Deep learning is creating computer systems we don't fully understand
theverge.com
theverge.com
Last computer system I can somewhat fully understand is OCS Commodore Amiga. CPU, chipset and software.
Obviously no one has a full mental model in their head of everything going on in their phone (for example) at a given moment in time. But when it really matters, and there is a specific question to answer, and a programmer with experience writing the kind of software in question has full access to all of the source code, its possible to get an answer to why a program behaved in a particular way.
I'm a programmer, and this is what I do every day.
And in those rare cases where knowledge of the specific hardware is required, there are experts who understand the hardware and can answer the relevant questions.
So for machine learning approaches, what is the equivalent of the "source code"? What resource is there to examine to determine how the system will respond to a given set of inputs, and why? Evidently, there are not clear answers, at least for some machine learning approaches.
Phones have a bit fewer, but I think you'll find at least about 10.
A simple cheap SD memory card has an embedded core, usually ARM7TDMI. Even most lithium batteries have an integrated microcontroller.
I'm also a programmer, I do bare metal embedded, kernel drivers and such. I often don't have anyone else to ask from, except occasionally hardware guys. They're also often stumped.
So in a high profile case, like the Tesla Autopilot fatality, if the behavior was traced back to a specific micro-controller, it might be just as hard to determine why the system behaved as it did, as with a complex neural network? Even with a company's reputation at stake, it's not possible to hire a consultant with a deep understanding of obscure micro controllers?
I only deal with the source code level, don't really have experience digging into the hardware level.
http://www.edn.com/design/automotive/4423428/Toyota-s-killer...
Components have always a set of intended behavior and some, often unknown or poorly understood, set of unintended behavior.
It's very important to understand that abstractions mask both intended and unintended behavior.
You deal with unintended side the best you can, with sanity checks and so on. But some of it inevitably leaks through.
Those outcomes that result from multiple systems being individually in good state but collectively causing undesired behavior are often very tricky to understand. On the flip side, when you do understand them, you tend to gain new surprising insights into the whole system.
Also timing makes understanding hard. When it comes to a complex network of systems, it's hard to concurrently reproduce the state where the effect occurs.
> So in a high profile case, like the Tesla Autopilot fatality, if the behavior was traced back to a specific micro-controller, it might be just as hard to determine why the system behaved as it did, as with a complex neural network?
If Tesla has good, exhaustive, event logging facilities, a "black box", I guess they'll be able to figure out what happened by analyzing all of the available data with the systems knowledge they have.
Anyways, if it's not systematic, why should Tesla autopilot mistake be handled so differently from a human being making a mistake?
You can't analyze why a human being caused an accident either. The best you can usually do is to develop scenarios "theories" based on the available evidence, but that's it.
Way, way back, I had an 8 bit computer that I built myself from chips and solder. The "permanent" memory was 16kB; I knew most of that code pretty well. Even then the CPU itself required quite a bit of study to comprehend.
There are pretty significant limits to our conscious, logical understanding of the world.
In both cases we can only fix a particular instance of an error and effectively have no way to show that similar errors don't exist, or even that we didn't introduce new bugs with the fix.
My point was not something about manipulating weights directly vs. augmenting the training set. The analogy is roughly, find a misclassification : add particular input to training set :: find a bug : add a patch/test case. I'm playing devil's advocate, of course, and there are obvious, important differences.
> The difference is that all human written systems was once understood by a human
But that's not one of them! In practice, we create software systems by getting some rough intuition about an interface, wiring together some components, then checking the results empirically. Basic components that everything else is built on, C compilers for instance, are based on ambiguous, even contradictory specifications.
Unfortunately, adding that particular input to the training set may mess up the ANN when given a different input (i.e. you fix one bug but create one or more new bugs, like a game of whackamole but where people die)
Can we actually do this, for complex ANNs?
Additionally, the weights are constant in production, which makes analyzing production issues just a matter of logging the inputs and the outputs. Of course, for most convolutional nets, the inputs can get quite large (for example, a camera feed).
I think the key problem right now is that there are not too many models for which tracing the execution lets us know much about what's happening. A single decision tree. A linear model. But add ensembles to trees or layers to linear models and the ability to understand what's happening through logging disappears.
In this sense I think we're essentially in the pre-assembly language state of machine learning. For most programming languages, we assume correctness as a given. If it produces the wrong result on correct inputs, that's really bad.
With machine learning, we don't have perfect correctness. Aka accuracy. We focus so much on improving accuracy that we rarely consider the trade-off in debuggability or interpretability.
I think this will have to change. Advances in NLP and image stuff from deep learning are very exciting. Ignoring the evil AI concerns, I think the best step will be trying to understand better why the models did what they did. I think it will be a while before we can just ask the model and get a reliable answer. In the meantime, I think we'll work on a kind of meta programming where we integrate models better with their broader context. An example might be sequence to sequence models that have a way to expose sentence fragments and documents in the training set that influenced a word choice. Or automatic labeling of clusters in image representation layers that have semantic meaning to a human. There's some interesting work already on debugging tools for image classification that I really liked. On the phone so I'll try to provide the link later.
The trouble is it all seems very domain specific. Figuring out how do it in a way that's not so domain specific is i think a very interesting thing to work on.
We don't understand computer systems. Like you said, We won't even understand software.
What we understand is a small slice of it and we make mistakes even with that. Or fail to consider some cases or scenarios.
Then we put those small slices together to form a module and make the same mistakes again.
Take this python program:
print "Hello, world!"
