Heh, sure.
Promote from a function to a linear operator on the space of functions, . The action of this operator is just “multiply by ”. We’ll similarly define meaning to multiply by the first, second integral of , etc.
Observe:
Now we can calculate what we get when applying times. The calculation simplifies when we note that all terms are of the form . Result:
Now we apply the above operator to :
The sum terminates because a polynomial can only have finitely many derivatives.
Use integration by parts:
Then is another polynomial (of smaller degree), and is another “nice” function, so we recurse.
Other people have mentioned sites like Mechanical Turk. Just to add another thing in the same category, apparently now people will pay you for helping train language models:
https://www.dataannotation.tech/faq?
Haven’t tried it yet myself, but a roommate of mine has and he seems to have had a good experience. He’s mentioned that sometimes people find it hard to get assigned work by their algorithm, though. I did a quick search to see what their reputation was, and it seemed pretty okay:
Linkpost for: https://pbement.com/posts/threads.html
Today’s interesting number is 961.
Say you’re writing a CUDA program and you need to accomplish some task for every element of a long array. Well, the classical way to do this is to divide up the job amongst several different threads and let each thread do a part of the array. (We’ll ignore blocks for simplicity, maybe each block has its own array to work on or something.) The method here is as follows:
for (int i = threadIdx.x; i < array_len; i += 32) {
arr[i] = ...;
}
So the threads make the following pattern (if there are threads):
⬛🟫🟥🟧🟨🟩🟦🟪⬛🟫🟥🟧🟨🟩🟦🟪⬛🟫🟥🟧🟨🟩🟦🟪⬛🟫
This is for an array of length . We can see that the work is split as evenly as possible between the threads, except that threads 0 and 1 (black and brown) have to process the last two elements of the array while the rest of the threads have finished their work and remain idle. This is unavoidable because we can’t guarantee that the length of the array is a multiple of the number of threads. But this only happens at the tail end of the array, and for a large number of elements, the wasted effort becomes a very small fraction of the total. In any case, each thread will loop times, though it may be idle during the last loop while it waits for the other threads to catch up.
One may be able to spend many happy hours programming the GPU this way before running into a question: What if we want each thread to operate on a continguous area of memory? (In most cases, we don’t want this.) In the previous method (which is the canonical one), the parts of the array that each thread worked on were interleaved with each other. Now we run into a scenario where, for some reason, the threads must operate on continguous chunks. “No problem” you say, we simply need to break the array into chunks and give a chunk to each thread.
const int chunksz = (array_len + blockDim.x - 1)/blockDim.x;
for (int i = threadIdx.x*chunksz; i < (threadIdx.x + 1)*chunksz; i++) {
if (i < array_len) {
arr[i] = ...;
}
}
If we size the chunks at 3 items, that won’t be enough, so again we need items per chunk. Here is the result:
⬛⬛⬛⬛🟫🟫🟫🟫🟥🟥🟥🟥🟧🟧🟧🟧🟨🟨🟨🟨🟩🟩🟩🟩🟦🟦
Beautiful. Except you may have noticed something missing. There are no purple squares. Though thread 6 is a little lazy and doing 2 items instead of 4, thread 7 is doing absolutely nothing! It’s somehow managed to fall off the end of the array.
Unavoidably, some threads must be idle for loops. This is the conserved total amount of idleness. With the first method, the idleness is spread out across threads. Mathematically, the amount of idleness can be no greater than regardless of array length and thread number, and so each thread will be idle for at most 1 loop. But in the contiguous method, the idleness is concentrated in the last threads. There is nothing mathematically impossible about having as big as or bigger, and so it’s possible for an entire thread to remain unused. Multiple threads, even. Eg. take :
⬛⬛🟫🟫🟥🟥🟧🟧🟨
3 full threads are unused there! Practically, this shouldn’t actually be a problem, though. The number of serial loops is still the same, and the total number of idle loops is still the same. It’s just distributed differently. The reasons to prefer the interleaved method to the contiguous method would be related to memory coalescing or bank conflicts. The issue of unused threads would be unimportant.
We don’t always run into this effect. If is a multiple of , all threads are fully utilized. Also, we can guarantee that there are no unused threads for larger than a certain maximal value. Namely, take then and so the idleness is . But if is larger than this, then one can show that all threads must be used at least a little bit.
So, if we’re using CUDA threads, then when the array size is 961, the contiguous processing method will leave thread 31 idle. And 961 is the largest array size for which that is true.
So once that research is finished, assuming it is successful, you’d agree that many worlds would end up using fewer bits in that case? That seems like a reasonable position to me, then! (I find the partial-trace kinds of arguments that people make pretty convincing already, but it’s reasonable not to.)
MW theories have to specify when and how decoherence occurs. Decoherence isn’t simple.
