One interpretation of this would be imitation learning: teaching a system to imitate human strategies, rather than optimize some objective of its own.
The problem with imitation learning is: since humans are pretty smart, a close imitation of a human strategy is probably going to involve planning in the deliberate service of some values. So if you set a big neural network on the problem of imitating humans, it will develop its own preferences and ability to plan. This is a recipe for an inner optimizer. Its values and planning will have to line up with humans in typical cases, but in extreme cases (eg adversarial examples), it could be very different. This can be a big problem, because the existence of such an AI could itself push us to extreme cases where the AI has trouble generalizing.
Another interpretation of your idea could be “approval-directed agents”. These are not trained to imitate humans, but rather, trained based on human approval of actions. However, unlike reinforcement learners, they don’t plan ahead to maximize expected approval. They only learn to take specific actions more when they are approved of, and less when they earn disapproval.
Unlike imitation learners, approval-directed agents can be more capable than human trainers. However, unlike reinforcement learning agents, approval-directed agents don’t have any incentive to take over control of their reward buttons. All the planning ahead comes from humans, looking at particular sorts of actions and deciding that they’re good.
Unfortunately, this still faces basically the same problem as imitation learning. Because humans are approving/disapproving based on complicated models of the world and detailed thoughts about the consequences of actions, a big neural network has good reason to replicate those faculties within itself. You get an inner optimizer again, with the risks of misalignment that this brings.
Thanks for this response. I heard a similar discussion recently, with someone talking about whether an algorithm’s reward function was activated because it got the answer correct or because it knew it was what the programmers wanted it to do. It’s not clear since the decision-making pathways are not always clear, especially with more complex machine learning.
The inner optimizer thing is really interesting; I hadn’t heard it coined like that before. Is it in AI’s interest (a big assumption that is has interests at all, I know) to become so human-specific that it loses its ability to generalize? Variability would decrease in the population and the probability mechanisms of machine learning would approach certainty, thus rendering the AI basically ineffective.
Is it in AI’s interest (a big assumption that is has interests at all, I know) to become so human-specific that it loses its ability to generalize?
There’s an approach called learning the prior through imitative generalization, that seemed to me a promising way to address this problem. Most relevant quotes from that article:
We might hope that our models will naturally generalize correctly from easy-to-answer questions to the ones that we care about. However, a natural pathological generalisation is for our models to only give us ‘human-like’ answers to questions, even if it knows the best answer is different. If we only have access to these human-like answers to questions, that probably doesn’t give us enough information to supervise a superhuman model.
What we’re going to call ‘Imitative Generalization’ is a possible way to narrow the gap between the things our model knows, and the questions we can train our model to answer honestly. It avoids the pathological generalisation by only using ML for IID tasks, and imitating the way humans generalize. This hopefully gives us answers that are more like ‘how a human would answer if they’d learnt from all the data the model has learnt from’. We supervise how the model does the transfer, to get the sort of generalisation we want.
One interpretation of this would be imitation learning: teaching a system to imitate human strategies, rather than optimize some objective of its own.
The problem with imitation learning is: since humans are pretty smart, a close imitation of a human strategy is probably going to involve planning in the deliberate service of some values. So if you set a big neural network on the problem of imitating humans, it will develop its own preferences and ability to plan. This is a recipe for an inner optimizer. Its values and planning will have to line up with humans in typical cases, but in extreme cases (eg adversarial examples), it could be very different. This can be a big problem, because the existence of such an AI could itself push us to extreme cases where the AI has trouble generalizing.
Another interpretation of your idea could be “approval-directed agents”. These are not trained to imitate humans, but rather, trained based on human approval of actions. However, unlike reinforcement learners, they don’t plan ahead to maximize expected approval. They only learn to take specific actions more when they are approved of, and less when they earn disapproval.
Unlike imitation learners, approval-directed agents can be more capable than human trainers. However, unlike reinforcement learning agents, approval-directed agents don’t have any incentive to take over control of their reward buttons. All the planning ahead comes from humans, looking at particular sorts of actions and deciding that they’re good.
Unfortunately, this still faces basically the same problem as imitation learning. Because humans are approving/disapproving based on complicated models of the world and detailed thoughts about the consequences of actions, a big neural network has good reason to replicate those faculties within itself. You get an inner optimizer again, with the risks of misalignment that this brings.
Thanks for this response. I heard a similar discussion recently, with someone talking about whether an algorithm’s reward function was activated because it got the answer correct or because it knew it was what the programmers wanted it to do. It’s not clear since the decision-making pathways are not always clear, especially with more complex machine learning.
The inner optimizer thing is really interesting; I hadn’t heard it coined like that before. Is it in AI’s interest (a big assumption that is has interests at all, I know) to become so human-specific that it loses its ability to generalize? Variability would decrease in the population and the probability mechanisms of machine learning would approach certainty, thus rendering the AI basically ineffective.
There’s an approach called learning the prior through imitative generalization, that seemed to me a promising way to address this problem. Most relevant quotes from that article: