Hacker News Reader

Machine learning is "learning from data." It is not the assumption that there are no dynamics, and that the future will simply be a repetition of the past. To the extent that the future is predictable, learning from data is the best that can be done.

The reality is that speech recognition, language translation, face recognition, object classification and detection, semantic segmentation, speech and image synthesis have improved by extraordinary leaps and bounds in recent years. If we used the logic that past failures inform a confident belief that future success in a challenge will inevitably fail, then we should've bet heavily against Alpha Go defeating one of the most accomplished human Go champions in the world. Self driving cars seem like a sci-fi fantasy until they become a mundane reality.

There's an irrational arrogance to human beings in general, and Wall Street types in particular, regarding the specialness / non-reproducibility of their intelligence. It's not unlike the belief that people had that organic molecules were somehow special, "vital," and not synthesizable from base elements.

Certainly, there's a long way to go to replicate the capabilities of a human brain, but I don't think we should exaggerate or fetishisize the human power to estimate and mitigate risk. We've seen many spectacular failures of that in recent years.

Would Alpha go beat the same human on 17x17 or 21x21? Or even 13x13? Let's not get started about self-driving cars…

Maybe not, because it has not been trained for those dimensions. Then again neither have humans (except maybe on 13x13). That could be interesting, but I'm not sure what we could learn from this.

Absolutely, both Fan Hui and Lee Sedol.

https://deepmind.com/research/alphago/

The link you provide does give no data on games played on 1717 or 2121. They were played on the standard 19*19 board.

The difference is that chess, go, etc are all essentially rules based. Finance has very few rules that do not break over time. Just look at QE. Arguably the financial market represents the collective intelligence of a huge amount of very clever people. Machines are only just starting to challenge a single human at a rules-based activity. We're very far from beating a brutally darwinian, impressively adaptive, human hive-mind whose main skill is figuring out when rules are about to get broken.

Interest rates cannot go negative was a commonly accepted rule at the beginning of my career (and interest rate option traders where using models that did not allow for negative interest rates).

QE?

Quantitative Easing

Here's my first result from Google:

https://en.m.wikipedia.org/wiki/Quantitative_easing

Google is getting really good.

What are the rules for image / video captioning? For natural sounding speech synthesis? For realistic image generation? For semantic segmentation? For determining perceptual visual similarity between images?

Frankly, rules based AI is basically a bust compared to learning from data approaches.

The hive was pretty amazing at destabilizing the entire global economy within a brief span following fundamental financial de-regulation.

People flatter themselves in ways which satisfies their egos and self-interest. Of course they want to believe in, and have everyone else credit, their magical superpowers. No government (backstop) wires necessary.

Automatic image and video captioning is in a very sorry state. Machines can't do it.

Machine learning is rule based. The difference is just that it creates many more rules for more and more granular cases. But it will never be intelligent, it will always be a dumb digital bureaucrat.

Given results from ex. neural networks, ML "rules" appear to be granular and flexible enough that I'm not sure they usefully count as being rules anymore.

There are rules in that you're able to create an exhaustive labeled training dataset. The rules may not be generic, but they exist through a specification by example.

Beating a human when there is no discoverability to the problem, and when building an exhaustive dataset is a problem in itself is not on the horizon yet.

> Automatic image and video captioning is in a very sorry state. Machines can't do it.

You may have missed various dramatic improvements over the past year:

https://scholar.google.com/scholar?as_ylo=2016&q=%22image+ca...

Definitely haven't. Where is captioning applied with human like accuracy?

Hmm. I suggest that recent SOTA results contradict your contention that the field is in "a very sorry state" and you up the standard to "human like accuracy". Feels like the goalposts are being moved.

Anyway, this paper demonstrates some impressive results:

https://arxiv.org/pdf/1611.06607v1.pdf

arguably the "crisis" was a perfectly rational response to inputs, namely, people no longer being able to service their debts.

But the idea that you can predict the future is absurd.

And even if you could, acting on that knowledge would change the future.

But the traders are trying to predict the future already! If it's impossible, then the machines are a shoe in, and if it is possible I'd still give the machines great odds.

