Neural networks and deep learning are the most exciting innovation in AI. However, in their current form they are often not practical for enterprise use.
Zoubin Ghahramani points out these limitations to deep learning:
» very data hungry (e.g. often millions of examples)
» very compute-intensive to train and deploy
» poor at representing uncertainty
» non-trivial to incorporate prior knowledge and symbolic representations
» easily fooled by adversarial examples
» finicky to optimize: non-convex + choice of architecture, earning procedure, initialization, etc.
» uninterpretable black boxes, lacking in transparency, difficult to trust
— Zoubin Ghahramani presenting at AAAI 2018
Moonshot projects have the resources and expertise to overcome these limitations. However, moonshots are not the model for the median enterprise machine learning project. To succeed in the enterprise we will have to develop practical approaches that fit within enterprise constraints. Often we will need to prioritize enterprise utility over model performance. We will need to prefer approaches that surface explanation friendly features. Generally we will need to deal with common fallacies that stop us from providing the tool set that enterprise usage requires.
Hat tip to Brian Ruttenberg for pointing out Zoubin’s presentation.