Think of it this way: if it were impossible to understand opaque, highly emergent systems designed by nonhuman processes, we might as well shut down every life sciences lab and toss all the biology textbooks into the recycle bin. In reality, of course, we can figure out how biological systems work, if we care to put in enough effort, and artificial neural networks are not only orders of magnitude simpler, but have the enormous advantage that we can run them on digital computers, which at least gives us full access to all the raw code and data.
Now, a 1000 lines Python program is generally comprehensible with a little effort. The equivalent neural network is completely opaque.
But yeah, let's be pedantic.
Didn't they though? I'm pretty sure a couple of cameras for vision would mean they consume a couple of gigabytes rather quickly. That doesn't include sound, motor stress feedback and any other included sensors.
Asimov's robots would have consumed massively more data to learn. Of course, it wouldn't have required the data be curated and classified by humans.
I think we limit ourselves quite a bit in the mainstream software industry by insisting that everything be hand-crafted by humans and understandable. (Not to mention how short we fall of making software "understandable").
Another thing is that understanding is a key component of trust. Animals come to mind here. We can't understand their firmware per se; by observing them over thousands of years, both passively and actively (through domestication, training regimens, etc.), we've learned to expect some things from them but not others. So I can be sure that e.g. my cat won't kill me in my sleep. But I can absolutely not trust it to go somewhere when I want it to go there, or to stay in place when I want it to stay, or even to eat food when I give it. We want to write software that will bear much greater, more important responsibilities than any animal in history. So we need to make them more understandable than animals if we're to ever trust them with those responsibilities.
But when we push up against the limits of the problems we can solve with software, it's usually because it has outgrown our ability to understand it.
That's exactly what is happening with this new burst of progress in machine learning. We're "teaching" the computer to do things, and we've let go of the ability to directly understand how the computer "represents its learnings."
Was anyone else bothered by the phrase "show their working"?
That's the kind of thing that grates on me because it's just slightly wrong. When I grew up it was always "show your work", not "working".
Is this a regional thing?
However, I think it's also possible that we may never be able to create machines that are self aware as people are unless we do allow them to build without our understanding of them. There could be a recursive relationship in that an intelligence is not capable of understanding itself.
The more serious problem is that once we make life-changing choices using machine learning models, the models should be just and accountable. Obviously, these are vague terms, but in terms of machine learning models they can have concrete meanings:
* Just: choices should be made on what we consider (morally) relevant inputs.
* Accountable: choices should be traceable to inputs and parameters (or eventually training data).
E.g. suppose that we find drones that stun people with a weapon acceptable. Such a drone would be just if it decides to stun a person because they have a weapon. Such a drone would be unjust if it decides to stun a person because they have a particular skin color [1].
Now, suppose that such a drone stuns a person robbing a supermarket with a fake weapon. Regardless of whether we consider this to be just, we probably want to know why the drone misrecognized the gun as real (the model should be accountable).
[1] There may be a non-uniform prior distribution p(weapon_use|race). But I hope we all agree that using such prior distributions would be absolutely unfair towards the individual.
As someone who uses deep networks for language processing, I completely agree. IMHO this implies that we should be careful making life-changing decisions using such models.
Though, I am not sure whether legislation at this point is useful either. The lack of understanding of new technology in legislative branches typically leads to bad laws.
No, so its probably a bad idea to trust such a system to make important, consequential decisions.
Tagging faces in photos, recommending books or movies, probably fine.
Making an independent decision to fire a weapon, probably very bad idea.
Maybe the machine wasn't sure if most important thing in the question is "window" or "cover", so it looked at a bed looking for bed cover?
I have to agree that the EU, despite normally being a bunch of martinets, is likely getting it right here. It's too easy to construct and overfitted DNN and have it produce bizarre biases that affect patients, payrolls, or patents. There needs to be an appropriate level of transparency baked in as a precondition for licensable IP, IMHO.
That isn't necessarily a bad thing. You get more done with them than you would with out them.
Is working used as a noun here or is that a typo? Just never heard that use of the word working. Like ya, you can see I'm working vs you can see my working.
Simply ‘work’ reads much better—is it really frequently used as ‘working’ in mathematics‽
> Using a gerund as a noun seems very unconventional and incorrect almost by definition.
In English, gerunds are nouns (by definition). In "his singing was terrible", singing is a noun.
Furthermore, gerunds do often become used to refer to physical objects/results of actions: "writing", "reading" (e.g. readings for a class), "drawing", "booking" (of a ticket), "crossing" (on a road), "wiring", "etching", and others. Using "working" to refer to the written steps of working out a maths problem is just another example in that vein.
Your examples are good, though the only ones that are moderately common in American English seem to be “drawing” and “crossing”.
FTFY
"LIKE GOOD STUDENTS, COMPUTERS NEED TO BE ABLE TO SHOW THEY'RE WORKING"
heh.
What I would guess what they do is run deep learning algorithms on the data produced by the car's systems back at their HQ to design more efficient and complete systems, which they then update all of their cars to use.
An intelligent machine that is interpretable in a straightforward manner is an impossibility, it is a contradiction in terms.
You can teach the machine to answer questions about it's internal state in order to analyse it's function or you can apply your own mind to painstakingly deduce it's operation. But I contend you cannot have a generalised rubric for determining it's operation.
And when it comes to the bias in the data - bias in the system thing, well, same thing goes for human judgement. Another way of putting this would be garbage in-garbage out. Humans might be able to tell you of bias in the inputs, rather than just using it, but this ability rapidly deteriorates when the dimensionality of the inputs rises. You can tell if there's something weird about combinations of 3 numbers, but you can't tell if there is something off about 150 number sequences.
btw, I vouched for and upvoted the parent. It makes an excellent point and, frankly, I'm baffled by the downvotes/dead status.