Study urges caution when comparing neural networks to the brain


Neural networks, a style of computing system loosely modeled on the organization of the human brain, form the premise of many artificial intelligence systems for applications such speech recognition, computer vision, and medical image evaluation.

In the sphere of neuroscience, researchers often use neural networks to attempt to model the identical form of tasks that the brain performs, in hopes that the models could suggest recent hypotheses regarding how the brain itself performs those tasks. Nevertheless, a bunch of researchers at MIT is urging that more caution ought to be taken when interpreting these models.

In an evaluation of greater than 11,000 neural networks that were trained to simulate the function of grid cells — key components of the brain’s navigation system — the researchers found that neural networks only produced grid-cell-like activity after they got very specific constraints that are usually not present in biological systems.

“What this implies is that so as to obtain a result with grid cells, the researchers training the models needed to bake in those results with specific, biologically implausible implementation decisions,” says Rylan Schaeffer, a former senior research associate at MIT.

Without those constraints, the MIT team found that only a few neural networks generated grid-cell-like activity, suggesting that these models don’t necessarily generate useful predictions of how the brain works.

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

Modeling grid cells

Neural networks, which researchers have been using for many years to perform quite a lot of computational tasks, consist of hundreds or thousands and thousands of processing units connected to one another. Each node has connections of various strengths to other nodes within the network. Because the network analyzes huge amounts of information, the strengths of those connections change because the network learns to perform the specified task.

On this study, the researchers focused on neural networks which were developed to mimic the function of the brain’s grid cells, that are present in the entorhinal cortex of the mammalian brain. Along with place cells, present in the hippocampus, grid cells form a brain circuit that helps animals know where they’re and how one can navigate to a special location.

Place cells have been shown to fireside each time an animal is in a selected location, and every place cell may reply to a couple of location. Grid cells, however, work very in another way. As an animal moves through an area akin to a room, grid cells fire only when the animal is at one among the vertices of a triangular lattice. Different groups of grid cells create lattices of barely different dimensions, which overlap one another. This permits grid cells to encode numerous unique positions using a comparatively small variety of cells.

This kind of location encoding also makes it possible to predict an animal’s next location based on a given place to begin and a velocity. In several recent studies, researchers have trained neural networks to perform this same task, which is referred to as path integration.

To coach neural networks to perform this task, researchers feed into it a place to begin and a velocity that varies over time. The model essentially mimics the activity of an animal roaming through an area, and calculates updated positions because it moves. Because the model performs the duty, the activity patterns of various units inside the network will be measured. Each unit’s activity will be represented as a firing pattern, much like the firing patterns of neurons within the brain.

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

Nevertheless, the MIT researchers found very different results. In an evaluation of greater than 11,000 neural networks that they trained on path integration, they found that while nearly 90 percent of them learned the duty successfully, only about 10 percent of those networks generated activity patterns that may very well be classified as grid-cell-like. That features networks during which even only a single unit achieved a high grid rating.

The sooner studies were more prone to generate grid-cell-like activity only due to constraints that researchers construct into those models, in keeping with the MIT team.

“Earlier studies have presented this story that in the event you train networks to path integrate, you are going to get grid cells. What we found is that as a substitute, you may have to make this long sequence of decisions of parameters, which we all know are inconsistent with the biology, after which in a small sliver of those parameters, you’ll get the specified result,” Schaeffer says.

More biological models

One in every of the constraints present in earlier studies is that the researchers required the model to convert velocity into a singular position, reported by one network unit that corresponds to a spot cell. For this to occur, the researchers also required that every place cell correspond to just one location, which just isn’t how biological place cells work: Studies have shown that place cells within the hippocampus can reply to as much as 20 different locations, not only one.

When the MIT team adjusted the models in order that place cells were more like biological place cells, the models were still in a position to perform the trail integration task, but they now not produced grid-cell-like activity. Grid-cell-like activity also disappeared when the researchers instructed the models to generate various kinds of location output, akin to location on a grid with X and Y axes, or location as a distance and angle relative to a house point.

“If the one thing that you simply ask this network to do is path integrate, and also you impose a set of very specific, not physiological requirements on the readout unit, then it’s possible to acquire grid cells,” Fiete says. “But in the event you loosen up any of those features of this readout unit, that strongly degrades the flexibility of the network to provide grid cells. In truth, normally they do not, although they still solve the trail integration task.”

Subsequently, if the researchers hadn’t already known of the existence of grid cells, and guided the model to provide them, it will be most unlikely for them to look as a natural consequence of the model training.

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

“Whenever you use deep learning models, they generally is a powerful tool, but one must be very circumspect in interpreting them and in determining whether or not they are truly making de novo predictions, and even shedding light on what it’s that the brain is optimizing,” Fiete says.

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

“Neural networks generally is a useful source of predictions. If you would like to find out how the brain solves a computation, you’ll be able to train a network to perform it, then test the hypothesis that the brain works the identical way. Whether the hypothesis is confirmed or not, you’ll learn something,” says Harris, who was not involved within the study. “This paper shows that ‘postdiction’ is less powerful: Neural networks have many parameters, so getting them to duplicate an existing result just isn’t as surprising.”

When using these models to make predictions about how the brain works, it’s necessary to bear in mind realistic, known biological constraints when constructing the models, the MIT researchers say. They at the moment are working on models of grid cells that they hope will generate more accurate predictions of how grid cells within the brain work.

“Deep learning models will give us insight in regards to the brain, but only after you inject loads of biological knowledge into the model,” Khona says. “In case you use the proper constraints, then the models can provide you with 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.


What are your thoughts on this topic?
Let us know in the comments below.


0 0 votes
Article Rating
1 Comment
Newest Most Voted
Inline Feedbacks
View all comments

Share this article

Recent posts

Would love your thoughts, please comment.x