The principle of Free Energy applied to neurobiology by Professor of Neuroscience Karl Friston at University College London is probably one of the most groundbreaking and influential concepts in the development of Artificial Intelligence.
Karl Friston introduced the concept of free energy as the main mechanism that explains the functioning of the human brain, and therefore the concept of intelligence. This concept goes far beyond the human brain and can be applied to other biological systems as well as to artificial intelligence itself. This step is of tremendous importance if we think about how to design the future of artificial intelligence and its implications for everyone.
Let’s start at the beginning. What is this about Free Energy? Let’s think about the famous equation that comes to express that human satisfaction is equal to the real result minus the expectations we had. Behind this expression, there is something profound that reflects our way of processing the environment. On the one hand, we feel, and we do so through literally billions of sensors spread throughout our body. On the other hand, the brain receives data from this universe of sensors and needs to organize them efficiently in order to have as accurate a “vision” of that environment as possible. That is to say, it has to predict, look for patterns, which help it to form an “image” of the environment. This wonderful process has no end. It is a continuous path of trial and error that tries to minimize the differences between what it predicts to be (its “image” of the environment), and what the “sensors” transmit to it that they feel. That difference is what is called Free Energy.
Let’s see some implication of this concept. If Free Energy were zero, our systems would be perfectly “calibrated” and would be giving us exact predictions of what surrounds us and its expected behavior. Let’s say we would be very satisfied. However, let’s think about what would happen if this Free Energy were large. That means we didn’t get it right. In other words, what the sensors transmit, the brain is not capable of associating it with a pattern that allows it to adequately predict the environment. This is fundamental and implies that any system that pretends to self-organize (biological or not) must look for a state of minimum Free Energy. This applies fully to both human and artificial intelligence.
This theory is not “merely” a psychological theory. Its roots come from physics and mathematics. As some may have perceived, this concept is an extension of the Bayesian model that conceives the human brain as a machine to make predictions and adjust their errors. Sometimes Free Energy is also called Predictive Error. However, the Bayesian concept is only about the interaction between what we believe and what we perceive. The human brain goes much further. This wonderful organ minimizes those errors through action. That is, the human brain needs to interact with the world. It needs to move through it and thanks to that interaction it corrects what the senses transmit to it. It is absolutely fantastic, and it is called Active Inference. He feels, predicts, reviews the prediction, assumes the error, updates the prediction model, and moves on. Does this sound like some self-adjusting artificial intelligence algorithms to you? Indeed, this goes beyond Machine Learning and goes beyond even the most advanced Deep Learning models.
This is where in my opinion the concept of Free Energy can contribute most to the future development of Artificial Intelligence. The brain, in its constant search to minimize Free Energy, not only adjusts the predictive model to satisfy the inputs of the sensors. It goes much further and is able to act to check if it is the predictive model that has to prevail over the sensors. I give the same example that Shaun Raviv gives in an article of his published in Wired. Let’s imagine that my brain infers that I’m touching my nose with my index finger of my right hand, but at the same time our motion sensors (for technicians, they are called proprioceptors) tell me that my arm is hanging down, what does the brain do? First, it picks up that there is a discrepancy that generates Free Energy. Secondly, and this is the wonderful thing, in his predictive model he doubts the sensors and to confirm that doubt, he makes me raise my arm and have my index finger touch any part of my face. By confirming this initial sensory error, it recalibrates the sensor, not the predictive model. It is a continuous process of perception, action, planning, and problem-solving with one goal, to minimize Free Energy.
Let us not think that this is something of the distant future, it is rather of the immediate present. While I was at the Mobile World Congress in Barcelona I had the chance to dine with a Spaniard who is a reference in the world of artificial intelligence in Palo Alto and one of the people who knows more about this self-learning approach in Artificial Intelligence. He told me that one of the first and most lucrative applications of this concept was in the area of audiovisual media (“Media”) to learn from the viewer and propose content tailored to their needs. Julie Pitt, head of machine-learning at Netflix, has already been incorporating these Free Energy principles since 2014. According to them, the beauty of these approaches lies in the fact that they not only serve to make the algorithms developed work in a known environment, but they are also able to act in any environment, even new and unknown.
Let’s go back to the human brain. Let’s look at the neurobiological implications of this theory of Free Energy. When Free Energy is high, the brain tries to minimize it, but if it cannot because it is not capable of assigning an adequate predictive model to what the sensors transmit to it, mental illnesses appear. Someone with schizophrenia may be the result of his inability to assign a model of the world that matches what his eyes transmit to him, provoking hallucinations and delusions. In fact, in recent years, the concept of Free Energy is being applied for a better understanding of diseases such as anxiety, depression, psychosis, or even certain symptoms of autism, Parkinson’s or psychopathy. It’s not just about knowing which areas of the brain are malfunctioning or which neuronal connections are badly calibrated, according to Friston, it’s about having a mathematical model of calculation that deals with what we perceive.
The latter is of paramount importance in Artificial Intelligence and lies in the following reasoning. If the principle of Free Energy offers a unified explanation of both how the brain works and when it malfunctions, it is reasonable to think that it could be a path that would allow us to create a “brain” from scratch. The perspectives are extraordinary.
Spanish version published at: