The limitations of Machine Learning/AI

As per usual, there has been much discussion about the role AI might begin playing in finance and white collar professions at large. I came across a fantastic article about the limitations of this subject that I think many of you might find informative.

Here's a somewhat long but interesting excerpt:

In short, deep learning models do not have any understanding of their input, at least not in any human sense. Our own understanding of images, sounds, and language, is grounded in our sensorimotor experience as humans--as embodied earthly creatures. Machine learning models have no access to such experiences and thus cannot "understand" their inputs in any human-relatable way. By annotating large numbers of training examples to feed into our models, we get them to learn a geometric transform that maps data to human concepts on this specific set of examples, but this mapping is just a simplistic sketch of the original model in our minds, the one developed from our experience as embodied agents--it is like a dim image in a mirror.

Current machine learning models: like a dim image in a mirror.

As a machine learning practitioner, always be mindful of this, and never fall into the trap of believing that neural networks understand the task they perform--they don't, at least not in a way that would make sense to us. They were trained on a different, far narrower task than the one we wanted to teach them: that of merely mapping training inputs to training targets, point by point. Show them anything that deviates from their training data, and they will break in the most absurd ways.

Comments (23)

Jul 21, 2017

I think the old "People fear what they don't know" applies here. But would you regard automatization in factories as the same thing as machine learning? or is automatization even more narrow.

Jul 21, 2017

People fearing what they don't know is 99% of it. Just think: people demonize "Wall Street" ostensibly because "bankers are thieves." Try explaining market efficiency, capital markets, or really anything beyond high school algebra outside a classroom to your average Joe.

Now you want to explain "machine learning" to Joe? You may as well try to explain The Matrix to someone from the 17th century. They'd first want to know why the machines don't use steam instead of humans.

Jul 21, 2017

As per usual is essentially the same thing as saying as as usual.

    • 1
Jul 21, 2017

lol fair enough

Jul 22, 2017

This is like a post in 1899 speaking about the limits of the motor car and how it's still very much a niche product with narrow applications.

Why is our understanding superior to the understanding a computer can attain in one specific subject? There's certainly already a great deal of things that computers can now do that no human or even a collection of humans could accomplish in a lifetime. The "dim image in the mirror" is pretty damn good compared to no image at all.

There will come a point when this changes...we have narrow AIs for now, true. At some point we will have AGI - general AI. It will have global understanding and be able to "understand" new topics on the basis of what it already knows. This may be 20 years off, it may be 50 years off. Really, it depends on what the semiconductor industry does in that period of time. Silicon is dying. We're reaching the end of its rapid expansion in capability. Once graphene becomes viable...watch out.

"When you stop striving for perfection, you might as well be dead."

    • 3
Learn More

Side-by-side comparison of top modeling training courses + exclusive discount through WSO here.

Best Response
Jul 21, 2017

Our understanding is superior because right now all we have are a few equations that we can stack on top of one another and translate to different applications to yield a targeted result. I don't think you're taking into account the amount of effort required to make even the simplest machine learning task a fruitful endeavor; your 1899 example would be more correct if you had said 1850.

Graphene is just one potential path that computing power might go down - I personally disagree that graphene will be it due to difficulties associated with charge retention and what that means for building an architecture on top of it. We're nearing the end of silicon dominance, but we're not nearly there yet. Perhaps by 2030 this starts becoming an issue at the industry level, perhaps by 2035 we get the first non-silicon processors available at massive scale, and even then it will take a long time to adapt programming and development processes enough to even begin utilizing the new technology reasonably well.

I see your point, and these predictions are as much bullshit as any, but I think OP's point is very much applicable to today. There's a lot of fuss over things that people don't understand, but when you step into industry you see just how limited these applications are. The counterpoint to that is that we're absolutely in the embryonic, perhaps even merely conceptual stage of big data + machine learning. There is a very long road ahead that encompasses many different moving parts of our world (processor tech, statistical learning techniques, programming development, big data pipelines, data transmission network capacity, +more), so who knows how soon or how far away a true AI is.

Long story short I agree with you and OP, just a lot of nuance involved with every perspective on this.

    • 6
Learn More

Side-by-side comparison of top modeling training courses + exclusive discount through WSO here.

Jul 24, 2017

ML used in conjunction with various other tools, both hardware and software, might result in the creation of a megacomputer. Think about it, average human tasks such as living like a factory worker in Ohio can be broken down into a specific sequence of tasks that can each be trained into the system.

Most secretarial work (arranging meetings, booking tickets, etc) can already be automated with modern tech to replace secretaries. Of course, extending it further to include non-mundane tasks such as analyzing a company through and through for PE would be quite impossible. But that does not stop us from using them to help us analyze insane amounts of data, fill in the gaps and form meaningful connections between them.

Intellectual work may require more work to be put in, and will require a rules-based system, to prevent overextension of machine knowledge - which can easily be subverted by a hacker/cracker. Which is essentially what happened in iRobot. But unless the machine has a means of obtaining new experiences, and is ingrained with a system to learn in a specific way from each experience for different types of problems, I don't think ML will become a problem.

GoldenCinderblock: "I keep spending all my money on exotic fish so my armor sucks. Is it possible to romance multiple females? I got with the blue chick so far but I am also interested in the electronic chick and the face mask chick."

    • 1
Jul 24, 2017
  1. Who cares if deep learning models do not understand what the inputs mean in real life, as long as they produce correct outputs? One of those outputs being human readable reports with quality as good as those written by senior analysts?(Kensho+) Many bankers dont have the slightest idea what theyre talking about and still they fake it well enough to earn a deal.

2."Show them anything that deviates from their training data, and they will break in the most absurd ways."

That is why you do cross validation and try to get models that generalize well. What you say only applies to bad ML models, not all ML models. Same thing can be said about bad bankers.

  1. Whatever limitation they might have, will not stop non-quantitative backgrounds from losing their job. (Unless youre already senior enough)
    • 1
    • 1
Jul 21, 2017

the whole point of me posting this was to sort of disprove #3 that you have listed. Contextual understanding (in the human sense) is still really important, even for non-quant positions.

    • 3
    • 1
Jul 24, 2017

If you work in an area where you use ML/DL approach to improve/validate an existing task that was previously done by experienced professionals who use "contextual understandings", you will soon find out how much bs exist there. Of course there are "a few" people that genuinely have those rare qualities, but most ppl who allege they have unique irreplacable skills actually do not possess them. I know I am not speaking kind words that will earn me bananas, but face the reality.

    • 1
Jul 24, 2017

This is a pretty good point, but I think most people who know deep learning/neural networks are well aware of this fact. An ML model is only as good as its training data. As the old saying goes, the only thing worse than no data is bad data.

Jul 25, 2017

This brilliantly written "Wait But Why" article articulates why AI, as apparently everyone here sees it, may not be as immediately (<50 yrs) plausible as most may think.

Neuralink and the Brain's Magical Future - Wait But Why

Jul 26, 2017

While I definitely agree with the op with regards to what we currently call AI and its limitations vs what's referred to as a "general artificial intelligence", I think stilarie is also right in that a lot of the jobs people do (especially in middle and back office) don't actually require much (if any) of that. Even more cynically, it's not clear to me that the recent advances in deep learning/ AI are even necessary to replace these jobs; you might be surprised by how much time people still take doing things that could easily be automated with a python script. In this scenario the AI stuff just serves as a buzzword to get institutional buy-in for otherwise obvious improvements and most of the savings actually come from setting up a system and processes that are tight enough that their outputs would be suitable input for any sort of algorithm.

I have a similar view on the applications of block chain in financial institutions. Unlike with crypto currencies and that company that's trying to use it to establish reliable and verifiable identities for people in unstable countries, it's not clear to me why a bunch of highly regulated multimillion dollar companies couldn't get together and designate somebody to hold a centralized database. So rather than the block chain being this great new thing that lets them do stuff they couldn't do before it's really just a fancy buzzword that gets people and companies excited enough to overcome their institutional inertia and actually do what was always obviously a good idea in the first place (it might also make these companies more attractive to young tech talent that would otherwise be drawn towards Silicon Valley).

Despite my lofty theses above, I acknowledge that I'm a bit of a neophyte on all this stuff so I'd be interested if somebody with more knowledge/experience could tell me why I'm wrong.

