The Uncanny Intuition of Deep Learning to Predict Human Behavior

Credit: https://unsplash.com/search/poker?photo=GikVY_KS9vQ

One of the most promising applications of Deep Learning (DL) is in its use to enhance interaction with computers. DL is particularly suited for this since it has an intuitive capability that is similar to biological brains. It is able to handle the inherent unpredictability and fuzziness of the natural world.

This however does imply that improving AI to human interaction will require some level of capability of an AI to predict human behavior. This is a requisite capability and interesting enough there has been some recent research on this topic.

The conventional behavioral model approach is to have social scientists build analytic models (statistical models in most cases) of human behavior and to employ a distribution fitting exercise to verify validity of the model. The DL approach however is different since a prior model is not required to begin with. Rather, the DL system learns to predict by observing the behavior of human participants. We can get a glimpse of this ground breaking technique by examining some recent research work that uses this approach.

MIT trained a Predictive Vision system on YouTube videos from shows like “The Office” and “Desperate Housewives” to predict whether two persons will hug, kiss, shake hands or slap a five. The trained the Deep Learning system on 600 hours of video. The system was able to predict an action 43 percent of the time. This compares to previous algorithms that could only predict 36 percent of the time.

One pragmatic use of predict human behavior is to do so in the confines of a car. Brains4Cars uses a sensor fusion deep learning system based on LSTMs to anticipate driver behavior 3.5 seconds before it happens. It uses a collection of sensors such as cameras, tactile sensors and wearable devices to make its predictions:

A recent NIPS 2016 paper uses Deep Learning to predict strategic behavior. In most systems the assumption is that the participants perform in a perfectly rational manner and are based from insights from cognitive psychology and experimental economics. However in this system, one that is based on Deep Learning, the system learns a cognitive model without the need for expert knowledge. This system is able to outperform system that are built from expertly constructed features:

Extending beyond just making predictions, Deep Learning systems have been used to assist in contexts with human negotiation. In a paper, “Reinforcement Learning of Multi-Issue Negotiation Dialogue Policies” the authors used Reinforcement Learning and a hand-crafted agenda based policy and evaluated them by having each negotiate against each other in different settings. What was discovered was the RL model consistently outperformed the hand crafted agenda based model. In addition, human’s were asked to rate both systems and the result was the RL based approach was rated to be more “rational”.

Finally, in an impressive act of engineering, a team from the Czech Republic and Canada created a Poker playing system that played 33 professional pokers players from 17 countries and gained a win rate that was an order of magnitude better than a good player rating. The team coined their creation DeepStack:

DeepStack takes a fundamentally different approach. It continues to use the recursive reasoning of CFR to handle information asymmetry. However, it does not compute and store a complete strategy prior to play and so has no need for explicit abstraction. Instead it considers each particular situation as it arises during play, but not in isolation. It avoids reasoning about the entire remainder of the game by substituting the computation beyond a certain depth with a fast approximate estimate. This estimate can be thought of as DeepStack’s intuition: a gut feeling of the value of holding any possible private cards in any possible poker situation. Finally, DeepStack’s intuition, much like human intuition, needs to be trained. We train it with deep learning using examples generated from random poker situations. We show that DeepStack is theoretically sound, produces substantially less exploitable strategies than abstraction-based techniques, and is the first program to beat professional poker players at HUNL with a remarkable average win rate of over 450 mbb/g.
The DeepStack algorithm is composed of three ingredients: a sound local strategy computation for the current public state, depth-limited lookahead using a learned value function over arbitrary poker situations, and a restricted set of lookahead actions.

There’s additional commentary about DeepStack in Technology Review:

The researchers compare DeepStack’s approximation technique to a human player’s instinct for when an opponent is bluffing or holding a winning hand, although the machine has to base its assessment on the opponent’s betting patterns rather than his or her body language. “This estimate can be thought of as DeepStack’s intuition,” they write. “A gut feeling of the value of holding any possible private cards in any possible poker situation.”

In this context, not only is DeepStack able to perform accurate predictions of the behaviors of its opponents, it does so in a way that its own behavior is not predictable!

To summarize, Deep Learning is able to make prediction on tacit behavior of humans as well as rational behavior. The ability to anticipate behavior, predict behavior or win in games of bluffing are extremely advantageous tools to have in one’s business arsenal. However, if you still aren’t convinced that Deep Learning can predict human behavior, then you might want wonder why Facebook has a job opening in this kind of work: Facebook’s Mysterious Job Listing Sounds Like It’s Working on How to Read Your Mind”.

Strategy for Disruptive Artificial Intelligence: https://gumroad.com/products/WRbUs