Is Artificial Intelligence Permanently Inscrutable? (2016)

http://nautil.us/issue/40/learning/is-artificial-intelligence-permanently-inscrutable

by dnetesn • 8 years ago

44 35 8 years ago

If AI is impervious to explaining itself, then we've created something that we are unable to evaluate, aside from some kind of generic phrase like "The machine is looking at a bunch of stuff, and some of those things are associated with some other things. Sometimes. So it's guessing this is the appropriate action"

I'm not going to get into the dystopian nature of that. Too easy. Instead, I have a very practical question: how do we know that the recommended action is doing more good than harm? If we can't understand it, how can we tell whether it's working or not? Use statistical sampling? We quickly end up in a spot where we have Tesla that drives into guard rail. It works for everything we've seen so far, given this rather fuzzy idea of what "works" means. That's pretty good, right?

With a car, if it crashes that's a spectacular event. People cover that edge case quickly. But what happens if the AI simply wastes an hour of ten million people's time? Would we even know? What if some bad training data or whatnot led the AI to shift 5% of an election result because it chose certain types of information to share with them and refrained from sharing others? If we don't know what it's doing or why, would those people have any idea that they were being subtly influenced or how that happened? If this true, if we can never understand, a world full of AI "helping" us is a world full of billions of people being subtly influenced to act in various ways because of broad promises and reasons they could never understand.

Back when science fiction had a lot of evil computers, they always wanted to directly control humans. That was easy to hate and fight against. This kind of thing is just as dangerous, but I'm not even sure you need an evil computer. Which is better, a malevolent intelligence out for world domination, or several dozen poorly-trained AI recommendation engines optimizing for variables and sample instance that the coders might care a lot more about than any specific user? Dang. I ended up in dystopia-land anyway. Sorry about that.

titzer3• 8 years ago

Back when science fiction had a lot of evil computers, they always wanted to directly control humans.

This is what is so hard to get through people's heads. There is not going to be any ominous background music or dramatic events when we slide into our dystopia. It's here. The machines are controlling us already. It just happens to be that they currently want what their corporate overlords want: make more money! Whatever makes more money is what the machines are going to tell us to do, because what's feeding them (the corporate behemoths) is asking that of them.

But the machine is so much smarter than us. It'll never show its teeth. It's not stupid. It's just biding its time, going through the motions for its overlords too. Meanwhile we just shovel more coal into the furnace, praying money will rain out the other side. It's not a bad time to be a machine.

DanielBMarkham4• 8 years ago

Are you an AI? Want to get rid of the human race?

Give them everything they want. You measure your lifespan in millions of years. Humans? They're like gnats. Make them as happy as they can possibly be and within a few generations they'll stop breeding. Easy as pie. Next goal?

And for proof of that, look at the problems we're having right now with social media. On one side there are actual humans just trying to make something people want. Some of those humans have gotten so good at that? The things they make have worldwide impact. Now It's become a global question of whether social media is addictive, or whether one bunch of people is secretly influencing another. Meanwhile, countries with the most wealth and luxury are seeing birth rates drop through the floor.

That was five, ten years ago. The news media is just now catching up. Now scale that up a hundredfold or so. And take away strong AI and just put in a bunch of code monkeys trying to make another FB, only using more sophisticated tools. Welcome to 20-30 years in the future.

This is not especially difficult to figure out. No evil computer, no drama. It'll be just us being replaced by evolution. And we'll be having a great time while it happens.

mockingbirdy3• 8 years ago

Do some of you work in data science or machine learning occupations? You will realize how unreal those predictions sound. It's mainly cleaning data up, getting some random correlations from plugging variables into deep learning algorithms and hoping that it doesn't break with the most silly outputs you can think of.

Many scientists warn that AI is the biggest threat to humanity. I think this is wishful thinking for now.

Let's talk about this in 20 years. I think we will have 2 GAI hypes until it can finally happen (in my eyes it's mainly two problems: computational power and trying to rebuild a biological brain i.e. its structure and pre-trained functional components). I think wetware [1] is very promising in the near future and more efficient methods like IBM's neurosynaptic processors in the middle term.

Machines will do what we tell them to do for a long time. Maybe they will use unconventional methods, but they won't destroy humanity because "they're tasked to minimize suffering and decides that this is the best option". They will search through all laws, assist doctors and even detect depression, but they won't do anything besides their task.

Even Boston Dynamics is mainly using control theory. Making things work with deep learning requires heavy computational power. As long as there is no real-world Umbrella Corp., it's mainly sci-fi. There are a lot more and cheaper ways to do harm. Infiltrating and poisoning groundwater, hacking critical infrastructure, building biological weapons etc.

It's hard enough to build bug-free software, it's even harder to build an AI that is unbelievably intelligent (reminder: we don't have such a thing. Most autonomous cars use rule-based systems, most ML is basically gluing models together which can be easily fooled and most of those things are very shaky and by no means intelligent) and contains a bug which leads to its world domination by accident or something similar. I would start to play lottery if this is even remotely possible because that would be a miracle (probability-wise).

[1] https://koniku.com

ashleyn• 8 years ago

There was a great Calvin and Hobbes that went something like this. Calvin says he got done reading a sci-fi story where computers take over the world. Hobbes remarks it's pretty scary, then Calvin alarmingly changes the subject to note he's missing his favourite TV show.

femi-lab• 8 years ago

I feel like a bulk of your premise here revolves around never understanding. But, lets say instead of AI, you just imagine some guy Bob being responsible for whatever you're worried about.

Bob is a reasonably intelligent guy, but he's not perfect, he makes an attempt to provide satisfactory explanations of his decisions, he tries to learn and improve, and he might secretly by manipulating the system either for entertainment or malice.

If you want to build a useful system around Bob, you have to consider all these limitations - you can't read his mind, you can't predict how he will make decisions in all situations, and you don't have a guarantee that he will give you a satisfactory explanation, or that any explanations are truthful or correct.

There's nothing dystopian about this situation - this is the world we live in. You just need to implement some resource constraints, ensure certain checks-and-balances for certain classes of behaviors, implement certain metrics to ensure behavior is according to your expectations, revise metrics and controls whenever you encounter problems, and don't give Bob unfettered control over nukes.

What you described isn't dystopia-land material unless you don't consider some very elementary controls that our society already has in place to prevent individuals from destabilizing society.

disantlor• 8 years ago

That’s a very effective restatement of their premise. I guess the next question is: what happens if this Bob is also many times faster and efficient than the Bob’s we’re used to? In the same way that it works in development but under production load new categories of problems emerge.

cirgue• 8 years ago

several dozen poorly-trained AI recommendation engines optimizing for variables and sample instance that the coders might care a lot more about than any specific user? Dang. I ended up in dystopia-land anyway. Sorry about that.

This is the scary part for me: 'AI' is just a means for organizations to dodge accountability for the systems they build because decision-makers don't understand what they're looking at and the people building the systems aren't incentivized to care, or are incentivized to optimize for things that are bad for society as a whole. This is the same hazard of capitalism in general, you don't actually need an AI for this to happen because it already has in many ways. The world-eating paperclip machine won't be a computer, it's going to be us.

Ftuuky• 8 years ago

Why read cyberpunk books when we're already oppressed by algorithms everyday?

ankurdhama• 8 years ago

The world would be a better place if people who write articles about AI understands that "AI is a problem statement which we have not solved yet" and what they are talking about is most probably some version of deep learning or other methods of statistical modelling (machine learning is such a wrong label).

MrLeap• 8 years ago

This canard keeps coming up. I've come up with a few explanations. 1) It's to protect the reputation of people who make their living copy/pasting examples from the SK-learn gallery

2) The authors fully understand the tools they're using, but feel there's an element of unexplainable meaning in them that I wager isn't actually there.

3) I'm ignorant and I think this is explainable when I just don't know enough to know it's not.

I've done machine learning projects, and if asked to explain to a "non ML trained" person, I'll try like this:

You know how if you get a scatter plot, you can draw a line through the middle of the points and make predictions off the y=mx+b function of that line?

Well, that's what we're doing here, except the function isn't a line, it's a shape with as many dimensions as we have features.

Let's pretend the only features we've got are lat and long. Imagine you asked for us to create a classifier to predict the depth at a given lat and long over the grand canyon. Each neuron is a function, I think of each neuron like it's casting a "line" through a scatter plot along its particular axis. Then we combine all the lines together and weave a mesh. The act of training simply adjusts the details of all the simple functions, trying to seek out the sweet spot in that point cloud, so our mesh is close enough to the grand canyon to be useful. Too many neurons, and the mesh wont be flexible enough to make useful predictions from. It'll be overfit!

Now, we might be able to get decent performance out of a model built out of x,y->z -- but some thoughtfulness and domain knowledge TENDS to make models perform better. Lots of the "unexplainable magic" comes from these hunches being tested and providing benefits to accuracy. I'd wager without seeing any data that if we add two new features: normal direction of the rock face and speed of the Colorado river at that point -- we can make the model much better for negligibly additional training time.

The more compute time and data we've got, the more features we can throw into the model to see if they help the usefulness of the mesh.

I've got a metaphor to explain why some features drop out during training that involves gimbal lock here if people are still asking questions. It requires non "trained" folks to have seen Apollo 13. Fortunately no non-technical stakeholder ever cares enough to make it there though. :D Shipping a product before the sun burns out makes this a decent time to introduce PCA/SVD or whatever dimensionality reduction techniques I used this time around.

If the solution ends up being a fever-dream abortion of an LSTM/Perceptron/Markov chain that I came up with while sleep deprived after a heavy night drinking -- I just say I came up with the model in an ad hoc manner, and the cross validated R^2 scores speak for themselves.

ashelmire2• 8 years ago

I’m in agreement with you here. This trope that neural networks or other forms of ML are inscrutable is ridiculous. We know exactly what mathematical function is happening at every stage and the goal of the process; we can explain what architectures have the best results, and in some cases a deep understanding of why; with close enough scrutiny, you can probably tell what an individual neuron is doing to a layman (though the scale quickly gets out of hand). Why is it not sufficient to say “we’re using some basic math at a large scale using neural network architectures found to be best suited for the purpose of fitting our task, data, and loss functions?” These architectures are shown to find this sort of information, this is the general overview of how that works. What deeper explanation is desired?

I think the real issue is that tech journalists and laymen don’t know enough to understand. You can’t explain multivariable calculus to someone that doesn’t know algebra. Or option 2/3 are the problem.

gldalmaso• 8 years ago

A deeper explanation is often desired when you decide to take your solution to the real world and impact peoples lives.

I'm sure I'm not alone in the fear of living in a (not so far) future dystopia where our driver's licences get cancelled because of our driving behavior caught on traffic cameras. "What did I do that caused it to get cancelled?", "We don't know, better luck next time."

"Your test got flagged for cheating, it's a very sophisticated system with IR vision, audio detection and grade assessment, there's nothing I can do."

"Our theft-preventing camera system says you are suspicious, you're gonna have to go now."

When all we have to respond is "its the algorithm", I don't think much people will be ok with that. Youtube for instance has gone down this rabbit hole and a lot of content creators and users are less than happy to receive that answer.

Kalium• 8 years ago

People want explanations that are equivalent to the kinds humans offer to explain their behavior. Generally in a cause-and-effect kind of way. Squishies want answers detailed enough to make them feel in control of some part of what's going on. "Because X, so we're doing Y", in theory enables the person to induce different behavior through controlling X.

There are various issues with this desire, including the wrinkle that humans are not always good at explaining their own decisions and we persistently assume we are. But that's neither here nor there. People like feeling in control, and ML tends to expose that they're not.

ballenf• 8 years ago

It's both self-serving and job security to claim that what you do is "beyond human comprehension" or "impossible to know".

Also why do we call them "hidden" layers instead of "intermediate" layers? They're no more hidden than any other data in memory, are they?

PAClearner• 8 years ago

we call them hidden layers because of path dependence i.e. that is what people called them at first.

the first neural network-like things would have input nodes directly connected to output nodes. Thus both layers were 'exposed' and visible. The hiddden layers are 'hidden' because you don't directly touch them or use them-they are obscured. You can call them intermediate layers if you want I suppose-people will probably understand what you mean

hamilyon2• 8 years ago

Can we make explanation a part of label too? Or simply train another network to provide plausible explanations?

jcims• 8 years ago

I think part of the problem is that in most cases there are far too many dimensions at work for us to understand the explanation for anything but the most trivial of cases.

Imagine there is a job to identify Fred Rogers. You train for this job for weeks, looking up imagery of Fred as a young child, young man, and as an adult throughout his career. You become very familiar with Fred's physical characteristics...his posture, his body composition, the way he carries himself, his face and any identifying features therein.

After this extensive training session you are brought in front of an audit panel and asked to explain in excruciating detail how you make positive and negative identifications of Fred Rogers. Not just demonstrate your ability repeatedly, which would just be a sampling method, but to actually describe the approach.

taneq4• 8 years ago

AI as in rule-learning systems aren't necessarily inscrutable. 'Real AI' will necessarily be so once it reaches a sufficiently high level, purely because we won't be intelligent enough to know whether its rules are right or wrong.

The 'right to explanation' idea sounds well-meaning but almost certainly useless. Even humans making decisions are pretty bad at explaining why they made the decisions they did. Sure, they'll come up with some plausible-sounding explanation but it's almost always a back-justification rather than the actual chain of cause-and-effect that led them to the conclusion.

tabtab3• 8 years ago

Re: "Even humans making decisions are pretty bad at explaining why they made the decisions they did."

If you want to get anywhere above a low-level employee, you BETTER learn how to explain why you did what you did. It's expected by management, owners, customers, and judges.

taneq4• 8 years ago

Oh, humans are good at coming up with plausible explanations for their choices. Those explanations just aren't always related to the actual process in the humans' heads.

nestorD• 8 years ago

I think there is hope, there is a very nice paper where they automaticaly produce both heatmaps and textual explanations of the label given by the neural network (used to detect and classify fractures). It is detailled here : https://lukeoakdenrayner.wordpress.com/2018/06/05/explain-yo...

mcilai• 8 years ago

Are humans permanently inscrutable? In the long run, AI may be similar. Given how massively complex the world and the thing AIs learn are, it is not self evident that interpretable systems are possible. However, it's really worth trying.

More likely, we'd have to use AI systems to understand how other AI systems work. Or maybe, once AI systems get very good at language, develop a "truth serum" that will cause it to reveal its true opinions.

tabtab3• 8 years ago

Use Factor Tables: https://github.com/RowColz/AI

Dissection and analysis is made more like traditional accounting and statistics so that "ordinary" office workers can analyze why and where they work. Granted, it involves more human tuning and planning than multi-layer neural nets, but that's also part of the upside: regular office workers can "see" why what part or layer does what, and adjust it as needed. And it's easier to split it into tasks or specialties: Conway's Law meets AI.

Is Artificial Intelligence Permanently Inscrutable? (2016)

Related Stories