Old man 1: Life is one trouble after another. I’d be better off dead, better yet, I wish I was never born
Old man 2: True, true, but who has such luck ?.. maybe one in a thousand.
My blog: https://blog.cerebralab.com
I’m also building an open source generic ML library: https://github.com/mindsdb/mindsdb & https://github.com/mindsdb/lightwood …. which I guess might be of interest to some people here
I don’t think I ever said “running experiments”, I said looking at data was relevant (i.e. the data other people had collected about the movement of the planets, Earth’s moon and objects here on earth)
If I implied otherwise my bad, please point it out, I will correct it.
I believe the thing we differ on might just be a semantic, at least as far as redefinition goes. My final conclusion is around the fact that the term is bad because it’s ill-defined, but with a stronger definitions (or ideally multiple definitions for different cases) it would be useful, it would also, however, be very foreign to a lot of people.
Corrected the wording to be a bit “weaker” on that claim, but also, it’s just a starting point and the final definition I dispute against doesn’t rest on it.
1. The problem with theories along the vein of AIXI is that they assume exploration is simple (as it is, in RL), but exploration is very expensive IRL
So if you want to think based on that framework, well, then AGI is as far away as it takes to build a robust simulation of the world in which we want it to operate (very far away)
2. In the world of mortals, I would say AGI is basically already here, but it’s not obvious because it’s impact is not that great.
We have ML-based systems that could in theory do almost any job, the real problem lies in the fact that they are much more expensive than humans to “get right” and in some cases (e.g. self driving) there are regulatory hurdles to cross.
The main problem with a physical human-like platform running an AGI is not that designing the algorithms for it to perform useful tasks is hard, the problem is that designing a human like platform is impossible with current technology and the closest alternatives we’ve got are still more expensive to build and maintain than just hiring a human.
Hence why companies are buying checkout machines to replace employees rather than buying checkout robots.
3. If you’re referring to “superintelligence” style AGI, i.e. something that is much more intelligent than a human, I’d argue we can’t tell how far away this is or if it can even exists (i.e. I think it’s non obvious that the bottleneck at the moment is intelligence and not physical limitations, see 1 + corrupt incentives structures, aka why smart humans are still not always used to their full potential).
I was asking why because I wanted to understand what you mean by “decomposition”.
a system is a decomposer if it can take a thing and break it down into sub-things with a specific vision about how the sub-things recombin
Defines many things.
Usually the goal is feature extraction (think Bert) or reducing the size of a representation (think autoencoders or simpler , PCA)
You need to narrow down your definition, I think, to get a meaningful answers.
Why is the literature into reversible encoders/autoencoders/embedding generators not relevant for your specific usecase ?
Give an answer to that it might be easier to recommend stuff.
I don’t want to get into the whole CW thing around this topic, *but*:
1. Since you so off handedly decided not to use p values, why do you:
a) Use linear models for the analysis provided such low r2 scores
b) Why use r2 at all ? Does it seem meaningful for this case, otherwise, if your whole shtik is being intuitive why not use mae or even some pct based error ?
c) Are you overfitting those regression models instead of doing cross validation ?
d) If the answer to c is no, then: provide nr of folds and variation of the coefficients given the folds, this is an amazing messure to determine a confidence value regarding the coeficient associated not being spurious (i.e of the variation is 0.001-0.1 then that means said coeficient is just overfitting on noise).
f) If the answer is no, why ? I mean, cross validation is basically required for this kind of analysis, if you’re just overfitting your whole dataset that basically makes the rest of your analysis invalid, you’re just finding noise that can be approximated using a linear function summation.
Also, provided the small effect sizes you found, why consider the data relevant at all ?
If anything this analysis shows that all the metrics you care about depends mostly on some hidden variable neither you nor the pseudoscientists you are responding to have found.
Maybe missing something here though, it’s 3:30am here, so do let me know if I’m being uncharitable here or underspecfying some of my questions/contention-points.
The more specific case I was hinting at was figuring out the loss <--> gradient landscape relationship.