They don’t actually. One could equally well say: “Fundamental theories of physics have to specify when and how increases in entropy occur. Thermal randomness isn’t simple.” This is wrong because once you’ve described the fundamental laws and they happen to be reversible, and also aren’t too simple, increasing entropy from a low entropy initial state is a natural consequence of those laws. Similarly, decoherence is a natural consequence of the laws of quantum mechanics (with a not-too-simple Hamiltonian) applied to a low entropy initial state.
Good post, and I basically agree with this. I do think it’s good to mostly focus on the experimental implications when talking about these things. When I say “many worlds”, what I primarily mean is that I predict that we should never observe a spontaneous collapse, even if we do crazy things like putting conscious observers into superposition, or putting large chunks of the gravitational field into superposition. So if we ever did observe such a spontaneous collapse, that would falsify many worlds.
Amount of calculation isn’t so much the concern here as the amount of bits used to implement that calculation. And there’s no law that forces the amount of bits encoding the computation to be equal. Copenhagen can just waste bits on computations that MWI doesn’t have to do.
In particular, I mentioned earlier that Copenhagen has to have rules for when measurements occur and what basis they occur in. How does MWI incur a similar cost? What does MWI have to compute that Copenhagen doesn’t that uses up the same number of bits of source code?
Like, yes, an expected-value-maximizing agent that has a utility function similar to ours might have to do some computations that involve identifying worlds, but the complexity of the utility function doesn’t count against the complexity of any particular theory. And an expected value maximizer is naturally going to try and identify its zone of influence, which is going to look like a particular subset of worlds in MWI. But this happens automatically exactly because the thing is an EV-maximizer, and not because the laws of physics incurred extra complexity in order to single out worlds.
Right, so we both agree that the randomness used to determine the result of a measurement in Copenhagen, and the information required to locate yourself in MWI is the same number of bits. But the argument for MWI was never that it had an advantage on this front, but rather that Copenhagen used up some extra bits in the machine that generates the output tape in order to implement the wavefunction collapse procedure. (Not to decide the outcome of the collapse, those random bits are already spoken for. Just the source code of the procedure that collapses the wavefunction and such.) Such code has to answer questions like: Under what circumstances does the wavefunction collapse? What determines the basis the measurement is made in? There needs to be code for actually projecting the wavefunction and then re-normalizing it. This extra complexity is what people mean when they say that collapse theories are less parsimonious/have extra assumptions.
Disagree.
If you’re talking about the code complexity of “interleaving”: If the Turing machine simulates quantum mechanics at all, it already has to “interleave” the representations of states for tiny things like a electrons being in a superposition of spin states or whatever. This must be done in order to agree with experimental results. And then at that point not having to put in extra rules to “collapse the wavefunction” makes things simpler.
If you’re talking about the complexity of locating yourself in the computation: Inferring which world you’re in is equally complex to inferring which way all the Copenhagen coin tosses came up. It’s the same number of bits. (In practice, we don’t have to identify our location down to a single world, just as we don’t care about the outcome of all the Copenhagen coin tosses.)
This notion of faith seems like an interesting idea, but I’m not 100% sure I understand it well enough to actually apply it in an example.
Suppose Descartes were to say: “Y’know, even if there were an evil Daemon fooling every one of my senses for every hour of the day, I can still know what specific illusions the Daemon is choosing to show me. And hey, actually, it sure does seem like there are some clear regularities and patterns in those illusions, so I can sometimes predict what the Daemon will show me next. So in that sense it doesn’t matter whether my predictions are about the physical laws of a material world, or just patterns in the thoughts of an evil being. My mental models seem to be useful either way.”
Is that what faith is?
If a rationalist hates the idea of heat death enough that they fool themselves into thinking that there must be some way that the increase in entropy can be reversed, is that an example of not seeing the world as it is? How does this flow from a lack of the first thing?
To be clear, I’m definitely pretty sympathetic to TurnTrout’s type error objection. (Namely: “If the agent gets a high reward for ingesting superdrug X, but did not ingest it during training, then we shouldn’t particularly expect the agent to want to ingest superdrug X during deployment, even if it realizes this would produce high reward.”) But just rereading what Zack has written, it seems quite different from what TurnTrout is saying and I still stand by my interpretation of it.
eg. Zack writes: “obviously the line itself does not somehow contain a representation of general squared-error-minimization”. So in this line fitting example, the loss function, i.e. “general squared-error-minimization” refers to the function , and not .
And when he asks why one would even want the neural network to represent the loss function, there’s a pretty obvious answer of “well, the loss function contains many examples of outcomes humans rated as good and bad and we figure it’s probably better if the model understands the difference between good and bad outcomes for this application.” But this answer only applies to the curried loss.
I wasn’t trying to sign up to defend everything Eliezer said in that paragraph, especially not the exact phrasing, so can’t reply to the rest of your comment which is pretty insightful.
