Our brain consolidates some
experiences more efficiently than the other experiences that occur at random. A memorable experience has more elements of predictability that aid generalization.
Why is it
that we find it difficult to remember arbitrary things that happen randomly but
on the other we could easily remember things that seem to make sense like the
taste of the coffee? This has to do with the fundamental mechanism that the
brain relies on to store new information from our experiences as memories. The
mechanism of why and how the brain prioritizes certain memories over others has
been largely unknown until quite recently a significant part of the mystery
seems to have been uncovered.
What is generalization?
Generalization
is a process by which our brains find similarities or patterns in previously
learned information on the experiences with the environment to predict new
properties in the environment that were not seen before. This is the same
process by which we can predict that a storm is brewing the moment we get a
pungent aroma in a coastal area.
Brain regions responsible
for memory storage
The
hippocampus is a small region in the brain that stores episodic memory, that is
the memory of different episodes of our life with varying emotional weight. So
it acts like a notebook that records all our memories in chronological order
like episodes in a major television series. Whereas the neocortex is the
outermost surface of the brain with considerable thickness it is the most
recently evolved part of the brain. The neocortex is the primary part of the
brain that deals with logic, critical thinking, and cognition.
What is memory
consolidation?
Memories are
initially encoded in the hippocampus rapidly by various factors. But, storing
memories in such a localized region may be prone to lesion formation or the
rapid encoding of multiple memories can lead to the formation of false memories
or could lead to a ‘distorted reality’. Hence our brain moves some memories to
distribute to different regions in the neocortex. This process is called memory
consolidation.
The research in memory studies done till now suggested that memories remain in the
hippocampus for a short period until it is further consolidated into the
neocortex. However, recent research suggests that some memories remain in the hippocampus and
are never consolidated.
What’s new?
According to
new research, a team of neuroscientists was
trying to find out how system consolidation is affected by the ability of the
brain to generalize or to find recognizable patterns useful for future
prediction. To do this they created a set of artificial neural networks that
mimics the parts of the brain that are responsible for memory retrieval and
storage and mainly consolidation. These areas are most notably the hippocampus
and the neocortex.
Artificial neural networks
The
neuroscientists built an artificial neural network similar to the one used in
machine learning and artificial intelligence known as the
teacher-notebook-student network. The teacher network is built with fixed and
unchangeable ‘weights’ between the nodes. ‘Weights’ are the strengths of the
connections between the nodes.
They used a student
network that was of the same size as that of the teacher network to maximize learning
but was built with learnable and variable weights between the nodes in contrast
to the teacher. The student network learns these weights from the fixed number
of input-output pairs that are generated by the teacher network. This is
similar to how our brains understand the environment from the inputs-output
generated by our environment. For example, when we are seeing a fish, we recognize
it as a ‘fish’ by seeing its streamlined body, gill slits, and fins here the
‘fish’ is considered as the output, and the various features that help us
recognize it are considered as the input. The ability of our brains to
recognize a fish is what is considered complete learning in a neural network.
They built a
notebook network that was directly to the student network just like how the
hippocampus is finely interconnected with the neocortex. The notebook was a recurrent
neural network whose main specificity is to directly encode the input-output
data that is been fed to the student by the teacher. This is similar to how the
hippocampus rapidly encodes our everyday events as episodic memory into its
structure as soon as our neocortex (student) experiences them. So the
artificial neural network that was used in this experiment was structurally
similar to the brain areas that are responsible for memorization and learning.
The research
The
neuroscientists modeled system consolidation in the artificial neural network
by activating the notebook network repeatedly which further activated student neurons
therefore facilitating student learning. This is similar to the consolidation
of memories encoded in the hippocampus into the neocortex. In previous theories, all of the hippocampal memories are consolidated hence to
test this theory the scientist optimized the notebook activation for maximum
memorization.
They found
that the consolidation to transfer all the hippocampal memories into the
neocortex leads to bad generalization of the memories. They found that complete
consolidation can take our ability to make meaningful predictions and the
ability to make broad assumptions and statements. It also means losing our
ability to react to changes that cannot be controlled by us or the changes of
the immediate future.
But this
does not seem to be happening in real life we can predict near future events to
a reasonable degree of accuracy and we can assume certain things that could
turn out to be fairly right. So they found the existing model to be inaccurate
so they formed a new model where the consolidation of memories occurs only when
it is useful for generalization.
According to
this model, they found that the system consolidated most of the memories when
the teacher provided inputs that were significantly predictable and followed a
certain pattern that the student network was able to replicate the structure of
the teacher network. When the teacher network was unpredictable with added
noise then the student network couldn't completely consolidate the memories of
the notebook because it would lead to internalization of the noise in the student
network hence leading to bad generalization which may lead to improper
prediction.
Conclusion
The research
done above has given us the idea that memory consolidation from the hippocampus
to the neocortex only facilitates generalization. When the memories as
unpredictable elements such as the weather forecast of every day you may
remember the weather column of the particular newspaper that you read every day.
This is because the hippocampus ‘picks’ the predictable elements from the
memories to consolidate into the neocortex making it easily accessible.
What remains inconclusive?
Although the
research is experimentally verified with the artificial neural network the real
mechanism in which the brain measures the predictability of various experiences
is unknown. Research on biological subjects should give us an idea on this
subject soon.