Can algorithms tell our dreams?
How neuroscientific methods help us learning more about memory processing during sleep.
We spend around one third of our lives asleep — but the functions of sleep are in large parts still a mystery. What we know is that sleep is a regularly occurring shift in consciousness, leaving us incapable to interact with our surroundings or to consciously focus on ourselves. Despite the apparent inactivity, though, sleep does not seem to be a passive state at all. Using machine learning approaches, neuroscientists confirmed signs of active memory processing during sleep.
WHY DO WE SLEEP?
There are a couple of ideas out there, trying to explain our need for sleep. Sleep is for example thought to restore the immune system, or to recalibrate neuronal activity in the brain.
We are not quite sure which mechanisms regulate our need for sleep and how bodily functions interact with our inner circadian rhythm. And we are only beginning to understand the intense effects that sleep deprivation has on our emotional and physical health.
However, everyone has experienced that a lack of sleep makes it difficult to concentrate or to remember things.
Another key function of sleep therefore seems to be its role in memory consolidation — the formation of stable and useful memories.
HOW IS MEMORY STORED IN THE BRAIN?
Let us first take one step back and think about our understanding of memory. On the one hand, memory helps us remember what we had for supper two days ago or how we celebrated our last birthday. This type of memory is called episodic as it enables us to recall personally experienced episodes from our past. On the other hand, memory refers to knowledge that we acquire throughout our lives. By experiencing similarities across different events, we learn to separate a dog from a cat or how to distinguish a friendly from a hostile person.
If you have a background in computational neuroscience, you will know that teaching an artificial intelligent system, or allowing it to learn by itself, requires large datasets … and time. When acquiring the concept of a dog, for example, the algorithm needs many encounters with different dogs to slowly incorporate the statistics of what makes a dog a dog. If connection weights are adjusted too quickly (the algorithm thinks it knows everything after having seen only a couple of dogs) — the presentation of contradicting information leads to catastrophic interference, a term that describes the severe disruption of existing knowledge.
Let’s now zoom into the human brain to understand how we avoid this problem. According to Demis Hassabis, neuroscientist and founder of DeepMind, deep neural networks work in some ways similar to human neocortex. Consisting of neuronal cell bodies and fibers, this multilayered cerebral structure slowly acquires stable concepts of our surroundings and stores semantic knowledge.
At the same time, however, our brains must be capable to rapidly log single experiences, enabling us to remember single episodes. To satisfy both needs — stable long-term storage and rapid learning — our brains has two learning systems that are linked by bidirectional connections: the above-mentioned neocortex and the hippocampus, a structure that lies deep inside the medial temporal lobe.
THE CONCEPT OF MEMORY REACTIVATION
When and how do these memory systems interact? Let’s say we would like to remember our last birthday. The encoding of our birthday occurs in different specialized brain regions in the neocortex, that take care of emotions, perception and cognition. Our experience is made up by the joy that we experienced when our friends came over, or by the smell of a birthday cake and the flowers we got.
Next to this original, distributed brain activity, a second memory trace is established in the hippocampus, binding together the information from the distributed neocortical areas. That way, neocortical information is integrated into one coherent memory trace. To make the memory of our birthday more stable, our neural activity within this hippocampal-cortical network is reactivated.
You can imagine it a bit like hippocampus holding a miniature version of the memory that is linked to the original memory. Whenever this miniature version of the memory is activated, the associated memory trace in the neocortex gets stronger. Step by step, the reactivation of these connections leads to a strengthening of the neocortical trace. Ultimately, memory in the neocortex becomes independent of the hippocampus.
REACTIVATION DURING SLEEP?
In a groundbreaking study from 1994, the two neuroscientists Matthew Wilson and Bruce McNaughton observed the activity of cells [1] when laboratory rats were exploring a new environment. Groups of cells that were active when the animal went to particular locations in this environment, were also active during the animals’ upcoming sleep. In other words — the sleeping brain rehearsed what the animals had learnt during the day. This reactivation of neural connections during sleep is thought to help transfer memories from the hippocampus to the neocortex.
Up to now, it was hard to demonstrate such a reactivation in humans, because the activity of individual neurons cannot be directly recorded inside the brain and most memories will activate entire networks of brain regions, not only a couple of isolated cells that would be easier to surveil. To get around those problems, researchers from the university of Tübingen applied new statistical pattern detection methods to show that the human brain also actively reprocesses previously learned information during sleep.
HOW ALGORITHMS DETECT REACTIVATION IN THE SLEEPING BRAIN
In their study, brain activity during sleep was recorded, using electrodes that were applied on participants’ head and face [2]. They measured electrical activity in sleeping participants who had previously studied different kinds of pictures under varied learning conditions. They then trained a support vector machine — a computerized algorithm for pattern recognition — to differentiate the learning conditions from the sleeping brain’s activity only. Testing the algorithm with unknown sleep activity data, it was able to predict which kinds of pictures participants had studied before sleep.
Showing that an algorithm is able to link brain activity during sleep with the content of previous learning, clearly supports the idea that learning related information is worked on at night. On top of that, the researchers could show that the amount of reactivation correlated with memory — the more the studied items were “rehearsed” during sleep, the better was participants memory during the next day.
“The stronger the signs of reprocessing in sleep, the better our subjects remembered the material in the morning. Thus, our study provides initial evidence that sleep-dependent processing of memories actually leads to their stabilization”
explains Dr. Monika Schönauer, who has implemented the study and analyzed the data together with her colleagues. According to the researchers
“it is likely that multivariate pattern classification will enable more breakthroughs especially in cognitive neurosciences, allowing the investigation of previously hidden processes like dreams and spontaneous thought processes. “
Even though the mysteries of sleep remain to be fully uncovered, the above mentioned study stresses that sleep is not a passive state at all. Instead, our brain uses this offline state to reprocess information and to strengthen memories. Algorithms cannot tell our dreams — but they sure help us to understand what’s going on inside our heads during sleep.
[1] By the way — the cells that we are talking about here are so called place cells. These are cells which code for a specific location in a specific environment. John O’Keefe, May-Britt Moser and Edvard Moser got a Nobel Prize in 2014 for their research on these cells!
[2] This method is called electroencephalography (EEG) and is able to record the summed activity of the outer layers of neocortical pyramidal cells, but grasping only a fraction of what is really going on inside the brain!
About the author: Dorothee is a CDTM student from Fall 2016 and has just finished her studies in clinical psychology and cognitive neurosciences at the LMU. She has been a member of the above mentioned research group for three years.
