Study urges caution when comparing neural networks to the brain | MIT News

Neural networks, a type of computer system loosely modeled on the organization of the human brain, form the basis of many artificial intelligence systems for applications such as speech recognition, computer vision and image analysis. medical.

In the field of neuroscience, researchers often use neural networks to try to model the same kind of tasks the brain performs, in the hope that the models might suggest new hypotheses about how the brain itself performs. these tasks. However, a group of MIT researchers urge that more caution be taken when interpreting these patterns.

In an analysis of more than 11,000 neural networks that were trained to simulate the function of grid cells – key components of the brain’s navigational system – the researchers found that the neural networks only produced activity like grid cell only when given very specific stresses that are not found in biological systems.

“What this suggests is that to get a result with grid cells, researchers train the models needed to integrate those results with specific, biologically implausible implementation choices,” says Rylan Schaeffer, former associate of senior researcher at MIT.

Without these constraints, the MIT team found that very few neural networks generated grid cell-like activity, suggesting that these models do not necessarily generate useful predictions about brain function.

Schaeffer, who is now a graduate student in computer science at Stanford University, is the lead author of the new study, which will be presented at the 2022 Conference on Neural Information Processing Systems this month. Ila Fiete, professor of brain and cognitive sciences and fellow at MIT’s McGovern Institute for Brain Research, is the lead author of the paper. Mikail Khona, an MIT graduate student in physics, is also an author.

Grid Cell Modeling

Neural networks, which researchers have used for decades to perform various computational tasks, consist of thousands or millions of processing units connected together. Each node has connections of varying strengths with other nodes in the network. As the network analyzes huge amounts of data, the strengths of these connections change as the network learns to perform the desired task.

In this study, the researchers focused on neural networks that were developed to mimic the function of brain grid cells, which are found in the entorhinal cortex of the mammalian brain. Along with place cells, found in the hippocampus, grid cells form a brain circuit that helps animals know where they are and how to navigate to another location.

Place cells have been shown to fire whenever an animal is in a specific location, and each place cell can respond to more than one location. Grid cells, on the other hand, work very differently. When an animal moves through a space such as a room, grid cells only trigger when the animal is at one of the vertices of a triangular lattice. Different groups of grid cells create networks of slightly different dimensions, which overlap. This allows grid cells to encode a large number of unique positions using a relatively small number of cells.

This type of location coding also makes it possible to predict the next location of an animal based on a given starting point and speed. In several recent studies, researchers trained neural networks to perform this same task, known as path integration.

To train the neural networks to perform this task, the researchers introduce a time-varying starting point and speed. The model essentially mimics the activity of a wandering animal in a space and calculates updated positions as it moves. As the model performs the task, the activity patterns of the various network units can be measured. The activity of each unit can be represented by a firing pattern, similar to the firing patterns of neurons in the brain.

In several previous studies, the researchers reported that their models produced units with activity patterns that closely mimic the gating patterns of grid cells. These studies concluded that grid-cell-like representations would emerge naturally in any neural network trained to perform the path integration task.

However, the MIT researchers found very different results. In an analysis of over 11,000 neural networks they trained in path integration, they found that while nearly 90% of them learned the task successfully, only about 10% of those networks generated activity patterns that could be classified as grid-like cells. . This includes networks where even a single unit has achieved a high grid score.

According to the MIT team, earlier studies were more likely to generate grid cell-like activity solely because of the constraints the researchers build into these models.

“Previous studies have presented this story that if you train networks to integrate paths, you will get grid cells. What we’ve found is that instead you have to do this long sequence of choosing parameters, which we know are incompatible with biology, and then in a small portion of those parameters you’ll get the desired result,” explains Schaeffer.

More biological models

One of the constraints found in previous studies is that the researchers required the model to convert velocity to a single position, reported by a lattice unit that corresponds to a locus cell. For this to happen, the researchers also required that each place cell correspond to a single location, which is not how biological place cells work: studies have shown that place cells in the seahorse can respond to up to 20 different pitches, not just one.

When the MIT team adjusted the models to make the place cells more like biological place cells, the models were still able to perform the path integration task, but they no longer produced any activity of place. grid cell type. Grid cell-like activity also disappeared when researchers asked the models to generate different types of location output, such as location on a grid with X and Y axes, or location as distance and d angle from an origin point.

“If the only thing you’re asking this network to do is path integration and you put a set of very specific, non-physiological demands on the read unit, then it’s possible to get cells from grid,” says Fiete. “But if you relax any of these aspects of this read unit, it greatly degrades the ability of the network to produce grid cells. In fact, they usually don’t, even though they still solve the task of path integration.

Therefore, if researchers had not already known about the existence of grid cells and guided the model to produce them, it would be highly unlikely that they would appear as a natural consequence of model formation.

The researchers say their findings suggest greater caution is warranted when interpreting neural network models of the brain.

“When you use deep learning models, they can be a powerful tool, but you have to be very careful in interpreting them and determining whether they really make de novo predictions, or even shed light on what the optimizes the brain,” explains Fiete.

Kenneth Harris, professor of quantitative neuroscience at University College London, says he hopes the new study will encourage neuroscientists to be more careful when stating what can be shown by analogies between neural networks and the brain.

“Neural networks can be a useful source of predictions. If you want to learn how the brain solves a calculation, you can train a network to perform it and then test the hypothesis that the brain works the same way. Whether the hypothesis is confirmed or not, you will learn something,” says Harris, who was not involved in the study. “This paper shows that ‘postdiction’ is less powerful: neural networks have many parameters, so getting them to replicate an existing result isn’t as surprising.”

When using these models to make predictions about brain function, it is important to consider realistic and known biological constraints when building the models, explain the MIT researchers. They are currently working on grid cell models that they hope will generate more accurate predictions of how grid cells work in the brain.

“Deep learning models will give us insight into the brain, but only after injecting a lot of biological knowledge into the model,” says Khona. “If you use the right constraints, the models can give you a brain-like solution.”

The research was funded by the Office of Naval Research, the National Science Foundation, the Simons Foundation through the Simons Collaboration on the Global Brain, and the Howard Hughes Medical Institute through the Faculty Scholars Program. Mikail Khona was supported by the MathWorks Science Fellowship.

Leave a Comment