Which yes, a highschooler can do for a 5 cell network, but for any real network it seems like it’s fairly hard to say anything about it… I.e. I’ve read a few paper delving into the subject and they seem complex to me.
Maybe not PhD level ? I don’t know. But hard enough that most people usually choose to stick with a loss that makes sense for the task rather than optimize it such that the resulting gradient is “easy to solve” (aka yields faster training and/or converges on a “more” optimal solution).
But I’m not 100% sure I’m correct here and maybe learning the correct 5 primitives makes the whole thing seem like childplay… though based on people’s behavior around the subject I kinda doubt it.
TL;DR Please provide references in order for me to give a more cohesive reply, see papers bellow + my reasoning & explanation as to why you are basically wrong and/or confusing things that work in RL with things that work in SL and/or confusing techniques being used to train with scarce data for ones that would work even when the data is large enough that compute is a bottleneck (which is the case I’m arguing for, i.e. that compute should first be thrown at the most relevant data)
Maybe you do this, but me, and many people in ML, do our best to avoid ever doing that. Transfer learning powers the best and highest-performing models. Even in pure supervised learning, you train on the largest dataset possible, and then finetune. And that works much better than training on just the target task. You cannot throw a stick in ML today without observing this basic paradigm.
I would ask for a citation on that.
Never in any ML literature have I ever heard of people training models on datasets other than those they wanted to solve as a more efficient alternative to training on the dataset itself. Of course, provided more time once you converge on your data training on related data can be helpful, but my point is just that training on the actual data is the first approach one takes (obviously, depending on the size of the problem you might start with weight transfer directly)
People transfer weights all the time, but that’s because it shortens training time.
New examples of unrelated data (or less-related data) does not make a model converge faster on validation data assuming you could instead create a new example of problem-specific data.
In theory it could make the model generalize better, but when I say “in theory” I mean in layman’s terms since doing research on this topic is hard and there’s scarce little in supervised learning.
Most rigorous research on this topic seems to be in RL, e.g.: https://arxiv.org/pdf/1909.01331.pdf and it’s nowhere near clear cut.
Out of the research that seems to apply better to SL I find this theory/paper to be most rigorous and up to date: https://openreview.net/pdf?id=ryfMLoCqtQ … and the findings here as in literally any other paper by a respected team or university you will find on the subject can be boiled down to:
“Sometime it helps with generalization on the kind of data not present in the training set and sometime it just results in a shittier models and it depends a lot on the SNR of the data the model was trained on relative to the data you are training for now”
There are GAN papers, among others, which do pretty much this for inferring models & depth maps.
Again, links to papers please. My bet is that the GAN papers do this:
a) Because they lack 3d rendering of the objects they want to create.
b) Because they lack 3d renderings of most of the objects they want to create.
c) Because they are trying to showcase an approach that generalizes to different classes of data that aren’t available at training time (I.e. showing that a car 3d rendering model can generalize to do 3d renderings of glasses, not that it can perform better than one that’s been specifically trained to generate 3d renderings of glasses).
If one can achieve better results with unrelated data than with related data in similar compute time (i.e. up until either of the models has converged on a validation dataset/runs or in a predefined period of time), or even if one can achieve better results by training on unrelated data *first* and then on related data rather than vice versa… I will eat my metaphorical hat and retract this whole article. (Provided both models use appropriate regularization or that at least the relevant-data model uses it, otherwise I can see a hypothetical where a bit of high-noise data can serve as a form of regularization, but even this I would think to be highly unlikely)
No. You don’t do it ‘just’ to save computation. You do it because it learns superior representations and generalizes better on less data. That finetuning is a lot cheaper is merely convenient.
Again see my answers above and please provide relevant citations if you wish to claim the contrary, it seems to me that what you are saying here goes both against common sense. i.e. given a choice between problem-specific data and less-related data your claim is that at some point using less-related data is superior.
A charitable reading of this is that introducing noise in the training data helps generalize (see e.g. techniques involving introducing noise in the training data, l2 regularization and dropout), which seems kind of true but far from true on that many tasks and I invite you to experiment with it an realize it actually doesn’t really apply to everything nor are the effect sizes large unless you are specifically focusing on adversarial examples or datasets where the train set covers only a minute portion of potential data.
To address 2) specifically, I would say that philosophical “Rationalists” are a wider group but they would generally include the kind of philosophical views that most people on e.g. LW hold, or at least they include a pathway to reaching those view.
See the philsophers listed in the wikipedia article for example:
Pythagoras—foundation for mathematical inquiry into the world and mathematical formalism creating in general
Plato—foundation for “modern” reasoning and logic in general, with a lot of ***s
Aristotle -- (outdated) foundation for observing the world and creating theories and taxonomies. The fact that he’s mostly “wrong” about everything and the “wrongness” is obvious also gets you 1⁄2 of the way to understand Kuhn
René Descartes—“questioning” more fundamental assumptions that e.g. Socrates would have had problems seeing as assumptions. Also foundational for modern mathematics.
Baruch Spinoza—I don’t feel like I can summarize why reading “Spinoza” leads one to the LW-brand of rationalism. I think it boils down to this obsession with internal consistency and his obsession to burn any bridge for the sake of reaching a “correct” conclusion.
Gottfried Leibniz—I mean, personally, I hate this guys. But it seems to me that the interpretations of physics that I’ve seen around here, and also those that important people in the community (e.g. Eliezer and Scott) use are heavily influenced by this work. Also arguably one of the earliest people to build computers and think about them so there’s that.
Immanuel Kant—Arguably introduced the Game Theoretical view to the world. Also helped correcting/disproving a lot of biased reasoning in philosophy that leads to e.g. arguments for the existence of good based on linguistic quirks.
I think, at least in regards to philosophy until Kant, if one were to read philosophy following this exact chain of philosopher, they would basically have a very strong base from which to approach/develop rationalist thought as seemingly espoused by LW.
So in that sense, the term “Rationalist” seems well fitting if wanting to describe “The general philosophical direction” most people here are coming from.
But is there some functionality that this would provide that a wiki doesn’t ? (or some nice interface for that functionality that a wiki doesn’t).
Or is just the simplicity of installation and/or the simplicity of the data format ?
Do you think this is better than having e.g. a personal wiki ?
I mean, I basically agree with this criticism.
However, my problem isn’t that in the literal sense new theories don’t exist, my issue is that old theories are so calcified that one can’t really do without knowing them.
E.g. if I as a programmer said “Fuck this C nonsense, it’s useless in the modern world, maybe some hermits in an Intel lab need to know it, but I can do just fine by using PHP” then they can become Mark Zuckerberg. I don’t mean that in the “become rich as *** sense” but in the “become the technical lead of a team developing one of the most complex software products in the world” sense.
Or, if someone doesn’t say “fuck C” but says “C seems to complex, I’m going to start with something else” then they can do that and after 5 years of coding in high level languages they have acquired a set of skills that allowed them to dig back down and learn C very quickly.
And you can replace C with any “old” abstraction that people still consider to be useful and PHP with any new abstraction that makes things easier but is arguably more limited in various key areas (Also, I wouldn’t even claim PHP is easier than C, PHP is a horrible mess and C is beautiful by comparison, but I think the general consensus is against me here, so I’m giving it as an example).
In mathematics this does not seem to be an option, there’s no 2nd year psychology major that decided to take a very simple mathematical abstraction to it’s limits and became the technical leader of one of the most elite teams of mathematicians in the world. Even the mere idea of that happening seems silly.
I don’t know why that is, maybe it’s because, again, math is just harder and there’s not 3-month crash course that will basically give you mastery of a huge area of mathematics the same way a 3-month crash course in PHP will give you the tools needed to build proto-facebook (or any other piece of software that defines a communication and information interpretation & rendering protocol between multiple computers).
