Hi HN,<p>I wanted to showcase a quick reseach project I have done for a [lesswrong post](<a href="https://www.lesswrong.com/posts/y9tnz27oLmtLxcrEF/constructability-ai-safety-via-pull-request" rel="nofollow">https://www.lesswrong.com/posts/y9tnz27oLmtLxcrEF/constructa...</a>), where we tried to attempt to construct a flower recognizer in a way that is already understandable (the basic idea of constructability being to create neural networks in a way that is already understood instead of trying to reverse engineer them)<p>The idea of creating AI systems that we are able to explain the behaviour of and predict for their erratic behaviors seemed important, and so we wanted to prototype what it would look like to have an image recognizer we can explain in all their details.<p>Models differ in their performance, we have been able to achieve 75% F1 score pretty reliably independent of the architecture though. More importantly, the notion of decomposing a model into many submodel does seem to hold in a way that I did not expect (it keeps composing even with retraining the submodels).<p>I am overall quite glad with the results. Let me know if you have any questions!