duvenaud

Can you explain a bit what neuroscience adds to the picture? To use your example, once we know that humans can take a difference perspective, what does it help to know which structure in particular is responsible for that ability?

They should have cited Robin Hanson's excellent proposed explanation of the Fermi hypothesis, which is that the universe is probably already about half full of civilizations, but that because they're expanding near the speed of light, we won't see them until they're almost here:

https://grabbyaliens.com/

I'm confused - it sounds like you're simply agreeing with Paul here, while claiming to disagree.

> If IQ doesn’t further our knowledge of human behavior nor the human mind it is at best a useless construct.

I mean, I think it does further our knowledge of these things, but even if it didn't, that wouldn't make it useless, because it still predicts things we care about pretty well.

> GDP has the benefit here of being extremely easy to measure... IQ doesn’t have this quality

Which measures of intelligence do you think are easier to measure than IQ? You mention the SAT positively below, which is very similar to an IQ test.

> IQ advocates often have a vested interest in an IQ that is immutable and (to some degree) heritable. This is often assumed to be true...

I mean, every researcher would say that it's only immutable and heritable to a degree, and there's tons of research quantifying the degree of mutability and heritability (e.g. twin studies, admixture studies).

> But it is still used to explain group differences. This is where using IQ becomes scientifically invalid.

Is there a specific claim you have in mind that you consider invalid? Can you give me an example of an explanation of group differences that you would consider scientifically valid?

> with immutability and heritability, you don’t have to explain it further. The answer is in psychology and genetics.

Are you saying these claims happen to be false, or that they're categorically impossible / bad explanations? Is there any aspect of human behavior that you think can be explained by psychology or genetics?

It sounds like you're just saying that IQ is overhyped. You agree that IQ can be a useful tool. And everyone agrees that having more data than just IQ lets you make better predictions.

How is IQ not "explaining behavior" a problem? We also wouldn't say that GDP "explains the economy", but that's not necessary for it to be useful. Both of these statistics are relatively strongly correlated with many outcomes we care about, compared to other things that are about as easy to measure. I think that's the main reason people use them, and it's a good one.

It sounds like you've seen people over-use aggregate statistics as a way to downplay nuance. That must be annoying. But that doesn't make their use philosophically or scientifically invalid.

Thanks for elaborating. But I still don't understand in what sense a skewed distribution "fits the underlying data". I presume that for any given task, we would have to write down a model or likelihood that relates intelligence to the outcome of the task. No matter what transformation of intelligence we use when discussing population statistics, as long as it's monotonic then we can always undo that transformation when we write down a model.

Again, regarding earthquakes: Does a logarithmic scale "fit the underlying data better"? I don't think the question even makes sense. It's like asking which font fits the underlying data better. It's an arbitrary choice that doesn't need to affect our predictions or models at all. But you also said that "the only reason to do that is to prove the people doing the former wrong." so maybe we agree on this, too.

I think you might be misunderstanding psychometrics if your objection to IQ is the supposed implication that "intelligence is this one thing that you can measure". As far as I understand, psychometricians would agree that intelligence has many facets, and that most of them are hard to measure. I think the claim that they're making is simply that one can summarize estimates of these facets into a single scalar that will be correlated with all of them. Just like GDP correlates with many hard-to-measure and complicated facets of a country's wealth. Would you say the fact that GDP can't measure some important things about a country's wealth, or that it's a complicated thing, implies that there's no such thing as "general wealth"? The same analogy works for e.g. health, strength, or pretty much any property that we have a word for.

Thanks for explaining. It sounds like we agree that "OP is wrong because normal distribution is not a sufficient indicator for some underlying phenomenon". And that we're not sure that there's a "true distribution" of intelligence. But I don't know how you can then argue that "intelligence does not fall on normal distribution but rather on a skewed distribution". Again, how could we choose between calling skewed thing vs its log transform "intelligence"?

