What Learning Can Learn From AI

Scalable learning is a key differentiator for modern enterprise business. The theory states that the institutions most likely to thrive in today’s changing economic environments will be those that provide opportunities not only to learn faster as a whole organization, but also to learn from other individuals and organizations to create new knowledge. That sounds great, but how does one actually achieve scalable learning. One way learning platforms are working to scale learning is using artificial intelligence (AI) and machine learning in learning technology through content recommendations and personalized learning. But I’m not here to talk about a “Netflix of Learning.”

Instead, by understanding how artificial intelligence seeks to use algorithmic software solutions to recreate processes similar to those of a human brain, we can begin to understand how enterprise learning and development can evolve from these concepts. Because the underlying fields of study have related conceptual goals, it can be enlightening to look at AI research that's trying to recreate the process of learning in software can help us improve the process of human learning.

Let’s take, for example, the challenges AI faces when introducing and applying certain complexities at scale. For instance, creating a reinforcement learning algorithm to learn to play tic-tac-toe is simple compared to developing one to learn to play chess. In both cases, the desired end state is to win the game, but the paths to achieve that state are very different due to the sheer complexity of chess.

When this example is applied to how learning occurs in large enterprise, we can identify similarities—such as:

• a desired end state (i.e. greater competency or career development as opposed to winning the game), and
• elements of reinforcement (e.g. scores, badges, skills gained, or leaderboards as opposed to game pieces captured).

What Is Reinforcement Learning?

Based on the concept of the Markov decision process, reinforcement learning is the act of an agent interacting with an environment, and the interpretation of the agent’s new state following that interaction to determine a reward. The agent is now able to use the information about the state change and resulting reward to improve future outcomes and rewards.

The simplest explanation is the concept of trial and error. Is the agent now in a more favorable state than it was prior to taking action? It’s all about the reward, just like a learner might take a particular action if they feel incentivized or motivated by the potential outcome of learning a new skill.

Reinforcement Learning vs. Evolutionary Computation

If we think of reinforcement learning as trial and error, evolutionary computation is trial and error at scale. While reinforcement learning focuses on a single agent, evolutionary computation seeks to improve an entire population of agents and explores different possible solutions to a problem. Evolutionary computation also encourages novel solutions to problems that are notoriously difficult to solve and places extra focus on solutions that are not only good, but also varied, leading to specialization and comparative advantage opportunities.

So, where reinforcement learning seeks to achieve a preferred state (such as a winning score or a stronger skill set), evolutionary computation adds emphasis to doing so uniquely. This process builds niche development and specialization to help create unique differentiators.

Compared to reinforcement learning, evolutionary computation provides three main improvements:

1. Diversity. Enabled by extreme parallelization via new hardware architectures, a diversity of agents leads to fascinating discoveries via neural networks.
2. Novelty. Operating without existing intuition and developing a niche, agents not only learn effective solutions, but also new and unique solutions.
3. Leaps forward, not incremental steps. Combining the scale, diversity, and novelty of the solutions output by evolutionary computation, progress is measured by significant leaps instead of small increments (i.e. the solutions become unique differentiators).

Reinforcing Scalable Learning as a Key Differentiator

If we replace the concept of an agent in artificial intelligence with human learners, we begin to understand how enterprise learning and development can evolve from these concepts.

So, how can we implement lessons learned from evolutionary computation in organizational learning?

• Embrace a modern learning ecosystem to provide for a variety of experiences. Just like new hardware architectures enable parallelization for evolutionary computation, new software architectures provide the diversity needed to enable new discoveries by human learners.
• Focus on rewarding diversity of thought. As a learning practitioner, investigate the unique paths some of the organization’s highest performers have taken and identify what you can learn from them.
• Create opportunities for collaboration. People with different experiences bring unique perspectives to a collaborative experience. As a result, everyone will learn much more from each other than they would independently.

Human learners may not iterate as fast as an algorithm, and that is why key concepts (such as measuring learning fitness) are important to help reinforce curiosity and, therefore, new discoveries. By rewarding unique and new discoveries—not just adequate solutions—we have the opportunity to replicate the novelty of evolutionary computation and build truly unique differentiators for our organizations. All this builds to create a leap forward in efficiency for an organization, not just an incremental step forward.

Recommended Resources

Because you’re as curious as a neural network, check out the following resources:

• The Data Skeptic podcast is a great resource for artificial intelligence and all things data in a digestible format.
• Check out John Hagel’s work at Deloitte’s Center for the Edge for more information about scalable learning as a differentiator, or what he calls “The Big Shift.”