The nuances of the game are different from strictly trying to predict the future.

Casinos don't try to predict the future of every spin of the wheel or roll of the dice. Every individual game is random. But over a large enough sample size (number of games played), the odds are not random.

The same principle is at play in trading the markets, except the "odds" are determined by the traders' skill - their "edge." Which is roughly some combination of access to the right people, access to better information (where better might but not necessarily mean quicker), and experience (gut feelings, hunches, what-have-you).

So the real question is, can machine learning algorithms develop an "edge", and if so, can they stay solvent long enough to do so.

No. In casinos the games are designed so that in the long run the house always wins. They don't need to care about odds.

The stock market isn't a game; it isn't designed and it continuously evolves.

> In casinos the games are designed so that in the long run the house always wins.

That's exactly what I meant by odds; they always win because the odds of the games are in their favor.

> The stock market isn't a game

It's a game in the same sense that life itself is a game.

> it isn't designed

The mechanisms that fit all the pieces together are very much designed by humans. True the behavior of the market isn't designed; it's an emergent property of the complex system we call "the markets."

> it continuously evolves

Yes, but change is also the constant. What underlies that change is supply and demand.

Of course you're free to disagree. But if you'd like to learn more about the not-100%-accurate analogy I described in abbreviated form, the source is here: http://www.goodreads.com/book/show/253516.Trading_in_the_Zon...

Edit: Good reading on "life as a game": http://www.goodreads.com/book/show/189989.Finite_and_Infinit...

>It's a game in the same sense that life itself is a game.

Life isn't a game.

There are varying levels of sophistication from which data can be learned from. Pigeons can learn non-trivial word concepts and statistics but they do so at a rate that is much slower than a human infant.

Thus far, machines have not done well in scenarios of low stationarity. Those are scenarios where the past is not so good a predictor of the future or where the data manifold is rapidly changing. This occurs for example, when the act of predicting changes what is being predicted. The related notion of Antipredictable Sequences is the most interesting consequence of the oft misconstrued notion of the No free lunch theorem.

Humans are far from the ideal machine for this scenario but they vastly outmatch current algorithms. This is because humans are much better in the kind of generalization that is based off theory building. While not perfect, it does well enough at generalizing from observations (think Newton deriving the laws of gravitation from observations and Brahe's data) to scenarios that are not so close to the data.

This means that for now and in complex environments, simpler models that are adaptable to change online are better. Additionally, their inspectability means they are more easily modified to better match any very changed dynamics.

So your examples of: `speech recognition, face recognition, object classification and detection, speech`

Are of the type where a learner, such as a deep network, whose generalization ability mostly hails from interpolating to fill in missing structure based on data (implying poor out of sample performance) does well when you can give lots of examples which provide good coverage.

Language translation is sort of inbetween (the space is not as smooth but is still fairly stationary at the scale of years) and so we also find that the advantages of deep learning have not been as pronounced and dramatic as has been the case for image and speech.

The digital machine advantage is two. Their memory capacity is outmatched. Second, and related to the first, energy needs are not so pressing a concern.

No one is suggesting that AI is a solved problem. I simply cited a handful of areas such as computer vision where progress in ML has been remarkably rapid after many years where the problems seemed daunting and skeptics had written off the prospects of the field.

Citing Newton as a typical example of the superiority of human inferential abilities perhaps represents a case of cherry picking. How often do most human beings come to exhibiting the level of insight and reasoning power required to construct the calculus and discover the laws of classical mechanics? Inventing supernatural explanations for natural phenomena, burning sacrifices / witches / books (on Evolution or other heresies) seem more par for the course than Newtonian level revelations.

Having said that, learning the laws of classical physics from observational data strikes me as a pretty natural task for the right kinds of machine architectures, since they represent essentially geometric symmetries which hold over a vast range of scales, space and time.

> Inventing supernatural explanations for natural phenomena

If our mathamechanical minions ever start musing about 'The Supreme Eigenvector' that would be the time to start worrying about machines taking over.

My point was that you've underestimated the problem difficulty, when the signal is not of a smooth manifold and where the dynamics are far from stationary, it is incorrect to infer that techniques which have worked well in stable and smooth scenarios will continue to work well. In fact they have not and it looks, not for a while yet either.

> Citing Newton as a typical example of the superiority of human inferential abilities perhaps represents a case of cherry picking

Are hyperparameter searches across a sea of AWS instances cherry picking?

Anyways, I don't think I cherry picked. There is not a single machine that can do the same yet. And more importantly, a learning algorithm that notices model failure and works out the surgery required to correct or even completely replace it.

Newton is an example of the human brain at its best. But he was far from the only: Maxwell, Green, Einstein, Noether, Archimedes and so on the list goes. But we don't need to go that far since Crows are already capable of generalization that out matches what machines can do for the near future.

I would add a third big advantage to the digital machine: the ability to seamlessly connect to "classical" algorithms. For example, AlphaGo also used a classic MonteCarlo algorithm.

A hypothetical "robot-cat-AI" could for example have access to a really precise physics engine, and also could run at a much higher "frame-rate", so it could see the world in slow-motion and execute really exquisite moves as a result.

How do you know that cats, flies, and other lifeforms run their brains at the same "frequency" as humans?

When it comes to comparing the sophistication, adaptivity and functions of life to the in-silico mashines we produce and use very, very primitive.

I bet Google "could" write some algorithms that can predict if a stock will go up or down seconds before change based on live search data. All it takes is for an article or something to come out, then watch people search for "Will IBM stock drop" - and perform live sentiment analysis across all such live queries involving the stock name.

In a sense, it's all about what information you have access to.

So the rumor is Goldman Sachs got out of the financial crash first because their systems saw the market move before anyone else. I don't know if that is apocrypha, but it is more to the point Google would not be predicting the market, but instead see market movements before anyone else.

I highly doubt that it is as technical as you seem to make it.

GS is very well connected, and they are constantly out there asking people for their opinions. Ten years ago they probably had some sort of database where people could share experiences, and that is most likely what led to them guessing the financial crash.

But it's not like they had some machine that one day told them to short mortgages.

I checked you bio are developing your own HFT strats? I ask because I am connected with a couple startup HFs. Ping me on Symphony {Brian Hewes} if your up for chatting about it.

The trouble with stock market systems is that any successful one is defeated by its own success - as its logic gets factored into everyone else's strategy.

This self-defeating problem is not present in the other AI applications you mentioned.

Any strategy a human can think of, they can also code as a program. The difficulty is in expressing one's strategy clearly.

Again, I'm not trying to provide an exhaustive survey of AI or ML applications. Merely point out that the fallacy of using past failures of AI and ML to conclude that they will continue to fail at tasks against which they have thus far made limited headway.

Bear in mind that in the 90's only a handful of supercomputers had teraflops of computing power, where now you can get an 11 TFLOPS Titan X Ultimate for $1200. Compute power continues to grow exponentially, yet it has only recently reached a level where certain kinds of approaches are truly practical. As Heinlein said, "When it's time to go railroading, people go railroading."

It's interesting that you should talk about antagonistic systems, since Actor-Critic Models, dueling architectures, Generative Adversarial Networks (GANs) are an extremely hot area of AI/ML research at the moment.

A question - with the answer left as an exercise for the reader:

What do speech recognition, language translation, face recognition, object classification/detection have in common that is not true about predicting the future price of a security?

> because traders are used to seeing such predictions fail.

There was a joke from the 80s: soon the whole trading floor will be replaced by a computer, a man and a dog. The man presses the button to turn on the computer every morning. The computer operates all of the transactions and settlements automatically. And the dog is there to bite the man if he touches any other button.

30 years later, still no dog on the floor!

But on the flip side, most physical trading floors are closed. Can't see the last few lasting much longer.

yeah, i really don't know what these guys are talking about. the business day to day has been massively revamped by automation. there are hft ops that are blasting millions and millions of orders a second and probably account for 70%+ of all trading volume.

I think you have cash equity in mind. This is a very small subset of what investment banks trade.

i trade financial futures and cash equity. i realize ibanks offer leveraged bets on all kinds of nonsense to provide a "service", but pretty much any strategy known to humanity can be achieved on an exchange with negligible counterparty risk

Being centrally cleared to reduce counterparty risk and being traded electronically on an exchange are two distinct things.

Electronic trading does change dramatically the shape and profile of a trading floor. Introducing central clearing with margining not so much.

You won't achieve electronic trading anytime soon on products where there are perhaps at most 200-300 buyers in the world, must of who are only interested in buying big chunks.

> Can't see the last few lasting much longer.

I believe the remaining ones are effectively TV studios for business channels, so they have a completely different purpose now.

There are several floors that have "real" action. A couple of the options pits that are truly specialist and the CBOE VIX pits for instance come to mind.

But I don't think anyone believes those have more than a couple of years left in them.

I think the joke was that the man is there to feed the dog, and the dog is there to keep the human from touching the computer.

This is a paraphrase of a semi-well-known quote from a respected business professor named Warren Bennis, about the "factory of the future" - who knows whether he coined it or adapted it from a common joke though. Here's his version of the quote, which I like a bit more:

The factory of the future will have only two employees, a man and a dog. The man will be there to feed the dog. The dog will be there to keep the man from touching the equipment.

The NYSE floor traders are only there for window dressing to be seen by the TV cameras. They are entirely redundant and watch movies for much of the day when there is no need to show a flurry of activity after opening or before closing. The entire building is no longer of any real importance.

Most of those floor traders is running a small business (2/3 guys/girls) based on past contacts and trader know-how (such as it is - some better, some worse.) It's really more of a 'bazaar' for small, likely well connected, trading firms rather than anything else. It has been 'hard' for them in recent years as order flow has moved to the largest, too-big-to-fail brokers (GS, JMP etc). But certainly the NYSE gets a lot out of the media coverage, but the brokers themselves are not owned by the NYSE.

> Also, high finance is not just about what you know. It's inevitably about who you know, about "illogical" factors such as salesperson charisma, entertainment, and most importantly, a credible personality type that understands the edge case risks

That explains the LIBOR scandal.

And virtually every scandal in human history.

We were talking about the finance industry, which some say that it should be this cold, rational thing where no "personal", charisma-based decisions are ever made, it's all based on offer and demand. Otherwise I do agree with you, our particularities as a species have caused lots of "scandals" in the past (like world-conquerors getting drunk and killing their best friends because of it), but that's to be expected, we're not robots.

> It's inevitably about who you know, about "illogical" factors such as salesperson charisma, entertainment, and most importantly, a credible personality type that understands the edge case risks.

And this is why AI/ML will never be very disruptive in law.

The low-dow is: It's not about the facts, it's a social status game.

The trouble with stock market systems is they are great at predicting the past and pretty much useless for predicting the future.

A fellow student at Caltech was developing a stock market AI in the 1970s. He was very secretive about it, and was sure it was going to make him rich. I sometimes wonder whatever happened to him.

One thing we fail to consider when talking about automation is that, at the higher end, tasks are complex and (man+machine) is much much better than machine alone. Further, the time-growth of (man+machine) will asymptotically dominate (machine). This will mean people with newer skills to master (man+machine) will be hired rather than jobs just going away. So predicting a particular job going away is a futile exercise at the higher end. The lower end is different, jobs there are simple enough to fully be automated away.

Though one of the new big players in FX doesn't have any traders [1].

[1] https://www.bloomberg.com/news/articles/2016-10-13/this-bank...

There's a duality of thought going on here (with apologies to the false dichotomy). One says that all intelligence is ultimately materialistic/mechanical and that all we must do is retrace these steps with electronics and we can achieve whatever we want including high-value traders. The other seems to leave room for... let's call it the "irrational" side of things.

My only point here is that just about all values themselves are irrational. The same problem in trying to get a computer to recognize that the Mona Lisa (or poched eggs or Mozart or whatever) have value, is the same one that makes it hard to teach them to figure out if Tesla is still a good long-term investment etc.