I don't understand what one could mean by the true marginal distribution of a latent quantity that's only defined in terms of its joint distribution with measurements. What difference would it make if we transformed IQ scores to a skewed distribution?

To give an uncontroversial example, we use logarithmic scales for earthquakes precisely because the underlying phenomenon is heavy-tailed. It's just for convenience, it doesn't imply anything about the phenomenon being quantified.

I'm not sure I understand the objection. I agree that standardizing to a normal distribution might be arbitrary, but how could one argue what the true transform of the distribution is? After all, a log-normally distributed variable is just exp() of a normally distributed one. Would anything be different if instead of IQ, we called it log-IQ?

Thanks for actually answering my question! I guess no one wants to be a downer, but I'm pretty shocked that no one else in this thread, or that I've talked to about this issue, or anything I've read about fusion, has mentioned that current designs aren't on track to be remotely cost effective, which is what you seem to be saying.

Okay, but the other commenters are saying that fuel is around 6% of the cost of running a fission plant. Do you really think fuel availability will be a decisive factor in say, the next 60 years?

It sounds like you're pretty knowledgeable on this topic - care to speculate about the economic pros and cons of fusion vs fission?

Finally some numbers! Thank you. So again, this sounds like more evidence against an economic advantage of fusion over fission, is that correct in your opinion?

Thanks. So it sounds like there isn't an economic advantage of fusion over fission, except indirectly through less political opposition.

Thanks. So it sounds like you're saying that, unless there is some radically different plant design in the future, that there won't be an economic advantage to fusion over fission, at least in the near term.

Okay, but is the fuel a major cost for fission plants? And it's not clear to me how nuclear waste can both be extremely long-lived and also very dangerous that entire time.

Thanks, but is the fuel cost a significant factor in fission plants? And the waste disposal? There's already lots of radioactive material in the ground, after all, for example in uranium mines.

This is a bit off-topic, but can anyone here explain the expected economic advantages of fusion power over fission? My understanding is that fusion plants would also use heat-driven turbines, and be pretty capital intensive. Would they fill a different niche, or are they expected to be fundamentally more cost-effective?

Sorry, what's the proposed limitation? It's presumably not a rule against saying "fire" at all, since that what fire alarms are for.

I presume you mean that the limitations on speech take the form of punishments after the fact for saying "fire" when there wasn't one. I agree this is sensible.

But I think it's an important point that we can't (and shouldn't, and don't) ban people entirely from shouting "fire" in a crowded theatre, because if they're right, it's important to know.

I think it's worth discussing this more fleshed-out example, because it shows that in order to limit speech without shutting down useful information, you need to have some sort of trusted fact-finding apparatus to punish false alarms, which then become an important locus of power.

Theatres do indeed have an app to let someone yell "fire", which is the fire alarm.

It funny to me that the standard example of speech that shouldn't be protected is the only one that's required to be protected by building codes.

They were all unblinded weeks ago, the authors are listed on the workshop website: https://beyondbackprop.github.io/

Good question. I do think there are some bogus workshops, and at least a couple of the papers at this workshop seem to be written by people who don't understand the area very well. However, there's also a lot of great work in this workshop. It's also asking an important and fundamental question, which we have reason to believe has better answers than our current ones. I also think quantum ML is a perfectly sensible line of inquiry (although imo unlikely to pay off for a long time, if ever), even if its results are consistently overhyped in the media.

I don't really know what to do about the state of journalism, however. I thought I was being pretty hard on the author of this piece elsewhere in this thread, actually. I guess you could blame everyone who upvoted the article. On the other hand, we can't let the perfect be the enemy of the good.

There is a screening process for NIPS workshops (which I've been part of the last two years) where more experienced researchers rate the proposed workshops. There are usually a couple that get through that I think shouldn't be. However, I'd rather let 1000 stupid workshops through than shut down the one seemingly weird one that's actually a promising direction. Because if not here, where can we nurture and support weird ideas?

