Between the time you read your Wi-Fi password on the coffee shop menu board and the time you can return to your laptop to enter it, you need to keep it in mind. If you’ve ever wondered how your brain does it, you’re asking a question about working memory that researchers have struggled for decades to explain. Now, MIT neuroscientists have published a key new insight to explain how it works.
In a study in PLOS Computational Biology, scientists at the Picower Institute for Learning and Memory compared measurements of brain cell activity in an animal performing a working memory task with the output from several computer models that represented two theories of the underlying mechanism for taking information into account. . The results strongly favored the new idea that a network of neurons stores information by making short-lived changes in the pattern of their connections, or synapses, and contradicted the traditional alternative that memory is maintained by neurons that remain active. persistently (like an idling engine). ).
Although both models allowed information to be taken into account, only the versions that allowed synapses to transiently change connections (“short-term synaptic plasticity”) produced patterns of neural activity that mimicked what actually happened observed in real brains at work. The idea that brain cells maintain memories by always being “on” may be simpler, acknowledged lead author Earl K. Miller, but it doesn’t represent what nature does and can’t produce the sophisticated flexibility of thought. which can arise from the flashing. neuronal activity supported by short-term synaptic plasticity.
“You need these kinds of mechanisms to give working memory activity the freedom it needs to be flexible,” said Miller, the Picower Professor of Neuroscience in MIT’s Department of Brain and Cognitive Sciences (BCS). “If working memory were just a sustained activity, it would be as simple as a light switch. But working memory is as complex and dynamic as our thoughts.”
Co-lead author Leo Kozachkov, who earned his PhD from MIT in November for theoretical modeling work that includes this study, said matching computer models with real-world data was crucial.
“Most people think that working memory ‘happens’ to neurons; persistent neural activity gives rise to persistent thoughts. However, this view has come under recent scrutiny because it doesn’t really agree. with the data,” said Kozachkov, who was co-supervised. by co-senior author Jean-Jacques Slotine, professor of BCS and mechanical engineering. “Using artificial neural networks with short-term synaptic plasticity, we demonstrate that synaptic activity (rather than neuronal activity) can be a substrate for working memory. com, in a quantitative sense, and also have functional advantages additional in terms of robustness”.
Combining models with nature
Along with lead author John Tauber, a graduate student at MIT, Kozachkov’s goal was not just to determine how information from working memory could be taken into account, but to shed light on how nature does it. This meant starting with “ground truth” measurements of the electrical activity of “points” in hundreds of neurons in an animal’s prefrontal cortex while it played a working memory game. In each of the many rounds the animal was shown an image that then disappeared. A second later he would see two images including the original and had to look at the original to earn a small reward. The key moment is that second intervention, called the “delay period”, in which the image must be taken into account before the test.
The team consistently observed what Miller’s lab has seen many times before: neurons spike strongly when they see the original image, spike only intermittently during the delay, and then spike again when the images must be recalled for the test (these dynamics are governed by an interaction of beta and gamma frequency brain rhythms). In other words, crossover is strong when information must be initially stored and when it must be recalled, but only sporadic when it must be maintained. The spike is not persistent during the delay.
In addition, the team trained software “decoders” to read information from working memory from measurements of spiking activity. They were highly accurate when the peak was high, but not when it was low, such as in the delay period. This suggests that the spike does not represent information during the delay. But this raised a crucial question: If the rally doesn’t take information into account, what does?
Researchers such as Mark Stokes of the University of Oxford have proposed that changes in the relative strength or “weights” of synapses could store information. The MIT team tested this idea by computationally modeling neural networks that incorporated two versions of each main theory. As with the real animal, the machine learning networks were trained to perform the same working memory task and to generate neural activity that could also be interpreted by a decoder.
The result is that the computational networks that allowed short-term synaptic plasticity to encode information increased when the real brain increased and did not when it did not. Networks that exhibited constant spiking as a method of maintaining memory increased all the time, including when the natural brain did not. And the decoder results revealed that accuracy decreased during the delay period in the synaptic plasticity models, but remained abnormally high in the persistent spiking models.
In another layer of analysis, the team created a decoder to read information from synaptic weights. They found that during the delay period, synapses represented information from working memory that spiking did not.
Among the two versions of the model that featured short-term synaptic plasticity, the more realistic one was called “PS-Hebb,” which features a negative feedback loop that keeps the neural network stable and robust, Kozachkov said.
Working memory functioning
In addition to better matching nature, models of synaptic plasticity also conferred other benefits that likely matter in real brains. One was that the plasticity models retained information in their synaptic weights even after half of the artificial neurons were “ablated.” The persistent activity models broke down after losing only 10-20 percent of their synapses. And, Miller added, just hitting occasionally takes less energy than going up persistently.
Also, Miller said, fast crossing bursts instead of a persistent spike leave room in time to store more than one item in memory. Research has shown that people can hold up to four different things in working memory. Miller’s lab plans new experiments to determine whether models with intermittent spikes and information storage based on synaptic weight adequately match real neural data when animals must consider multiple things rather than a single image.
In addition to Miller, Kozachkov, Tauber and Slotine, the other authors of the paper are Mikael Lundqvist and Scott Brincat.
The Office of Naval Research, the JPB Foundation and ERC and VR Starting Grants funded the research.