Joar Skalse

My name is pronounced "YOO-ar SKULL-se".

I'm a DPhil Scholar at the Future of Humanity Institute in Oxford.

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

This part of the argument is indeed quite general, but it’s not vacuous. For the argument to apply you need it to be the case that the function space is overparameterised, that the parameter-function map has low complexity relative to the functions in the function space, and that parameter-function map is biased in some direction. This will not be the case for all learning algorithms.

But, I do agree that the Levin bound argument doesn’t really capture the “mechanism” of what’s going on here. I can of course only speak for myself (and not the other people involved with this work), but I think of the Levin bound argument as essentially a more formal way to state this intuition. I.e., it is a loose argument for why we needn't be surprised that simple functions have larger measures in parameter-space, even if the argument doesn’t identify all the precise details of how this works in any particular case. For example, while the argument doesn’t apply to all ML algorithms, it does apply to all neural network architectures in exactly the same way, but clearly there are important differences in the inductive bias of eg CNNs and RNNs.

(Also: I don’t think the notion of “low effective parameterisation” really captures what’s going on here, but for the reason you point out yourself.)

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

Like Rohin, I'm not impressed with the information theoretic side of this work.

Specifically, I'm wary of the focus on measuring complexity for

functions between finite sets, such as binary functions.Mostly, we care about NN generalization on problems where the input space is continuous, generally R^n. The authors argue that the finite-set results are relevant to these problems, because one can always discretize R^n to get a finite set. I don't think this captures the kinds of function complexity we care about for NNs.

We’re not saying that discrete complexity measures fully capture what we care about for NNs! We do however think that they are sufficiently relevant to be informative for the bigger picture, even if just as a proxy for what we actually care about.

Most complexity measures give roughly similar values for the (relative) complexity of most objects, so our assumption is that if something is the case for a bunch of different tractable complexity measures, then this is also likely to be the case for whatever the “ideal” complexity measure would be in the relevant case. In particular, if regardless of whether K represents Boolean complexity, or LZ complexity, etc, then this is also likely to be true for the “ideal” complexity measure for neural networks.

Also: since we’re estimating various probabilities by sampling, we basically need to discretise the function space. If you have any concrete suggestions for how to get around this then we’re all ears!

As for the rest of your comment -- what you’re saying here seems true to me, but I’m not sure I see how any of this is a counterpoint to anything we’re saying?

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

Yes, it does of course apply in that sense.

I guess the question then basically is which level of abstraction we think would be the most informative or useful for understanding what's going on here. I mean, we could for example also choose to take into account the fact that any actual computer program runs on a physical computer, which is governed by the laws of electromagnetism (in which case the parameter-space might count as continuous again).

I'm not sure if accounting for the floating-point implementation is informative or not in this case.

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

There is a proof of it in "An Introduction to Kolmogorov Complexity and Its Applications" by Ming Li & Paul Vitanyi.

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

Ah, I certainly agree with this.

I do not wish to claim that *all* functions with low Kolmogorov complexity have large volumes in the parameter-space of a sufficiently large neural network. In fact, I can point to several concrete counterexamples to this claim. To give one example, the *identity function* certainly has a low Kolmogorov complexity, but it's very difficult for a (fully connected feed-forward) neural network to learn this function (if the input and output is represented in binary form by a bit string). If you try to learn this function by training on only odd numbers then the network will not robustly generalise to even numbers (or vice versa). Similarly, if you train using only number in a certain range then the network will not robustly generalise outside this range. This is because a pattern such as "the n'th input neuron is equal to the n'th output neuron" lacks a simple representation in a neural network (and hence this function has a small parameter-space volume, even though it has low Kolmogorov complexity). The same goes for the function that recognises palindromes, and etc.

So, I agree that there are certain functions with low Kolmogorov complexity that a neural network normally cannot "see" properly. I also think one could frame a lot of the research on developing new neural network architectures as being about making neural networks able to "see" more kinds of functions. For example, NALUs give neural networks the ability to "see" arithmetic relations more easily. I hence certainly think it's a very relevant question which complexity measure best describes the bias in neural networks (and I think this actually matters for practical problems). Note that the identity function is very smooth.

This is a bit of a tangent, but the Levin bound is actually about Kolmogorov complexity. It's a fairly simple theorem; the proof is constructive, and basically shows that a given function f which corresponds to many parameters in the parameter-space cannot be too complex, by constructing a simple program which computes f. Very roughly, if the parameter-space is finite and discrete, then we could construct a Huffman code for the function space (where the distribution over the function-space is the distribution that corresponds to the uniform distribution over the parameter-space). We can then make a computer program that computes f by concatenating the Huffman code of f and the parameter-function map m (which gives an upper bound on the Kolmogorov complexity of functions with large volumes). Of course, this theorem does not *per se* actually apply to neural networks, since it assumes that the parameter-space is finite and discrete, so in this context it's essentially just an intuition pump.

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

I'm not sure I agree with this -- I think Kolmogorov complexity is a relevant notion of complexity in this context.

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

Yes, I agree with this.

I should note that SGD definitely does make a contribution to the inductive bias, but this contribution does seem to be quite small compared to the contribution that comes from the implicit prior embedded in the parameter-function map. For example, if you look at Figure 6 in the *Towards Data Science* post, you can see that different versions of SGD give a slightly different inductive bias. It's also well-known that the inductive bias of neural networks is affected by things like how long you train for, and what step size and batch size you use, etc. However, these effects seem to be quite small compared to the effect that comes from the parameter-function map.

I should also note that I think that the fact that Gaussian processes even work at all already in itself gives us a fairly good reason to expect them to capture most of what makes NNs work in practice. For any given function approximator, if that function approximator is highly expressive then the "null hypothesis" should be that it basically won't generalise at all. The fact that NNs and GPs both work, and the fact that there is a fairly strong correspondence between them, means that it would be kind of weird if they worked for completely different reasons.

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

I believe Chris has now updated the *Towards Data Science* blog post to be more clear, based on the conversation you had, but I'll make some clarifications here as well, for the benefit of others:

The key claim, that , is indeed not (meant to be) dependent on any test data *per se*. The test data comes into the picture because we need a way to granuralise the space of possible functions if we want to compare these two quantities empirically. If we take "the space of functions" to be all the functions that a given neural network can express on the entire vector space in which it is defined, then there would be an uncountably infinite number of such functions, and any given function would never show up more than once in any kind of experiment we could do. We therefore need a way to lump the functions together into sensible buckets, and we decided to do that by looking at what output the function gives on a set of images not used in training. Stated differently, we look at the partial function that the network expresses on a particular subset of the input vector space, consisting in a bunch of points sampled from the underlying data distribution. So, basically, your optimistic guess is correct -- the test data is only used as a way to lump functions together into a finite number of sensible buckets (and the test labels are not used).

Why Neural Networks Generalise, and Why They Are (Kind of) Bayesian

We do have empirical data which shows that the neural network "prior" is biased towards low-complexity functions, and some arguments for why it would make sense to expect this to be the case -- see this new blog post, and my comment here. Essentially, low-complexity functions correspond to larger volumes in the parameter-space of neural networks. If functions with large volumes also have large basins of attraction, and if using SGD is roughly equivalent to going down a random basin (weighted by its size), then this would essentially explain why neural networks work.

I haven't seen the paper you link, so I can't comment on it specifically, but I want to note that the claim "SGD is roughly Bayesian" does not imply "Bayesian inference would give better generalisation than SGD". It can simultaneously be the case that the neural network "prior" is biased towards low-complexity functions, that SGD roughly follows the "prior", and that SGD provides some additional bias towards low-complexity functions (over and above what is provided by the "prior"). For example, if you look at Figure 6 in the post I link, you can see that different versions of SGD do provide a slightly different inductive bias. However, this effect seems to be quite small relative to what is provided by the "prior".

I have a few comments on this:

The main point, as I see it, is essentially that functions with good generalisation correspond to large volumes in parameter-space, and that SGD finds functions with a probability roughly proportional to their volume. Saying that SGD is “Bayesian” is one way of saying the latter, and the Kolmogorov complexity stuff is a way to formalise some intuitions around the former.

This has been done with real neural networks! See this, for example -- they use Gaussian Processes on stuff like Mobilenetv2, Densenet121, and Resnet50. It seems to work well.

We have done overtraining, which should allow SGD to penetrate into the region. This doesn’t seem to make much difference for the probabilities we get.

I basically agree with what you say here.