Short YouTube explainer about AI explainability and trust. The narrator argues that classifiers must provide human-interpretable explanations (highlighting supporting and counter-evidence) so humans can judge and augment AI decisions across images, tabular data, and text. The technique described is model-agnostic, improves human decision-making in tests, and has source code available.
Questions this video answers
LARRUZ (Two Minute Papers)
2018-03-03
Chapters
Dear fellow scholars, this is two minute papers with through over 200 episodes of this series, we talked about many learning based algorithms that are able to solve problems that previously seemed completely impossible. They can look at an image and describe what they depict in a sentence, or even turn video game graphics into reality and back. Amazing new results keep appearing every single week. However, an important thing that we need to solve is that if we deploy these neural networks in a production environment, we would want to know if we are relying on a good or bad AI's decision. The narrative is very simple. If we don't trust a classifier, we won't use it. And perhaps the best way of earning the trust of a human would be if the AI could explain how it came to a given decision. Strictly speaking, a neural network can explain it to us, but it would show us hundreds of thousands of neuron activations that are completely unusable for any sort of intuition, intuitive reasoning. So what is even more difficult to solve is that this explanation happens in a way that we can interpret. An earlier approach used decision trees that described what the learner looks at and how it uses this information to arrive to a conclusion. This new work is quite different. For instance, imagine that a neural network would look at all the information we know about a patient and tell us that this patient likely has the flu. And in the meantime, it could tell us that the fact that the patient has a headache and sneezes a lot contributed to the conclusion that he has the flu. But the lack of fatigue is notable evidence against it. Our doctor could take this information and instead of blindly relying on the output, could make a more informed decision. A fine example of a case where AI does not replace but augment human labor. An elegant tool for a more civilized age. Here we see an example image where the classifier explains which region contributes to the decision that this image depicts a cat and which region seems to be counter evidence. We can use this not only for tabulated patient data and images, but text as well. In this other example, we try to find out whether a piece of written text is about Christianity or atheism. Note that the decision itself is not as simple as looking for a few keywords. Even a mid tier classifier is much more sophisticated than that, but it can tell us about the main contributing factors. A big additional selling point is that this technique is model agnostic, which means that it can be applied to other learning algorithms that are able to perform classification. It is also a possibility that an AI is only right by chance. And if this is the case, we should definitely know about that. And here in this example, with the additional explanation, it is rather easy to find that we have a bad model that looks at the background of the image and and thinks that it is the fur of a wolf. The tests indicate that humans make significantly better decisions when they lean on explanations that are extracted by this technique. The source code of this project is also available. Thanks for watching and for your generous support and I'll see you next time.
