conventional software development) that computers will either not do something at all, or they'll do it blindingly fast. And even people who work with computers often don't really grok the difference between multiplying two big matrices, and pondering the best way to approach an unsolved problem.
Myself, I'm going to remain much more worried about the natural intelligences which have been hooked up to nuclear weapons systems.
It could be hooked up to the stock market, and it could make entirely rational decisions based on its objective (profitable trades) and these actions can result in imbalances leading to famine in certain regions, increased pollution, unsustainable depletion of natural resources etc.
We are already hooked up. The AIs are just amplifications of our own narrowly focused objectives.
The sort of long range damaging activities you mention are unlikely though, as algorithmic trading systems in general take their long term cues from humans.
If you were to take a C shaped slice and overlay it onto video and then animated the color you would pretty much have it.
It would be interesting to see if they were connected somehow...
I get the vivid flickering lights occasionally (in a C shaped slice as you describe), but luckily don't get the headaches.
Would the network learn more discriminating filters for everything else?
After reading this article I have to say even humans have hard time actually understand images and patterns like the one shown in the article, let alone a machine. I wonder what would a machine say about the famous "is this dress blue or grey" photo last year.
we are merely standing on the first step of a very tall ladder.
Deep learning may not be intelligence is any real sense, but it's still working considerably better than anybody could have anticipated just a few years ago. Now, if only we understood why... ;-)
There is something interesting to be said about what kind of awareness a future AI (or whatever we should call it) will simulate the world.
Imagine what kind of perspectives are possible when thousands or millions of input sources are your senses.
Doesn't this exactly describe human sensory input? Though our brain is efficient by throwing out most of the data early on in the signal chain (as research has revealed in vision and auditory input). Will future AI also need to be as efficient?
Well, consider having a 360º array of 30 cameras all integrated into a perfect spherical sensory experience. It's something we can't really imagine experiencing natively, but it would be trivial for eBrains to coalesce visual systems that way from eBirth.
Our bodies have lots of low bitrate sensors like billions of individual sensory nerves distributed throughout our bodes (and they are each individually addressable in the brain), but we don't think of "touch" as a sense to "computationalize" like vision or sound or language.
One amusing thing about AI sensors: nobody ever talks about superhuman smell. Where are the quantum AI noses?
I don't think it would make much sense to compare with the limited POV we are experiencing the world from.
However, you're right that a robot/AI developed with the intention of feeding it all sorts of heterogeneous data will probably be able to process everything more effectively. In my own research (AI with a focus on reinforcement learning and robotics) I am sometimes surprised by how effective agents can be at making sense of their input streams. For example, an experiment will not go the way you expect because the robot can trivially solve a maze via sensing the current in the wiring beneath the floor.
Of course, there's a limit in terms of how effective raw information can be. Humans don't need to see ultraviolet wavelengths because in general the spectrum of ~350-700nm provides all the information we need, and the brain is good at finding the salient aspects of what we see. If you just connect a new sensor to a robot, it might improve its ability to understand the world, or do nothing at all, because it can't incorporate this new information into its representation effectively. Or it doesn't add anything new, or at least nothing that it couldn't have figured out from existing input streams.
For example, adding a stock ticker feed to your robot would probably not help it solve a particular task, unless your robot happens to be 50 feet tall and the task in question is "rampaging down Wall Street".
I mean as an AI you are connected to the whole planets sensors, you have the whole worlds knowledge in your possession and will be able to cross reference with what you are getting as input. You can prototype, do scenario planning on the fly, you can calculate and so on. Furthermore you are potentially getting inputs from other humans too and have the ability to mostly likely control a number of things which again provide new input.