AI models and Neuroscience. Concepts, approaches, and perspectives.

Studying the human brain hasn’t always been the intention of people in artificial intelligence (AI). Yet neuroscience plays a crucial role in AI research. The number of scientists who agree with the validity of the direct connection between the study of the human brain and the success of artificial intelligence has progressively increased over the last decades.

Does everyone agree that the key to a successful model depends on understanding how the brain works?

No. Is this the one and only, universally accepted approach? Also no. However, there is no doubt that the two fields complement each other. In the future, they will only continue to assimilate even more.

The human brain is still far from being completely unraveled, which is why the forthcoming research is going to be such an incredibly complex project. While studying our minds, many experts have found incredible similarities between the brain and the Universe. Many of us already know that our bodies, or the atoms of which our bodies consist, were made of ashes and the stars. This discovery has inspired many poets, writers, and artists in general. Even now, it is rather unimaginable. Nonetheless, naturally, this has drawn the attention of astrophysics and neuroscientists. The visual resemblance between a neuron in the brain and clusters of galaxies is undeniable.

A: Hippocampal mouse neuron. B: Cosmic web

Yet, is it objective?

Several discoveries have been made to objectively compare how alike these two matters genuinely are.

Research by Vazza and Feletti was focused on finding quantifiable correlations, in which 4 μm thick slices of the human cortex and 25 megaparsecs thick “pieces” of the Universe were magnified and analyzed. At 40x magnification, the network of neurons has shown distinct structural similarities that can be measured and examined via two techniques. One is the “network degree centrality”, which has proved that the nucleus’s radius is considerably shorter than the length of connecting elements in a network. Curiously, this ratio is very much identical to that of galaxy clusters and filaments, respectively. The second one, the “clustering coefficient,” has also been calculated and revealed “remarkable” similarities between the two objects of research. Vazza and Feletti have stated that the degree of similarity is higher than other fundamental structures (biological and physical).

Vazza F and Feletti A (2020) Figure 3: Above: reconstructed connections between nodes for three examples of networks (blue lines, superimposed to the density contrast maps). Below: are distributions of clustering coefficient and of degree centrality for all slices.


Is the brain actually an appropriate sample for artificial systems?

Many scientists believe it is. One of them is Henry Markram, a neuroscientist whose goal is to build a complete simulation of the human brain inside a supercomputer. However, to make this grand project as detailed as possible, enormous resources are needed. Markram’s goal is to truly understand how our brain works in order to further use it in almost any field you could imagine.

Although constructing a model does not necessarily require replicating the original natural system, the practice of such a method is far simpler, given the fact that biology is already complex, and, indeed, it has demonstrated some of the major achievements in both AI and neuroscience. As of today, plenty of outstanding studies have been conducted concerning brain signals as a basis. Its outcomes have largely contributed to today’s artificial intelligence models.

Modern artificial intelligence models

IBM Research model

In their paper called “Context-Attentive Bandit: Contextual Bandit with Restricted Context,” the IBM Research team has focused on humans’ ability to pay selective attention and adapt quickly. They wanted to implement the idea of selecting the most important current information from the endless data and instantly making decisions in their model. The team created an algorithm that could focus on the most valid material based on their so-called “reward” or external feedback. One significant thing to mention is that the algorithm could analyze the data in real-time, while new information was being constantly delivered. Consequently, the system learned from its past experience.

The comparison of the IBM team’s online dictionary learning algorithms and the standard learning methods

Jeff Hawkins model

Another example of copying human algorithms has been explained by Jeff Hawkins, a neuroscientist and tech entrepreneur, who challenged today’s AI intelligence. Hawkins claims to know what true intelligence is and explains why AI should recreate it. From his standpoint, there is a “baseline” to intelligence and it consists of four main attributes: learning by moving, or embodiment; building up an overall point; continuous learning; and structuring knowledge via reference frames. As Jeff says, the vast majority of people do not find applying the brain to AI important. Although many researchers are inspired by the brain, as he admits, they aren’t replicating it, which is their biggest mistake. Jeff Hawkins truly believes that actually understanding neuroscience and what it means to be intelligent is going to lead to the most anticipated breakthroughs in intelligent machines.

Christos Papadimitriou model

According to Christos Papadimitriou, professor of computer science, whose presentation on speech and word representations in the brain was included in an online workshop on the understanding of deep learning by Google, the way in which our neural activity is turned into human speech, logic, reasoning, and planning is still uncovered. To understand the role of sets of neurons that represent something, or “assemblies”, Papadimitriou introduces his mathematical model called “interacting recurrent nets”. As stated in the model, the brain is supposedly divided into multiple different areas with neurons interacting inside each area. The model suggests three characteristics: randomness, plasticity, and inhibition.

The professor is convinced that these actions can be mathematically proven using assembly calculus and can be the reason for some of the human cognitive functions mentioned above. Testing assembly calculus to parse English sentences was one of the implementations of Christos and his team. The engine is able to produce sentences with a given sequence of words. The algorithm itself works just like the brain does, through simulated neuron spikes. Besides, it is said that the model’s speed and frequency could be compared to the ones that happen in the brain.

Other areas the team is working on are learning and planning which are done by children at a young age. Papadimitriou is certain that something like assembly calculus can be used as an exact copy of our brain while doing computations.

Evolwe model

Our company, Evolwe, is currently working on improving our conversational agent, or empathetic AI-based bot, by means of the methods listed above. Getting inspiration from cognitive science, neurobiology, neuropsychology, consciousness, etc., Evolwe is seeking after the almighty artificial coach who’s capable of identifying and reducing stress and anxiety, providing personal guidance, setting goals, and building confidence, and so much more. The methodology draws on diagnostic materials from psychology to determine the current state of a person. Understanding the signs allows AI to identify them and offer recommendations and meditation practices.

To draw a conclusion, several crucial limitations need to be mentioned.

One is, undoubtedly, ethics. While studying the animal brain might certainly help to understand some core principles, animals certainly don’t have some human behaviors, such as speech. Extra freedom in the exploration of the human brain is going to allow researchers to observe how speech happens and teach a network to reproduce it. However, “researchers are limited by ethical considerations in terms of how much they can intervene in processes in the healthy human brain.”

Another obstacle is the unanswered question of how the process of learning occurs. According to Tomaso Poggio, current networks’ ways of learning are nothing like a human baby’s. Teams of computational neuroscientists still haven’t figured out how to replicate it.

Additionally, some human mechanisms are considered to be evolutionary aspects of our genes, such as morality. The concept is so new and obscure that it will require many years to master.

The Evolwe team strives to actively use the latest developments. And this, of course, will allow our customers to comfortably use the most advanced products in this area, including our empathetic coach.

The new generation of AI is expected to go further and thoroughly explore the concepts of human intelligence. And while some of the present theories might sound vague and pretentious, the best is yet to come.