Finally, I think workshops play a vital role in onboarding newcomers to the field. One of the main ways you learn is by writing papers, and workshops lower the barrier to getting something out the door and getting feedback on it.

To be charitable, you could call it misleading laziness. But the article was published long after all the papers were de-anonymized, so it seems perverse to explain that the submissions were originally anonymous instead of just giving the authors' names.

Furthermore, the journalist managed to write an entire article about a workshop without once naming it, giving a link, or even defining backpropagation.

Edit: I want to say that I understand it's hard to cover an area that you're not an expert in. But if the journalist had googled the title of the articles he was writing about, he would have found the authors' names. Instead he gave the reader the impression that the articles are still anonymous.

Yes, good point!

The names of all the authors are clearly listed on the workshop website [1]. I think this is just a case of incompetence on the part of the journalist.

https://beyondbackprop.github.io/

Thanks for your question! But as I said, no one is really worried about the asymptotic time complexity of reverse mode differentiation, although there is scope for improving constants). The main scope for improvement is in the space complexity.

There is a lot of work on trying to speed up optimization, for example the K-FAC optimizer by Roger Grosse that uses second-order gradient information in a scalable way.

The lottery ticket pruning strategies do reduce space complexity, but I think the main reason people are interested in it is to reduce training time complexity, or deployment memory requirements, but not so much training memory requirements.

As for whether memory-saving and time-saving approaches are disjoint, many methods (like checkpointing) introduce a tradeoff between time and space complexity, so no.

Yes, "tuning for criticality" was the paper I was thinking of ! But I'm afraid I'm a dilettante when it comes to neuroscience. I basically just know the basic theories about consolidating learning during sleep.

This is basically what the chain rule / backprop does. The only caveat is that it's pretty hard to make a network with only positive weights. And once you allow negative weights, you have to know whether your downstream neighbours should have had higher or lower output. The easiest way to do this is to propagate the gradient backwards from the output, hence backprop.

Sure. First of all, I want to say that backprop, by which I mean reverse-mode differentiation for computing gradients, combined with gradient descent for updating parameters, is pretty great. In a sense it's the last thing we should be trying to replace, since pretty much the whole deep learning revolution was about replacing hand-designed functions with ones that can be optimized in this way.

Reverse-mode differentiation has about the same time cost as whatever function you're optimizing, no matter how many parameters you need gradients for. This which is about as good as one could hope for, and is what lets it scale to billions of parameters.

The main downside of reverse-mode differentiation (and one of the reasons it's biologically implausible) is that it requires storing all the intermediate numbers that were computed when evaluating the function on the forward pass. So its memory cost grows with the complexity of the function being optimized.

So the main practical problem with reverse-mode differentiation + gradient descent is the memory requirement, and much of the research presented in the workshop is about ways to get around this. A few of the major approaches are:

1) Only storing a subset of the forward activations, to get noisier gradients at less memory cost. This is what the "Randomized Automatic Differentiation" paper does. You can also save memory and get exact gradients if you re-construct the activations as you need them (called checkpointing), but this is slower.

2) Only training one layer at a time. This is what the "Layer-wise Learning" papers are doing. I suppose you could also say that this is what the "feedback alignment" papers are doing.

3) If the function being optimized is a fixed-point computation (such as an optimization), you can compute its gradient without needing to store any activations by using the implicit function theorem. This is what my talk was about.

4) Some other forms of sensitivity analysis (not exactly the same as computing gradients) can be done by just letting a dynamical system run for a little while. Barak Pearlmutter has some work on how he thinks this is what happens in slow-wave sleep to make our brains less prone to seizures when we're awake.

I'm missing a lot of relevant work, and again I don't even know all the work that was presented at this one workshop. But I hope this helps.

I'm one of the speakers at the workshop mentioned [1]. The article is a bit of a concept salad. I'm not familiar with all of the papers mentioned but am happy to try to answer questions.

[1] https://beyondbackprop.github.io/