Conceptualizing the Limits of Artificial Intelligence
This week I will be writing a series of posts about the limits of artificial intelligence, better defined as machine learning, and how to spot dubious claims or uses of machine learning.
Over the past several years, applying a machine learning system to analyze large amounts of data has become a popular method for selling new systems and cloud based computing products. These approaches are extremely viable in certain areas such as in business to business transactions and nonlinear optimization problems, but there are several areas that have proven resistant to this treatment.
The application of machine learning to the physical environment remains the most challenging problem for data scientists, despite the large number of resources poured into those projects. That can be seen by the problems with automated driving systems and with sensor fusion projects such as the troubled F-35 program. Combining multiple sensors with differing data sets remains a problematic area. For example, combining a radar image with camera and lidar data remains a troubling area for self-driving auto programs. A broader point about the limits of machine learning can be gleaned from these struggles.
Machine learning systems have difficulty with dirty or incomplete data sets. Where the data view is not nearly omniscient, a machine learning program will struggle without an inordinate amount of data to analyze. When we try to model the physical world for a machine learning system, we are forced to provide limited data that has significant omissions and blind spots. The takeaway lesson is that when evaluating the success and cost-effectiveness of applying a machine learning system to a problem the primary question is the quality of the data that you can provide to the system. If the data set is not clean or complete, you will have to supply a machine learning system with much more data and more resources that may make it not cost-feasible.
Machine learning will be eminently useful in circumstances where the data presents a clear picture, such as analyzing MRI data and making diagnoses from that data. It will remain limited in situations where the data is not as forthcoming, such as driving on a snow-covered road in on a snowy night in Buffalo.
When a machine learning project is presented to you, make certain to rigorously analyze the data sources and the quality and quantity of the data needed to optimize the machine learning system. Doing so will quickly prove the quality and cost-effectiveness of the product or project.
Machine Learning is the future! this already has applications in areas such as management, marketing, sales and retail and is a promising branch that will bring even more innovations to the corporate market. By applying the right methodologies and using an appropriate data set, there is the possibility to predict - with good confidence - business opportunities that would hardly be discovered by human analysis. This is therefore a great competitive advantage, which will bring many advantages and stand out from your competitors.
Thanks for the push Jason, I think a lot of businesses need to look into this!
One of the biggest issues with ML is certainly the over hype of the technology but more so than that it’s the “black box” theory or more simply put not being able to discern why the algorithm came to the conclusion it did. We’re doing a lot of research on adversarial ML and how bad actors are using the technology. Happy to share more with you offline, if interested.