In the past few years, we witnessed a rise in the popularity of a particular Machine Learning technique known as Artificial Neural Networks. This technique was first envisioned in the 1940s and developed through the second half of the twentieth century. In the last decade, thanks to the increase in computer power and data availability, neural networks started showing promising results in many different areas, such as computer vision, natural language processing, and artificial intelligence. Recently, terms such as Deep Learning have become mainstream, causing many professionals to ask the question: Can those new technologies help me and my company make better decisions? To answer this, we first need to understand what a Neural Network is and how it can solve problems.
Neural networks are biologically-inspired systems that, when trained with machine learning techniques and a fair amount of data, can (hopefully) learn to predict outputs based on inputs. One of their main advantages is their generality, meaning that the same overall training algorithm can work on a wide variety of problems. On the other hand, Neural networks are what we call "black box" systems, which means that their internal states do not give any insights on why or how they are producing their outputs. In practice, this is the cause of two of the biggest problems in real-life applications of black box systems:
The name "black box" comes from the fact that you cannot "see'' what it is doing. You can only feed the system with inputs and get the outputs, no questions asked! The problem is that in many practical situations, the why is as important as the what.
As an example, let's consider a hypothetical company that produces a given number of different products that consume common resources. This company wants to know the ideal proportion between its products that will maximize its profit, and for that, it uses a black box system. The only thing this system can do is to propose, based on the inputs, a product proportion that the company will have to blindly accept. In contrast, transparent systems such as simulation and optimization models can provide the user with additional information such as:
Of course, this additional information will depend on how you built the model – but that is the key point! Transparent models use scientific knowledge and hypotheses from experts that can be analyzed and tested. These confrontations are, in some cases, vital for effective decision-making.
Because these systems are not auditable, hidden biases can result in unfair or wrong decisions. For example, more than 2,000 researchers from Google, MIT, Microsoft, and Yale have called on academic journals to halt the publication of studies on "criminal recognition" systems based on facial features. The researchers claim that this type of system can perpetuate racial biases present in the data. In general, it's hard to tell if a high-accuracy black box model trained on past historical data will perform well in new contexts or just perpetuate trends from the past.
But at the end of the day, should I use this type of model or not? Well, as always, the answer is: It depends! If your problem is well defined, you are better off using transparent solutions, such as mathematical models, that help you understand your context and test your hypothesis. If, on the other hand, you are dealing with an unstructured problem with no known rules (such as the problem of how humans see or understand language), a Deep Neural Network is probably the best option. Of course, you can always have "Gray Box'' solutions that incorporate subsystems from the black box and transparent models, trying to use the best of each technology. Either way, the best option to stay competitive is to choose a solution that better fits your context instead of blindly seeking the latest trend.
Author: Vinícius Mello - Consultant at Cassotis Consulting
Co-author: Fabio Silva - Senior Manager at Cassotis Consulting