At aimultiple.com we are always talking about how AI systems are already changing the way companies are run, however it is easy to forget that we are still at the beginning of the AI revolution. We looked into the current and potential future economical impact of learning algorithms to understand how AI will evolve.
Explore today’s most impactful learning algorithms
Andrew Ng masterfully summarizes current commercial impact of learning algorithms in this 2 minute video. We summarize his main points below after also incorporating the topics highlighted by Yann LeCun’s famous cake.
Most AI applications in the corporate world rely on supervised learning. In supervised learning applications, AI systems use a large set of labeled data, like sales data at individual customer level to make predictions like a customer’s potential likelihood of buying. Andrew Ng rightly says that this is the area with most commercial impact today.
However, future potential for supervised learning is unfortunately limited. Though supervised learning is a great learning algorithm, its hunger for data is its primary Achilles’ heel. Without labeled data, there can be no supervised learning. And creating labeled data is either expensive or infeasible in many important domains. Different AI approaches are needed to tackle this lack of data problem. Therefore self-supervised learning is emerging by eliminating the necessity of data labeling process.
Transfer learning involves using labeled data in one domain to make predictions in another domain. For example, an AI system drives for thousands of hours in a digital simulation and uses those learnings while driving in the real world.
This approach is widely used in autonomous cars and other autonomous systems that need to make physical real world interactions. Since real world experience is expensive, simulations provide a learning environment for AI systems until they are mature enough to handle the challenges of the real world. And need for simulation learning continues as systems mature. As AI systems interact in the real world and make mistakes, researchers create similar digitized environments, allowing AI systems to solve hard problems in the simulated world, without taking risks like damaging equipment. The Atlantic’s article on Google Waymo’s efforts to improve their self-driving technology explains this approach in detail.
Unsupervised learning includes a set of clustering algorithms that can understand patterns without any labels. This approach has the most potential areas of application as it can work without labeled data. However more scientific progress is required before unsupervised learning makes a significant commercial impact.
In reinforcement learning, software agents take actions to maximize a reward function. Reinforcement learning is a flexible approach that allows software agents to make decisions and explore in an unsupervised manner. However application areas of reinforcement learning beyond games are limited compared to other approaches like supervised learning. Yann LeCun famously summarized commercial potential of reinforcement learning saying that it is the cherry on a very large cake, cake representing potential of unsupervised learning.
Hope you enjoyed our view of AI approaches from the point of view of commercial impact. For more specific examples, you can check out AI applications in marketing, sales, customer service, IT, data or analytics. And If you have a business problem that is not addressed here:
How can we do better?
Your feedback is valuable. We will do our best to improve our work based on it.