AI is a “moving target,” in that the cutting edge keeps cutting deeper. That being said, AI is still in its infancy, and there are sizeable limitations to its implementation. In an article for McKinsey & Company, Michael Chui, James Manyika, and Mehdi Miremadi discuss five of these limitations at length:
- Data labeling
- Obtaining massive training data sets
- The “explainability” problem
- The generalizing of learning
- Bias in data and algorithms
Young and Dumb
For AI to learn, it needs to be told what types of data it is looking at—and right now, all that data labeling is being done manually. Two methods of getting around this limitation are reinforcement learning and generative adversarial networks (GANs). Reinforcement learning rewards AI for successful behaviors and punishes them for unsuccessful ones. GANs pit two networks against each other to improve both networks’ understanding of a subject.
Another limitation of AI is the fact that thousands or sometimes millions of data records are required for AI to perform at the required levels. “One-shot learning” is a potential workaround for this limitation:
In this still-developing methodology, data scientists would first pre-train a model in a simulated virtual environment that presents variants of a task or, in the case of image recognition, of what an object looks like. Then, after being shown just a few real-world variations that the AI model did not see in virtual training, the model would draw on its knowledge to reach the right solution.
Then there is the explainability problem. Explainability is “the use of a heuristic problem-solving hierarchy to facilitate the explanation of hypothesis-directed reasoning.” Techniques are being developed to identify which inputs of data are most relevant to the needed output.
Typically, the experiences an AI has while learning one task are not easily transferred to assist in learning a new task. Learning is not generalized in this regard. And one more limitation of AI stems not from the technology, but rather from the people creating it. Unconscious bias in the minds of people who build AI can create AI that is equally and insidiously biased.
The authors go on to make these suggestions about how to stay on top of AI:
- Do your homework, get calibrated, and keep up.
- Adopt a sophisticated data strategy.
- Think laterally.
- Be a trailblazer.
This is just a top-level explanation. For the deep, deep details on this subject, you can view the full article here: https://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/what-ai-can-and-cant-do-yet-for-your-business