Jul 26, 2017

Read an article recently, where the thinking is that right now, a robot that could be introduced into our world will be one that does things like learning time of day we're most active and adjust our AC accordingly. Real simple, rudimentary, specific tasks these robots are set to do. World changing stuff here.

I know this makes some mad to think they're not getting their UBI anytime soon. It really puts a damper on the socialism believers I guess.

Oct 2, 2018

You are describing a simple forward feed neural net. Yes, it is nothing but a complex function that maps point to point. However there is progress being made in the AI field that goes well beyond this simple mapping function. Search for "openai beats dota 2 pros" for one example.

Array
Jul 21, 2017

I'm not saying that impressive things haven't been accomplished by neural networks. The point here is that they are still extremely limited. For game playing AIs based off neural networks, these guys heavily architect their structure, training process, etc. until they get it exactly right. The process takes years. And this is in a game where there are relatively fixed rules and there is relatively small amount of noise in the game. Whatever the randomness injected by the game and decisions by other players. But see enough games and you'll learn the structure of that noise. Not a big deal. The point of this post was not to say that neural networks are bad or dumb just that they are still heavily limited and there's a lot of misinformation and misconceptions out there.

    • 1
Oct 2, 2018
DeepLearning:

never fall into the trap of believing that neural networks understand the task they perform--they don't, at least not in a way that would make sense to us.

This really hits the nail on the head. Although there's wicked smart people programming these things, from my latest understanding they are now programming the things to program the things (at least one level removed on the complex ones), and don't really know why the things do what they do. I don't know people's understanding of AI here, but as far as getting the best algorithm it's just hyperspeed evolution.

It's like a gladiator pit of algorithms fighting each other, except the coaches of the gladiators are the ones fighting. They can pump out almost infinite gladiators and see how each does against what kind of opponents, then they discover which factors make big differences and pumps out new gladiators that are better in those categories. When you have real big money/processing, (google/deepmind), you have even more supervision/oversight- aka the fight is now between VPs or CEO levels, each a further level of abstraction so it's impossible to get that CEO to know why one single trader is doing so well.

Because there's no good way to ask them why they decided on certain attributes, we say the machines don't understand why they do what they do. The funny thing is humans can't show how we understand things either on a good level. Try playing the 'why' game with a three year old.

Human intelligence/learning is really just us updating our decision making algorithms(or principles as Ray Dalio would call them). We really are just complex forms of monkey see, monkey do. We start to combine these principles of things we've learned and it appears that we're creative, or original, but really we're just mashing up these principles and patting ourselves on the back with some evolution-provided dopamine.

"Show them anything that deviates from their training data, and they will break in the most absurd ways." If you think this doesn't apply equally well to humans, you have lived such a sheltered and boring life. The fact that ML algos break weirder is simply they've been trained in a more narrow way and are way more fragile (lack of good try/catch blocks basically).

    • 1
Oct 2, 2018

urmaaam: Your reply is the best post I've read. Well said! I could not say it better! When I discuss trading AI with someone, some people object opaqueness of AI because you can't see what's going on and extract the decision rules. People want to understand the rules and be in control and don't feel confident to let black box make decisions. I think that's an old way of thinking since in the past all software rules were written by people. Going forward however I see many AI black boxes in our lives. And human brain is the biggest and most complex AI black box of them all.

Array
Oct 4, 2018
urmaaam:

Because there's no good way to ask them why they decided on certain attributes, we say the machines don't understand why they do what they do.

This a billion times. I wouldn't misunderstand our own inability to understand the AI choices as an alleged limitation of AIs themselves. It's actually a rather dangerous misconception.

Jul 21, 2017

I would not say that this is a dangerous misconception. Model causality is a fundamental concept, one in which there is a pretty big missing gap for AI. For linear regression analysis, you have a causal understanding of the variables in your model. The only difference between linear regression and "machine learning" is non-linearity. That's it! Non-linearity assumptions can have very dangerous consequences as it can cause very unusual behavior. Think about an non linear function such as f(x) = x^2. f(x) increases very quickly as x increases. Not so much for a linear function f(x) = x. When you make assumptions about non-linearity you can accidentally impose some very crazy properties on your model so that if you get unexpected input you can get very unexpected output. Discounting interpretability and explainability as important is beyond insane.

    • 1
Oct 3, 2018