Creation of Memory: The Low Level of Memory
I. Introduction
I used to spend an hour everyday to walk around the campus to enjoy the environment, the weather, and the atmosphere at around ASI Company. Until one day, I started to question myself, why did I not forget what I had learned after the walk? The answer might seem trivial, which is what I had learned for test transferred to my long-term memory. Yet, the question became even more interesting when we think about the events that only happened once, and we still remembered up to now without practicing or rehearsal. The reason why the question is interesting is because that we only look at the problem from a high level. That means we look at the memory as a chunk where long/short-term memory are models. Instead, we should look at the memory problem at, the “storage site for synaptic strength of neuron cells” — the dendritic spines.
II. The origin of memory
The first step is to look at how our memory is formed during learning to do a task. Many researchers have shown that dendritic spine formation corresponds to learning and memory when learning new tasks in both humans and advanced animals (Yi et al., 2005). Yang and colleagues conducted the first two-day length experiment in which the mice had to run on the rotarod (2007). First of all, the mice were placed on the motorized rod in the chamber, and the speed of the motor would be increasing from 0 to 100 r.p.m in 3 minutes. The performance of the mice were recorded based on the time latency and rotation speed in which the mice could not keep up. The total performance was the average of all the trials. At the same time, the researchers observed the level of dendritic spine (of layer V pyramidal neurons) formation in the subjects. Then, one group would be trained on the new task (reverse running) for another two days, and the other group would remain training on the same task. The result shows that the dendritic spine formation remained significantly high if they were trained on a different task (reverse running) compared to the one remained on the same (Yang, Pan, & Gan, 2009). Thus, this observation confirmed that our memory during learning the tasks is formed by the change in synaptic connections, especially the synaptic dendritic spines in our cortex (Yi, Yang, Kwon, & Gan, 2005).
III. Pruning the “memory”
After the course of learning or training, a part of the new spines would be pruned or eliminated to save space for the newer spines (Yang, Pan, & Gan, 2009). That is the reason why we only remember the general picture of the older memory instead of full details. Yang and colleagues also investigated this phenomenon by conducting another experiment to measure the spine elimination rate for the mice in different environments: a standard housing environment (SE) and an enriched environment (EE) (2009). First, all the mice were put in the SE condition, and then, some of them would be randomly switched to the EE condition. With each group of the environments, they would be given either skill learning or no skill learning. The skill learning here was referred to as running on the rotarod. They showed that most new spines (> 75%) were eliminated over 2 weeks regardless of the ages and conditions of the mouse. That implied that less than 25% of new spines were remained. However, this fraction would be higher if the mice were trained for longer (4–14 consecutive days). This observation shows that memory will fade away regardless of the environment or the condition that we are in. However, the rate will be slower if we train or learn that knowledge for a longer period. This can explain why studying or scrambling the knowledge the night before the exam does not work well for many of us.
IV. Memory Elimination
An interesting point that we should notice is that the dendritic spines elimination does not happen randomly. Indeed, it will be troublesome if the elimination happens randomly. Imagine we learn how to drive today, and suddenly forget how to drive the day after. Xu and colleagues had two big groups of mice in which one was trained to reach out for food while they were in the chamber, and the other was the control group (2009). The way the control group was organized was a bit complicated. There were three different levels: one receives food only; one received the “shaping”, which meant the mice were allowed to get familiar with chamber and task requirements but were not trained to perform it; and the one with both shaping condition and training phase but the food were out of their reach.
The trained mice were said to be successfully learned when they could reach and put the food to their mouth. The results had shown that the new spines were not stable for the control group. On the other hand, the trained mice’ pre-existing spines became less stable compared to their new spines after the task. Although the task in this experiment was about movement tasks, we can see that the elimination rate of the spines branch depends on the task and how we learn the task. With the trained mice, the new spines got stabilized while their old spines became less stable. On the other hand, in the control group, the new spines were not stable enough, which suggested that they would not fluently perform the task because motor memory (getting the food from the outside) stabilizes learning-induced new spines and destabilizes pre-existing spines (Xu et al., 2009).
The same research group also measured the remaining spines after 120 days. They found that the trained mice remained 42.3% (± 2.9%) new spines while the control group only had 13.5% (± 1.7%) new spines remained. It is worth noting that the way Xu and colleagues observed the dendritic spine formation is similar to Yang’s research (as show above). The conclusion here is that the spines are not randomly pruned, but it depends on how we learn and associate the task. The finding is in line with the phenomenon that sometimes we forget the newer thing compared to the older thing that we have learned.
V. Lifetime of Memory
Does that mean that none of the spines at our birth will remain throughout our life? The answer is no. Cichon and Gan showed that the formation of the dendritic spines was based on the context that we have learned (2016). They demonstrated this idea by measuring Ca2+ in the participants while they were involving in different motor learning tasks (Cichon and Gan 2015). As a result, depending on the task, different branch-specific Ca2+ spikes cause long-term potentiation (LTP) of postsynaptic dendritic spines during the formation. LTP is defined as a persistent increase in synaptic strength that can be induced rapidly by a brief burst of spike activity in the presynaptic afferents. It is also defined as the process of strengthening the connections between two neurons. LTP happens in the forebrain and are thought to be memory storage of the central nervous system.
The discovery suggests that there should be a mechanism that encodes the similarity between contexts so that the brain will “choose” which branch to grow or create new spines on. This also suggests that some memories will last longer than the other because that particular branch is often chosen to grow synaptic on. However, this is still an opened-ended question about how the brain can maintain a small portion of innate spines throughout our life.
VI. Conclusion
Eventually, I want to end this blog why these findings are valuable. First, it is valuable for us in enhancing the way we learn and understand our memory at the biological level. Besides, with the development of Artificial Intelligence (AI), most of the current AI models are suffering from “catastrophic forgetting” which means the model will learn task A until it meets the requirements and then move to task B. The problem with this setup is that the model will forget the previous knowledge from task A. Many publications have been using the inspiration from observations above and apply to their models such as EWC (Kirpatrick et al., 2017), XdG (Masse, Grant, & Freedman, 2018), and many others. However, we have not yet achieved the performance as we train the model on task A only. Therefore, I hope we will see more exciting explorations to improve the quality of AI models as well as having a deeper understanding of how our memory works.
References
Yang, G., Pan, F. & Gan, WB. Stably maintained dendritic spines are associated with lifelong memories. Nature 462, 920–924 (2009). https://doi.org/10.1038/nature08577
Xu, T., Yu, X., Perlik, A. et al. Rapid formation and selective stabilization of synapses for enduring motor memories. Nature 462, 915–919 (2009). https://doi.org/10.1038/nature08389
Cichon, J., & Gan, W. B. (2015). Branch-specific dendritic Ca(2+) spikes cause persistent synaptic plasticity. Nature, 520(7546), 180–185. https://doi.org/10.1038/nature14251
Masse, N., Grant, G. & Freedman, D.. (2018). Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization. Proceedings of the National Academy of Sciences. 115. 10.1073/pnas.1803839115.
Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A. A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., Hassabis, D., Clopath, C., Kumaran, D. & Hadsell, R. (2016). Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences. 114. 10.1073/pnas.1611835114.
