Machine learning methods are very effective for tasks such as image classification and voice recognition, but they are not so successful in predicting the actions of chaotic or random systems, like the weather. In this community lecture, External Professor Michelle Girvan describes an exciting new approach combining knowledge-free prediction systems with knowledge-based models that dramatically improves prediction.
Key message: New and exciting challenges remain about how to scale this approach up to ensembles and layers of coupled machines (1:01:48), pointing the way to ever more sophisticated uses of machine learning, even in the medical area.
19:29 What is the difference between artificial intelligence, machine learning and deep learning?
27:30 Reservoir computing provides a simple way to train artificial neural networks with feedback loops.
37:00 In reservoir systems, the predicted state looks very much like the true state even after the predictions become unstable.
40:00 Having many reservoirs acting in parallel is highly effective for very large chaotic systems like predicting the weather.
41:30 Combining ‘top down’ reservoir computing with a ‘bottom up’ system such as causal state modeling works much better on random, stochastic systems, like predicting user activity on Twitter.
47:40 A system with a hybrid architecture combining the reservoir system with other models and gives dramatically improved predictions even with a poor model and a small reservoir.
54:15 Ensembles of machines could produce ever more accurate predictions. And human intervention combined with machine learning could produce even better results.
59:30 The hybrid method could offer a step towards explainable machine learning, answering questions like what are the weaknesses in machine learning and knowledge-based models.