Everything is Artificial Intelligence
If your only tool is a hammer, then every problem looks like a nail. This overly-user cliché seems very apropos when thinking about the rush to suggest artificial intelligence for practically every system and flaunt machine learning as a silver bullet to solve whatever value the new product or service is purported to deliver.
Indeed, AI technology can be very powerful. Recent advances in machine learning, aided by ubiquitous connectivity and distributed cloud computing, demonstrate how potent this technology is and promise much more to come. But AI-based systems can also be difficult to build and even more so to perfect and scale and deploy.
- Artificial intelligence systems are very impressive under limited controlled testing. But people inexperienced with AI tend to underestimate the effort to generalize and validate prototypes for real-world situations.
- Certain types of machine learning algorithms are easily influenced by latent biases in the data used to train and validate them. Unbeknown to their users, these systems will repeat and amplify the same biases.
- The opaque nature of certain AI algorithms and software vendors that consider their algorithms proprietary and aren’t willing to discloses how they operate result in “unexplainable AI”: systems whose output cannot be reverse-engineered and explained. Errors, inaccuracies and biases are practically impossible to detect in advances and may be discovered only while the system is deployed.
AI-based systems are rapidly becoming increasingly powerful. But they can also be imperceptibly biased, hard to validate and control, and arguably unethical, threaten privacy and outright dangerous. No wonder the public opinion on AI is quite lukewarm. And similar concerns are often raised by AI scientists and in business circles.
But is artificial intelligence really the silver bullet it purports to be? Is it the only, or even the best technology to solve every problem and build every new product?
Or is AI becoming this proverbial hammer that we use indiscriminately just because it’s in vogue?
Are there other robust and proven approaches that have been around long enough and can accomplish the task at lower cost and risk? For instance, can clustering and regression analysis be used in place of a neural network to discover the relationships between data variables? Can statistical process control (SPC) and control charts provide the same level of hardware monitoring and improve overall equipment effectiveness (OEE) as promised by machine learning systems? (In fact, some predictive maintenance systems touting machine learning are nothing more than old-fashioned SPC systems with fancy user interfaces…)
“When you are fundraising, it’s AI;(Baron Schwartz)
When you are recruiting, it’s machine learning;
When you are implementing, it’s linear regression.”
Even when AI tools do offer superior classification, prediction and control policy planning capabilities, there may be instances in which AI can be used for experimentation and prototyping, but the actual implemented will benefit from using different methods and technologies that can be validated and scaled faster and easier that AI.
Choose your AI hammers prudently!