It’s the same thing for piecewise-linear functions defined by multi-layer parameterized graphical function approximators: the model is the dataset. It’s just not meaningful to talk about what a loss function implies, independently of the training data. (Mean squared error of what? Negative log likelihood of what? Finish the sentence!)
This confusion about loss functions...
I don’t think this is a confusion, but rather a mere difference in terminology. Eliezer’s notion of “loss function” is equivalent to Zack’s notion of “loss function” curried with the training data. Thus, when Eliezer writes about the network modelling or not modelling the loss function, this would include modelling the process that generated the training data.
Could you give an example of knowledge and skills not being value neutral?
(No need to do so if you’re just talking about the value of information depending on the values one has, which is unsurprising. But it sounds like you might be making a more substantial point?)
Fair enough for the alignment comparison, I was just hoping you could maybe correct the quoted paragraph to say “performance on the hold-out data” or something similar.
(The reason to expect more spread would be that training performance can’t detect overfitting but performance on the hold-out data can. I’m guessing some of the nets trained in Miller et al did indeed overfit (specifically the ones with lower performance).)
More generally, John Miller and colleagues have found training performance is an excellent predictor of test performance, even when the test set looks fairly different from the training set, across a wide variety of tasks and architectures.
Seems like figure 1 from Miller et al is a plot of test performance vs. “out of distribution” test performance. One might expect plots of training performance vs. “out of distribution” test performance to have more spread.
In this context, we’re imitating some probability distribution, and the perturbation means we’re slightly adjusting the probabilities, making some of them higher and some of them lower. The adjustment is small in a multiplicative sense not an additive sense, hence the use of exponentials. Just as a silly example, maybe I’m training on MNIST digits, but I want the 2′s to make up 30% of the distribution rather than just 10%. The math described above would let me train a GAN that generates 2′s 30% of the time.
I’m not sure what is meant by “the difference from a gradient in SGD”, so I’d need more information to say whether it is different from a perturbation or not. But probably it’s different: perturbations in the above sense are perturbations in the probability distribution over the training data.
Linkpost for: https://pbement.com/posts/perturbation_theory.html
In quantum mechanics there is this idea of perturbation theory, where a Hamiltonian is perturbed by some change to become . As long as the perturbation is small, we can use the technique of perturbation theory to find out facts about the perturbed Hamiltonian, like what its eigenvalues should be.
An interesting question is if we can also do perturbation theory in machine learning. Suppose I am training a GAN, a diffuser, or some other machine learning technique that matches an empirical distribution. We’ll use a statistical physics setup to say that the empirical distribution is given by:
Note that we may or may not have an explicit formula for . The distribution of the perturbed Hamiltonian is given by:
The loss function of the network will look something like:
Where are the network’s parameters, and is the per-sample loss function which will depend on what kind of model we’re training. Now suppose we’d like to perturb the Hamiltonian. We’ll assume that we have an explicit formula for . Then the loss can be easily modified as follows:
If the perturbation is too large, then the exponential causes the loss to be dominated by a few outliers, which is bad. But if the perturbation isn’t too large, then we can perturb the empirical distribution by a small amount in a desired direction.
One other thing to consider is that the exponential will generally increase variance in the magnitude of the gradient. To partially deal with this, we can define an adjusted batch size as:
Then by varying the actual number of samples we put into a batch, we can try to maintain a more or less constant adjusted batch size. One way to do this is to define an error variable, err = 0. At each step, we add a constant B_avg to the error. Then we add samples to the batch until adding one more sample would cause the adjusted batch size to exceed err. Subtract the adjusted batch size from err, train on the batch, and repeat. The error carries over from one step to the next, and so the adjusted batch sizes should average to B_avg.
I don’t think we should consider the centroid important in describing the LLM’s “ontology”. In my view, the centroid just points in the direction of highest density of words in the LLM’s space of concepts. Let me explain:
The reason that embeddings are spread out is to allow the model to distinguish between words. So intuitively, tokens with largeish dot product between them correspond to similar words. Distinguishability of tokens is a limited resource, so the training process should generally result in a distribution of tokens that uses this resource in an efficient way to encode the information needed to predict text. Consider a language with 100 words for snow. Probably these all end up with similar token vectors, with large dot products between them. Exactly which word for snow someone writes is probably not too important for predicting text. So the training process makes those tokens relatively less distinguishable from each other. But the fact that there are a 100 tokens all pointing in a similar direction means that the centroid gets shifted in that direction.
Probably you can see where this is going now. The centroid gets shifted in directions where there are many tokens that the network considers to be all similar in meaning, directions where human language has allocated a lot of words, while the network considers the differences in shades of meaning between these words to be relatively minor.
Oh, very cool, thanks! Spoiler tag in markdown is: