Staged feedback

Algorithms to live by: multi-fidelity Bayesian optimisation

2022-07-05

Status: draft, last updated 2022-09-29

Bayesian optimisation is an algorithmic approach to finding the input which maximises an unknown function, aiming to use the fewest possible evaluations of that function. The algorithm works by choosing an input $$x$$ that might be good, evaluating it $$y = f(x)$$, then updating a model of the unknown function $$f$$ based on the result $$y$$. Then, a new input is chosen to evaluate based on the updated model, and so on. Typically the function can’t be evaluated directly, and instead an emulator which approximates the function must be used. Various emulators might be available, with different properties. For example, some may be fast and cheap but give noisy, biased responses, and others may be slow and costly but give precise, accurate responses.

This set-up reminds me of the process of receiving feedback on an idea or product, which we would like to improve. The true quality of the idea is unknown, but we can learn about it by receiving feedback from others. We might receive feedback, tune the idea to improve it based on the feedback, then submit the idea for more feedback, and so on. There are many different avenues for recieving feedback with very different properties – which probably cannot be just reduced to their statistical noise and bias properties. It might make sense to structure the ordering of how we recieve feedback

To give an example, in the 100th episode of the Clearer Thinking podcast, Spencer described how he improved the podcast by gathering feedback in stages: first from close friends, then interview guests, then listeners to the podcast. Or, when writing a scientific paper, feedback can be elicited from your supervisor, external collaborators, peer reviewers and hopefully eventually the wider scientific community. Before even thinking about writing a scientific paper, it might be a good idea to incubate your ideas gradually by say talking to a friend at lunch, writing a rough blog post or more polished forum post.

Three properties which I think are strategically important when thinking about receiving feedback are: the costs of receiving feedback, the noise and bias of the feedback, and the stage at which the feedback “counts”.

How much does it cost to get feedback?

Sometimes querying different sources has different costs. If I have a question about a topic at work, I might prefer to ask a colleague of around similar seniority, rather than someone more senior, in part as I expect the more senior person’s time to be of higher value. I might prefer to spend a few hours watching videos about squat form online, hopefully getting most of the way there, before hiring a personal trainer for a shorter session to fine-tune my technique.

How noisy or biased is the feedback?

However, lower cost feedback is typically some combination of more noisy or more biased. Continuing the above examples, my colleagues might be less well placed to give accurate feedback than someone more senior. In the squat form case, watching videos online might ingrain incorrect bad habits early on, which may later be difficult to correct, even with the help of a personal trainer.

Does the feedback count?

By feedback “counting”, I mean that it is summative rather than formative1. That is, at what stage does the feedback have consequences, positive or negative? For example, in the case of a scientific paper, we could naively assume the feedback only begins being consequential when the paper is submitted and disseminated more widely2. For a product, it’s important that it is improved sufficiently during early testing to be strongly presented at launch, as the (financial) feedback of the consumer base is what “counts”. By presenting a bad version of an idea, we risk inoculating others against it, making it more difficult to advance in the future. Particular ideas may need to be incubated longer than others3 to make sure that the research is more solid.

1. I think that this is another way to present the difference between (Bayesian) quadrature and optimisation. In quadrature, all the points are summative, whereas in optimisation only the final point is.↩︎

2. Actually, it’s probably the case that all feedback is consequential here.↩︎

3. For example it could be irresponsible to present poorly thought-out versions of controversial ideas.↩︎

Citation

Howes (2022, July 5). Adam's blog: Staged feedback. Retrieved from https://athowes.github.io/posts/2022-04-24-staged-feedback-loops/
@misc{howes2022staged,
}