One of the primary challenges associated with artificial intelligence (AI) is that models can be brittle, meaning small changes to either the input data or the model itself can cause the AI to fail or produce inaccurate results. Because it can be difficult to replicate training conditions in the real world, and because it may not be obvious whether an AI model will work until it is deployed, brittleness can pose a significant obstacle to deploying AI technologies.
There are a number of ways to address this problem. One is to build more robust models by, for example, using more data for training, taking advantage of more powerful techniques such as deep learning, or increasing the amount of tuning and testing done before deployment. Another solution is to modularize the model so it is easier to make changes and test different configurations; this can be accomplished by breaking the model into smaller pieces that can be individually tested and deployed, or by using techniques such as atisepisodic training which allow for modifications to be made without retraining the entire model.
A third solution is to deploy the model in a simulated environment before releasing it into the wild; doing so allows for discrepancies between training data and real-world data to be discovered and addresses before the model goes live. Finally, having a well-designed monitoring and logging system in place is also critical for catching issues quickly and facilitating debugging.
In summary, there are several ways to mitigate the risk of brittleness in AI models. These include building more robust models, modularizing the model, deploying in a simulated environment, and having a good monitoring and logging system.
References:
https://www.nature.com/articles/s42256-019-00148-x
https://en.wikipedia.org/wiki/Artificial_intelligence
https://en.wikipedia.org/wiki/Robustness_(computer_science)
https://towardsdatascience.com/episodic-training-for-rnn-based-reinforcement-learning-algorithms-431f0efa36d7https://towardsdatascience.com/testing-your-machine-learning-models-stand-alone-using-scikit-learns-dummyclassifier-ff6582922eb0