TECH | Mar 9, 2017

Multidisciplinary collaboration in the era of Machine Learning

If it is the analyst who translates messages into symbols, it is the expert who helps to interpret the content

M. was a woman born in the late nineteenth century who gave birth to several children. She used to say that before he died one of her sons was replaced three times by an already moribund double. M. claimed to be a descendant of Louis XVIII, King of the Indies, to possess a fortune of 125 billion francs, and, feeling herself threatened, believed to be under surveillance, and that all the people she met were doubles, or doubles of doubles. Within a few years, she was diagnosed with a disorder known as “Capbras Syndrome”. The woman physically and cognitively identified her loved ones but did not recognize them as such, insofar as she was incapable of any emotional intuition and “sense of familiarity”, to the point of believing that her relatives were impostors.

Today we know that learning and decision-making processes in humans are complex and always the result of forms of interaction between the “cognitive” and the “emotional” areas of the brain.

Disorders complementary to the one described exist. Prosopagnosia, recognized in the early 90s, prevents recognition of individuals, even of relatives, although the people involved experience subconscious feelings of familiarity (for example an increase in heartbeat). In these cases, people have to rely – even unconsciously – on “mechanical” strategies that aid recognition: “if the face of the person who comes to visit me in the hospital has this particular shape, or has that particular birthmark on his face, he must be my husband.” The fact of relying on forms of “conscious” recognition of details that permit recognition of objects or individuals is fundamental in cases of “cognitive” deficit. The more complex a problem becomes, the more it is necessary to “break it up” into pieces that make it possible to understand it thoroughly.

The identification of a set of fundamental characteristics (or features), making it possible to retrace a new problem to a familiar one, is an essential prerogative of every learning system and, consequently, also of methods of machine learning, through which – by having an appropriate base information (training set) available – it is possible to teach a machine to make forecasts or calculations based on previously collected stimuli. In the first place, machine learning systems use training sets for learning about complex issues, simplifying them through structures of features, so that they can be identified and understood at a later stage.

Let us consider the Viola-Jones algorithm, one of the first algorithms for the automatic identification (detection) of faces – seen from the front – in real time. To recognize the presence of a face, the algorithm starts from an essential characteristic of faces that is well known both figuratively and in the image processing field: faces are grossly “schematized” in areas of shadow and light that involve groups of rectangular pixels. The area of ​​the nose, for example, is generally brighter than that of the eyes. These pixel areas are searched and processed (Haar Features), while an efficient algorithm (called “boosting”) selects only the features most suitable for characterizing the training set. Identification of the face comes when the presence of features similar to those defined are found within new images and the algorithm verifies that “they are comparable” with those of the training set.

The Viola-Jones algorithm starts from easily understandable assumptions (the characteristic light/shadow contraposition in faces). Its implementation assumes a solid knowledge of both classification techniques and of the world of image processing and photography.

Insiders know that identification of a set of base features that make it possible to characterize a problem, or to make predictions as generally as possible, is rarely a trivial task and is the “hard core” in implementing a successful Machine Learning system.

There exist certain criteria that enable choosing the “most important” features to be considered within the training set for obtaining results consistent with the data available. It is never a trivial choice or one to be made solely on the basis of choice of a general rule and, in many cases, the supervision of an analyst is required. The choice of criteria according to which this selection is made is also up to the analyst and it is crucial to define the quality of the algorithm.

It is clear that the choice of base features is not purely a technical activity insofar as the “point of view” the analyst him/herself might affect analysis of the problem. Because different points of view can lead to different conclusions, it may even happen that, starting from different models, nevertheless reasonable results are obtained, although they are not always concordant. This may be even more significant when the number of potentially useful features increases or when, for example, it is impossible to trace a problem back to a simple comparison between historical series of data.

And it is in cases like these that knowledge of the context becomes essential. While it is certainly desirable for an analyst to have the basic knowledge necessary to understand the phenomena in which he/she is involved for identifying the base features, it is equally important to establish, where possible, a constructive dialogue between analyst and domain expert, who possesses the set of skills required to provide the former with the “best point of view” on the problem.

If it is the analyst who translates messages into symbols, it is the expert who helps to interpret the content, communicating a vision and contributing to turning data into value.

This assumes that analysts and domain experts speak a common language. This dialogue, which assumes highly multidisciplinary connotations, may not always be easy, in Business as in Science. Yet this dialogue is a prerequisite for transforming sector knowledge into value added.

Through the cooperation between analysts and domain experts it is thus possible to recognize with certainty a principle of causality in a correlation link, or the simplified structure of a real process in the schematization of a model and, finally, objectives and concrete actions within a common vision.

Michele Gabusi

Recommended reading for learning more:
“To Understand Facebook, Study Capgras Syndrome”, by Robert Sapolsky, Nautilus, November 2016.
Paul Viola, Micheal J. Jones, “Robust Real-Time Face Detection”, International Journal of Computer Vision, 2004, 57(2), 137-154