2 Answers sorted by
120
While at a recent CFAR workshop with Scott, Peter Schmidt-Nielsen and I wrote some code to run experiments of the form that Scott is talking about here. If anyone is interested, the code can be found here , though I'll also try to summarize our results below.
Our methodology was as follows:
1. Generate a real utility function by randomly initializing a feed-forward neural network with 3 hidden layers with 10 neurons each and tanh activations, then train it using 5000 steps of gradient descent with a learning rate of 0.1 on a set of 1023 uniformly sampled data points. The reason we pre-train the network on random data is that we found that randomly initialized neural networks tended to be very similar and very smooth such that it was very easy for the proxy network to learn them, whereas networks trained on random data were significantly more variable.
2. Generate a proxy utility function by training a randomly initialized neural network with the same architecture as the real network on 50 uniformly sampled points from the real utility using 1000 steps of gradient descent with a learning rate of 0.1.
3. Fix μ to be uniform sampling.
4. Let be uniform sampling followed by 50 steps of gradient descent on the proxy network with a learning rate of 0.1.
5. Sample 1000000 points from μ, then optimize those same points according to . Create buckets of radius 0.01 utilons for all proxy utility values and compute the real utility values for points in that bucket from the μ set and the set.
6. Repeat steps 1-5 10 times, then average the final real utility values per bucket and plot them. Furthermore, compute the "Goodhart error" as the real utility for the proxy utility points minus the real utility for the random points plotted against their proxy utility values.
The plot generated by this process is given below:
As can be seen from the plot, the Goodhart error is fairly consistently negative, implying that the gradient descent optimized points are performing worse on the real utility conditional on the proxy utility.
However, using an alternative , we were able to reverse the effect. That is, we ran the same experiment, but instead of optimizing the proxy utility to be close to the real utility on the sampled points, we optimized the gradient of the proxy utility to be close to the gradient of the real utility on the sampled points. This resulted in the following graph:
As can be seen from the plot, the Goodhart error flipped and became positive in this case, implying that the gradient optimization did significantly better than the point optimization.
Finally, we also did a couple of checks to ensure the correctness of our methodology.
First, one concern was that our method of bucketing could be biased. To determine the degree of "bucket error" we computed the average proxy utility for each bucket from the μ and datasets and took the difference. This should be identically zero, since the buckets are generated based on proxy utility, while any deviation from zero would imply a systematic bias in the buckets. We did find a significant bucket error for large bucket sizes, but for our final bucket size of 0.01, we found a bucket error in the range of 0 - 0.01, which should be negligible.
Second, another thing we did to check our methodology was to generate simply by sampling 100 random points, then selecting the one with the highest proxy utility value. This should give exactly the same results as μ, since bucketing conditions on the proxy utility value, and in fact that was what we got.
70
My answer will address the special case where is the true risk and the empirical risk of an ANN for some offline learning task. The question is more general, but I think that this special case is important and instructional.
The questions you ask seem to be closely related to statistical learning theory. Statistical learning theory studies questions such as, (a) "how many samples do I need in my training set, to be sure that the model my algorithm learned generalizes well outside the training set?" and (b) "is the given training set sufficiently representative to warrant generalization?" In the case of AI alignment, we are worried that, a model of human values trained on a limited data set will not generalize well outside this data set. Indeed, many classical failure modes that have been speculated about can be regarded as, human values coincide with a different function inside the training set (e.g. with human approval, or with the electronic reward signal fed into the AI) but not on the entire space.
The classical solutions to questions (a) and (b) for offline learning is Vapnik-Chervonenkis theory and its various offshoots. Specifically, VC dimension determines the answer to question (a) and Rademacher complexity determines the answer to question (b). The problem is, VC theory is known to be inadequate for deep learning. Specifically, the VC dimension of a typical ANN architecture is too large to explain the empirical sample complexity. Moreover, risk minimization for ANNs is known to be NP-complete (even learning the intersection of two half-spaces is already NP-complete), but gradient descent obviously does a "good enough" job in some sense, despite that most known guarantees for gradient descent only apply to convex cost functions.
