In the last post we discussed how to help your management think about AI in the right way. But that doesn’t solve the problem of how to govern your latest AI project. Here we list some of the best practices we have learnt through trial and error.
- Define what your AI is going to do, in detail: Make sure that this definition is tight. Not "my AI will help users reach peak productivity." Define in detail what inputs will trigger your AI and what it will do.
- Define what it will learn and what it is going to learn from, in more detail: A lot of AI programs begin with self-learning and end up with a team of developers scripting every single use case and a code base straight out of hell. You don't even need AI for that. Define in detail the input set from which you expect your AI will learn. Chances are, you don't have enough data to train your AI, or the data is proprietary. Defining the input set helps you realistically evaluate what you need to do to build a corpus of data for training.
- Define what it won't do in greater detail: Getting a working AI platform is tough. Be ruthless about pruning the need for every single component to be self-learning. Maybe you can use Luis.ai or IBM Watson for recognizing intents, but you want to limit your AI to intelligently learn how to respond. Do not attempt to create the deep learning neural net of everything.
- Define your edge cases in greatest detail: And make sure that your edge cases don't just define technical parameters, but also social, cultural and linguistic parameters. Understand that your AI will continue to get exposed to more data in the wild, with multiple latent variables and will invariably learn the wrong results. Make sure you have a team that constantly revisits the data and retrains the AI. Having a diverse team to do this is crucial. A mono-culture here will miss out on important social variables and you may just find that your AI does this.
- Spend a lot of time training for the edge cases: A lot of the work within enterprise AI is occupied in training systems to get the right results. What gets missed out is training the system to positively identify the edge cases. What happens in this case is that the error rate in edge cases is high, and since there are an enormous number of edge cases out there, the chances of getting it wrong are extremely high. Make sure that your team is as focused on identifying edge cases as it is on identifying positive cases and rejecting negative ones.
- Gracefully degrade: Define how to gracefully degrade whenever your AI comes across these edge cases. An easy way of implementing a graceful degradation is to let the AI flag the edge cases to human beings for resolution. Having people in the loop to handle ambiguous results and edge cases may a better bet, if the cost of getting it wrong is high.
- Monitor, evaluate and refine: Once your AI gets into production, the amount of data coming its way is going to be of an order of magnitude higher than in training. Making sure you have a continuous process to monitor, evaluate and refine your AI, as defined in the six steps above is even more crucial to the continued success of your platform. Make sure you have it.
While it's easy to join the AI bandwagon and jump head first into development, a large number of failed AI projects happen because they don't have the right governance or the right expectations. Make sure you do or yours may just be the story about a modern day Hans.