Mathematics doesn’t have useful abstractions that allow the user to be blind to the lower level abstractions, nonstandard analysis exists but good luck trying to learn it if you don’t know a more kosher version of analysis already, you can’t start at nonstandard analysis… or maybe you can ? But then that means this is a very under-exploited idea and it gets back to the point I was making.
I’m using programming as the bar here since it seems that, from the 40s onward, the requirements to be a good programmer has been severely lowered due to the new abstraction we introduce. In the 40s you had to be a genius to even understand the idea of computer. In modern times you can be a kinda smart but otherwise unimpressive person and create revolutionary software or write an amazing language of library. Somehow, even though the field got more complex, the entry cost went from 20+ years including the study of mathematics, electrical engineering and formal logic to a 3-month bootcamp or like… reading 3 books online. In mathematics it seems that the entry cost gets higher as time progresses and any attempts to lower that are just tiny corrections or simplifications of existing theory.
And lastly, I don’t know if there’s a process “harming” math’s complexity that could easily be stopped, but there are obvious processes harming programming’s complexity that seems, at least in principle, stopable. E.g. if you look at things like coroutines vs threads vs processes, which get thought as separate abstractions, yet are basically the same **** thing if you move to all but a few kernels that have some niche ideas about asyncio and memory sharing.
That is to say, I can see a language that says “Screw coroutines vs threads vs processes nonsense, we’ll try to auto-detect the best abstraction that the kernel+CPU combination you have supports for this, maybe with some input from the user, and go from there” (I think, at least in part, Go has tried this, but in a very bad fashion, and at least in principle you could write a JVM + JVM language that does this, but the current JVM languages and implementations wouldn’t allow for this).
But if that language never comes, and every single programmers learn to think in terms of those 3 different parallelism abstractions and their off-shots, then we’ve just added some arguably-pointless complexity, that makes sense for our day and age but could well become pointless in a better-designed future.
And at some point you’re bound to be stuck with things like that and increase the entry cost, though hopefully other abstractions are simplified to lower it and the equilibrium keeps staying at a pretty low number of hours.
Basically, the way I would explain it, you are right, using a bell curve and using various techniques to make your data fit it is stupid.
This derives from two reasons, one is am artifact, the fact that distributions were computation-simplyfing mechanisms in the past, even though this is no longer true. More on this here: https://www.lesswrong.com/posts/gea4TBueYq7ZqXyAk/named-distributions-as-artifacts
This is the same mistake, broadly speaking, as using something like pearsonr instead of an arbitrary estimator (or even better, 20 of them) and a k-fold-crossvalidation in order to determine “correlation” as a factor of the predictive power of the best models.
Second, and see an SSC post on this that does the subject better justice (completely missing the point), we love drawing metaphorical straight line, we believe and give social status to people that do this.
If you were to study intelligence with an endpoint/goal in mind, or with the goal of explaining the world, the standard dist would be useless. Except for one goal, that of making your “findings” seem appealing, of giving them extra generalizability/authorizativeness that they lack, normalizing tests and results to fit the bell curve does exactly that.
When you use them to mentally sort things for general knowledge of what’s out there and memory storage like in biology, if it works it works. Kingdoms seem to work for this.
Could you expand this a bit ?
3000 is a bit of an exaggeration, seeing as the vast majority of mathematics was invented from the 17th century onwards, it’s more fair to call it 400 years vs programming’s 70-something.
Though, if we consider analogue calculators, e.g. the on leibniz made, then you argue programming is about as old as modern math...but I think that’s cheating.
But, well, that’s kind of my point. It may be that 400 years calcifies a field, be that math or programming or anything else.
Now, the question remains as to whether this is good or not, intuitively it seems like something bad.
I kinda of agree with this approach, I actually propose it (thought a different program related to biology) in my previous article I posted here.
The reason I haven’t gotten into describing it that much is because it’s not like this is an area where I have a lot of power to influence stuff, my only goal here is to figure out why the failure modes happen to better avoid them myself.