Personalization will revolutionize our lives: from medical treatments and medicines tailored to our unique physiology and psychology, to marketing and advertising that provides actual value by understanding our unique spending behavior. Large companies like Google understand this, as evidenced by CEO Sindar Pinchai’s statement about moving towards a “personal google for everyone.” Artificial intelligence (AI) and machine learning (ML) techniques are uniquely suited to personalization. But personalization requires technologies that are personal — that learn our individual behaviors. After all, as the saying goes — the devil’s in the details — and in the details we are all unique. As a result the technologies enabling true personalization will need to effectively perform two critical functions:

(1) 1-to-1 behavioral analysis of every individual, and

(2) continuously learn from past and present behavioral data

Function (1) is critical because every individual’s behavior, physiology, psychology, etc. is different. Only by zooming in to resolve behavior at the individual level can personalization have real value. Function (2) is critical because individual behavior continuously changes over time.

Artificial intelligence and machine learning technologies can perform these functions; however, many of the technologies used today are not up to the task for two reasons. First, existing techniques, like Neural Networks (the technology behind “Deep Learning” approaches) apply the same logic to every object they are trying to classify. To see this, one need look no further than Google’s Deep Dream picture generator, in which deep learning neural networks are used to redraw images. The resulting pictures contain elements resembling different objects, from fish to eyes, which lead people to conclude that the algorithm has reinterpreted the image in an intelligent way. However, the neural network is actually just trying to identify characteristics of the images it was initially trained on in the new input picture. If the neural network was trained on faces, it will try and find elements of faces in the input image and replace these elements with face-like elements — mouths, eyes etc. — and it will do this to every input image. We can see an example of this in the before and after images below.

                                       Comparison of different images before and after Deep Dream

The circled features in these images illustrate how Deep Learning neural networks attempt to find similar features, even in completely disparate input images. In this example, all the circled features exhibit significant “bird-like” similarities suggesting that the underlying neural network was trained to classify images of birds by looking at millions of images of birds. This simple example highlights the “when you’re a hammer, everything’s a nail” approach these algorithms take to classification. This would suggest that, if these techniques were applied to try and classify human behavior or preferences, they would be unable to resolve individual behavior, instead relying on recognizing similarities across individuals. Additionally, because they attempt to “see” similarities across input data, they might infer characteristics of individuals that do not, in fact exist, much as the neural network in the earlier example attempted to find birds in images where no birds existed.

These techniques have shown success when trying to classify items that do not change over time, like classifying images of objects (birds, cats etc.) or natural language or processing, these approaches can add value. However, this is the opposite of what is required for effective personalization tools, which, by definition, require each individual to be characterized according to their own behavior and wants. Not only will behavior differ between individuals, but individual behavior will also change over time. Methods relying exclusively on large sets of historical data will be unable to effectively characterize these changes.

What is required is a comprehensive approach that combines the strengths of multiple AI technologies with real-time individual behavioral analysis to characterize changing behavior. Only then will we realize the true potential of artificial intelligence and machine learning to deliver on the enormous potential of personalization.