Artificial intelligence’s capacity to do tasks comparable to those of humans is indisputable as it grows in strength. Artificial neural networks are capable of learning computer code, language, and picture recognition. They restructure themselves as they learn, modifying the strengths of their connections to enhance performance, just like the circuits in the brain.
Many neuroscientists wonder what they can learn about how the brain learns from these machine systems. What models are most likely to be informative?
Scientists at MIT’s McGovern Institute for Brain Research and the K. Lisa Yang Integrative Computational Neuroscience (ICoN) Center contend in a paper published on October 9 in the journal PNAS that the network’s capacity to self-organize without the assistance of a developer is a critical indicator of the biological plausibility of an artificial neural network’s approach to learning. They demonstrate that an artificial neural network may construct its own learning apparatus. A collection of randomly connected units can rearrange itself into an efficient information-processing network, learning from the data it is trained on, with a few straightforward instructions about how and when connections should change.
Learning Biologically
According to Qianli Liao, a K. Lisa Yang ICoN Center Postdoctoral Fellow in the labs of MIT professors and McGovern Institute members Tomaso Poggio and Mark Harnett, “current AI technology is working very well, but it doesn’t necessarily use a learning algorithm that is used by the brain.” Liao co-wrote the PNAS paper with graduate student Yulu Gan and MIT research scientist Liu Ziyin. According to Liao, when an artificial neural network analyzes training data and gets feedback about its errors, engineers create the rules governing how connections change.
The method is effective in teaching AI to perform its duties. However, networks created in this manner might not provide much insight into biological learning, Liao and his colleagues warn. According to Liao, “the brain does not have an external designer or architect.” The brain must self-organize. Furthermore, we must understand how basic laws might organize the brain.
various brain regions may have various rules that alter neuronal connections as an animal learns, or the rules may alter as the brain develops. To develop new treatments and preventive measures, neuroscientists are trying to understand how these rules work and how aging or illness may disrupt them. One possible use, according to Liao, is creating brain-computer interfaces that enable paralyzed patients to control neuroprosthetic limbs or restore communication when the brain’s language functions are lost.
According to him, “none of these can function without understanding the principles underlying neuron-neuron interactions.” “Just creating a circuit that functions well won’t accomplish that.” Instead, the ability of an artificial neural network to organize itself—not just its connections, but also the machinery that modifies those connections—should be used to determine its usefulness to neuroscience.
Self-organized Artificial Intelligence
Liao, Ziyin, and Gan set out to create and evaluate a self-organizing computational network with learning capabilities in collaboration with Poggio and Hartnett. According to Liao, the construction of the network starts with several layers of randomly connected units that resemble neurons. Every neuron-like component in each layer is governed by the same guidelines about the proper adjustment of connections. According to him, “they just interact with each other using that rule, and they develop some sort of behavior.”
As humans and other animals learn from their experiences, the team’s network applies the kind of basic rules believed to reorganize brain circuitry. For instance, “neurons that fire together wire together” in both the machine network and the brain. Put differently, the link between two connected neurons gets stronger when they are engaged simultaneously.
The group demonstrated how a self-organizing network controlled by such rules may learn to identify patterns and categorize pictures. According to their idea, learning is made possible by self-assembling motifs that develop in accordance with basic laws and subsequently work as building blocks for circuits. These motifs rearranged themselves as their network learned using an optimization technique called gradient descent, which is comparable to the method used in the majority of contemporary machine learning systems and involves gradually adjusting connections in response to failures in order to improve performance.
Gradient descent is implemented by the majority of contemporary AI using a back-propagation algorithm, which neuroscientists claim is unlikely to be used by the brain. The team discovered that its self-organizing network performed on par with networks built to learn in this manner.
Linking Machines And Brains
According to Poggio, discovering a biological equivalent of back-propagation could lead to important advancements in neurology and artificial intelligence.
“We will have a strong connection at the level of fundamental mechanisms between artificial and natural networks—between machines and brains—if the brain actually uses the synaptic motifs we propose to accomplish what backpropagation does for artificial neural networks,” he claims.
According to Liao, a self-organizing network is an essential first step toward biological plausibility even though it may not learn in the same manner as the brain.
“The first high-level framework is self-organization,” he claims. “We can discuss the next topic once we have mastered self-organization as a measure.”
Although the group’s primary goal is to gain a deeper understanding of the brain, Liao claims that information might eventually enhance machine learning systems as well, allowing engineers to create networks that function better or more like humans.