Where does the brain store each type of memory?

Memory represents not one scientific mystery, but many. Neuroscientists and psychologists have discovered that several types of memory coexist in our brain: episodic memories of past experiences, semantic memories of facts, short- and long-term memories, and more. These memories often have different characteristics and even appear to be located in different parts of the brain. But it has never been clear what property of memory determines how or why it is classified in a certain way.

Now, a new theory backed by experiments using artificial neural networks suggests that the brain may sort memories by evaluating how likely they are to be useful as evidence in the future. The study suggests, in particular, that many memories of predictable things, from facts to repeated, useful experiences — such as what we ate for breakfast or on the way to work — are stored in the neocortex, where they can contribute to generalizations about the world. .

Memories that are unlikely to be useful — like the taste of that special drink you had at that party — are stored in a seahorse-shaped memory bank called the hippocampus. Separating memories in this way, based on their usefulness and generalizability, can improve the reliability of memories to help us deal with new situations.

The new theory’s authors — neuroscientists Weinan Sun and James Fitzgerald of the Howard Hughes Medical Institute’s Janelia Research Park, Andrew Sacks of University College London, and their colleagues — described it in a recent article in Normal neuroscience. It updates and expands the well-established idea that the brain has two interconnected, complementary learning systems: the hippocampus, which rapidly encodes new information, and the neocortex, which gradually integrates it for long-term storage.

James McClelland, a cognitive neuroscientist at Stanford University who pioneered the idea of ​​complementary learning systems in memory but was not part of the new study, noted that it “addresses aspects of generalization” that his group had not considered when it proposed the new study. Theory in the mid-nineties.

Scientists have known that memory formation is a multi-stage process since at least the early 1950s, thanks in part to studies on patient Henry Mollison — who for decades was known in the scientific literature simply as HM because he suffered from uncontrollable seizures. on her. Surgeons treat it by removing most of the brain’s structure.

Afterwards, the patient seemed completely normal in many respects: his vocabulary was intact; He had childhood memories and remembered other details about his life before surgery. But he always forgot the nurse who took care of him. During the decade she cared, she had to report back every morning. He has completely lost the ability to create new long-term memories.

Molaison’s symptoms helped scientists discover that new memories were first formed in the hippocampus and then gradually moved to the neocortex. For a time, it was assumed that this happened with all the persistent memories. However, as researchers began to see an increasing number of examples of memories that remained dependent on the hippocampus long-term, it became clear that something else was going on.

To understand the reason behind this anomaly, the authors of the new paper turned to artificial neural networks, because the function of the millions of interconnected neurons in the brain is incomprehensibly complex. These networks are “an approximation of biological neurons” but much simpler than real networks, Sachs said.

Just like living neurons, they have layers of nodes that receive data, process it, and then provide weighted outputs to other layers of the network. Just as neurons influence each other through their synapses, nodes in artificial neural networks adjust their activity levels based on input from other nodes.

The team connected three neural networks with different functions to develop a computational architecture they called the teacher-laptop-student model. The network of teachers represents the environment in which an organism can find itself and provides information about experience. The network of notebooks represents the hippocampus, quickly encoding all the details of every experience the teacher provides.

The student network was trained according to the teacher’s standards, referring to what was recorded in the notebook. “The goal of the model is to find the neurons – the ganglia – and learn the connections [descrevendo] “How can they revamp their activity pattern,” Fitzgerald said.

The laptop network’s memory iterations trained the student network in a general pattern through error correction. But the researchers also noted an exception to this rule: If the learners’ network was trained using too many unexpected memories — noisy signals that deviated too much from the rest — this degraded the learners’ network’s ability to learn the generalized pattern.

“From a logical point of view, this makes a lot of sense,” Sun said. Imagine receiving packages at home, he explained: If the package contains something useful for the future, “like cups and plates,” it seems reasonable to bring it inside. And store it away permanently. But if the package contains a Spider-Man Halloween costume or a promotional flyer, there is no need to clutter the house with it. These items can be stored in a different location or disposed of.

The study provides an interesting convergence between the systems used in artificial intelligence and those used in brain modeling. It’s an example where “the theory of these artificial systems has given some new conceptual ideas for thinking about memories in the brain,” Sachs said.

There are parallels, for example, with the operation of computerized facial recognition systems. They could start by asking users to upload high-resolution photos of themselves from different angles. Connections within the neural network can create a general perception of what a face looks like from different angles and with different expressions.

But if you send a photo that “contains your friend’s face, the system won’t be able to determine the expected facial expression between the two,” Fitzgerald said. This impairs generalization and makes the system less accurate in recognizing normal faces.

These images activate specific input neurons and flow activity through the network, adjusting the connection weights. With more images, the model adjusts the weights of connections between nodes to reduce output errors.

But the simple fact that an experience is unusual and does not fit into a generalization does not mean that it should be discarded and forgotten. Conversely, it may be very important to remember exceptional experiences. This appears to be the reason why the brain categorizes its memories into different categories that are stored separately, with the neocortex being used for reliable generalizations and the hippocampus for exceptions.

McClelland said this type of research raises awareness about “the fallibility of human memory.” Memory is a limited resource, and biology had to make the most of the limited resources. Even the hippocampus does not have a perfect record of experiences.

Every time an experiment is recalled, changes occur in the network connection weights, causing the memory elements to become more weighted. This raises questions about the circumstances under which “eyewitness testimony” takes place. [poderiam] “It should be protected from bias and the influence of repeated rounds of questions,” he said.

The model can also provide insights into more fundamental questions. “How do we build reliable knowledge and make informed decisions,” said James Anthony, a neuroscientist at California Polytechnic State University who was not involved in the study. This shows the importance of evaluating memories to make reliable predictions: too much noisy data or unreliable information can be as unsuitable for training humans as it is for training AI models. / Translated by Renato Prilorenzo

Original story republished with permission from Quanta Magazine, an editorially independent magazine supported by the Simons Foundation. Read the original content at The usefulness of memory guides us to where the brain stores it