Explaining these paradoxes is an active field of research, and the problem is not solved. That said, there have been some results in recent years that IMO provide some very important insights. One type of result is, algorithms different from gradient descent that provably learn ANNs under some assumptions. Goel and Klivans 2017 show how to learn an ANN with two non-linear layers. They bypass the no-go results by assuming either learning a sufficiently smooth function, or, learning a binary classification but with a sufficient margin between the positive examples and the negative examples. Zhang et al 2017 learn ANNs with any number of layers, assuming a margin and regularization. Another type of result is Allen-Zhu, Li and Liang 2018 (although it wasn't peer reviewed yet, AFAIK, and I certainly haven't verified all the proofs EDIT: it was published in NIPS 2019). They examine a rather realistic algorithm (ReLU response, stochastic gradient descent with something similar to drop-outs) but limited to three layers. Intriguingly, they prove that the algorithm successfully achieves improper learning of a different class of functions, which have the form of somewhat smaller ANNs with smooth response. The "smoothness" serves the same role as in the previous results, to avoid the NP-completeness results. The "improperness" solves the apparent paradox of high VC dimension: it is indeed infeasible to learn a function that looks like the actual neural network, instead the neural network is just a means to learn a somewhat simpler function.
I am especially excited about the improper learning result. Assuming this result can be generalized to any number of layers (and I don't see any theoretical reason why it can't), the next question is understanding the expressiveness of the class that is learned. That is, it is well known that ANNs can approximate any Boolean circuit, but the "smooth" ANNs that are actually learned are somewhat weaker. Understanding this class of functions better might provide deep insights regarding the power and limitation of deep learning.
Human approval is a good proxy for human value when sampling (even large numbers of) inputs/plans, but a bad proxy for human value when choosing inputs/plans that were optimized via local search. Local search will find ways to hack the human approval while having little effect on the true value.
I do have this intuition as well, but it's because I expect local search to be way more powerful than random search. I'm not sure how large you were thinking with "large". Either way though, I expect that sampling will result in a just-barely-approved of plan, whereas local search will result in a very-high-approval plan -- which basically means that the local search method was a more powerful optimizer. (As intuition for this, note that since image classifiers have millions of parameters, the success of image classifiers suggests that gradient descent is capable of somewhere between thousands and trillions of bits of optimization. The corresponding random search would be astronomically huge.)
Another way of stating this: I'd be worried about procedure B because it seems like if it is behaving adversarially, then it can craft a plan that gets lots of human approval despite being bad, whereas procedure A can't do that.
However, if procedure B actually takes a lot of samples before finding a design that humans approve, then we'd get a barely-approved of plan, and then I feel about the same amount of worry with either procedure.
How do you feel about a modification of procedure A, where the sampled plans are evaluated by an ML model of human approval, and only if they reach an acceptably high threshold are they sent to a human for final approval?
Also, random note that might be important, any specifications that we give to the rocket design system are very strong indicators of what we will approve.
A) Use some fancy AI system to create a rocket design, optimizing according to some specifications that we write down, and then sample rocket designs output by this system until you find one that a human approves of.
This does not seem like a realistic thing someone might do, because if the AI doesn't output a rocket design that a human approves of after a few tries, wouldn't the AI user decide that the specification must be wrong or incomplete, look at the specification and the AI's outputs to figure out what is wrong, tweak the specification, and then try again? (Or alternatively decide that the AI isn't powerful enough or otherwise suitable enough for the task.) Assuming this iterative process produces a human-approved rocket design within a reasonable amount of time, it seems obvious that it's likely safer than B since there's much less probability that the AI does something unintended like creating a message to the user.
Given this, I'm having trouble understanding the motivation for the question you're asking. Can you give another example that better illustrates what you're trying to understand?
I am confused about how gradient descent (and other forms of local search) interact with Goodhart's law. I often use a simple proxy of "sample points until I get one with a large U value" or "sample n points, and take the one with the largest U value" when I think about what it means to optimize something for U. I might even say something like "n bits of optimization" to refer to sampling 2n points. I think this is not a very good proxy for what most forms of optimization look like, and this question is trying to get at understanding the difference.
(Alternatively, maybe sampling is a good proxy for many forms of optimization, and this question will help us understand why, so we can think about converting an arbitrary optimization process into a certain number of bits of optimization, and comparing different forms of optimization directly.)
One reason I care about this is that I am concerned about approaches to AI safety that involve modeling humans to try to learn human value. One reason for this concern is that I think it would be nice to be able to save human approval as a test set. Consider the following two procedures:
A) Use some fancy AI system to create a rocket design, optimizing according to some specifications that we write down, and then sample rocket designs output by this system until you find one that a human approves of.
B) Generate a very accurate model of a human. Use some fancy AI system to create a rocket design, optimizing simultaneously according to some specifications that we write down and approval according to the accurate human model. Then sample rocket designs output by this system until you find one that a human approves of.
I am more concerned about the second procedure, because I am worried that the fancy AI system might use a method of optimizing for human approval that Goodharts away the connection between human approval and human value. (In addition to the more benign failure mode of Goodharting away the connection between true human approval and the approval of the accurate model.)
It is possible that I am wrong about this, and I am failing to see just how unsafe procedure A is, because I am failing to imagine the vast number of rocket designs one would have to sample before finding one that is approved, but I think maybe procedure B is actually worse (or worse in some ways). My intuition here is saying something like: "Human approval is a good proxy for human value when sampling (even large numbers of) inputs/plans, but a bad proxy for human value when choosing inputs/plans that were optimized via local search. Local search will find ways to hack the human approval while having little effect on the true value." The existence of adversarial examples for many systems makes me feel especially worried. I might find the answer to this question valuable in thinking about how comfortable I am with superhuman human modeling.
Another reason why I am curious about this is that I think maybe understanding how different forms of optimization interact with Goodhart can help me develop a suitable replacement for "sample points until I get one with a large U value" when trying to do high level reasoning about what optimization will look like. Further this replacement might suggest a way to measure how much optimization happened in a system.
Here is a proposed experiment, (or class of experiments), for investigating how gradient descent interacts with Goodhart's law. You might want to preregister predictions on how experiments of this form might go before reading comments.
Proposed Experiment:
1. Generate a true function V:Rn→R. (For example, you can write down a function explicitly, or generate a random function by randomly initializing a neural net, or training a neural net on random data)
2. Generate a proxy function U:Rn→R, which can be interepereted as a proxy for V. (For example, you can generate a random noise function W, and let U=V+W, or you can train a neural net to try to copy V)
3. Fix some initial distribution μ on Rn, which will represent random sampling. (For example the normal distribution)
4. Define from μ some other distribution ^μ, which can be interpreted as sampling points according to μ, then performing some kind of local optimization according to U. (For example, take a point x according to μ, then perform k steps of gradient ascent on U, or take a point x according to μ, sample k more points all within distance ε of x, and take the one with the highest U value)
5. Screen off the proxy value by conditioning points sampled from μ and ^μ to be in a narrow high band of proxy values, and compare the corresponding distribution on true values. (For example, is E(V(x)|U(x)∈(y,y+ε)) greater when x is sampled from μ or ^μ?)
So, after conditioning on having a high proxy value, μ represents getting that high proxy value via sampling randomly until you find one, while ^μ represents a combination of random sampling with some form of local search. If μ does better according to the true value, this would imply that the optimization via gradient descent respects the true value less than random sampling.
There are many degrees of freedom in the procedure I describe above, and even more degrees of freedom in the space of procedures that do not exactly fit the description above, but still get at the general question. I expect the answer will depend heavily on how these choices are made. The real goal is not to get a binary answer, but to develop an understanding of how (and why) the various choices effect how much better or worse local search Goodharts relative to random sampling.
I am asking this question because I want to know the answer, but (maybe due the the experimental nature) it also seems relatively approachable as far as AI safety question go, so some people might want to try to do these experiments themselves, or try to figure out how they could get an answer that would satisfy them. Also, note that the above procedure is implying a very experimental way of approaching the question, which I think is partially appropriate, but it may be better to think about the problem in theory or in some combination of theory and experiments.
(Thanks to many people I talked with about ideas in this post over the last month: Abram Demski, Sam Eisenstat, Tsvi Benson-Tilsen, Nate Sores, Evan Hubinger, Peter Schmidt-Nielsen, Dylan Hadfield-Menell, David Krueger, Ramana Kumar, Smitha Milli, Andrew Critch, and many other people that I probably forgot to mention.)