Memory in Deep Learning Systems Part I: The Neuroscience Perspective
Memory modeling is an active areas of research in the deep learning space. In recent years, techniques such as Neural Turing Machines(NTM) have made significant progress setting up the foundation for building human-like memory structures in deep learning systems. In the past, I’ve written extensbily about the role of memory in artificial intelligence(AI) so I am not planning to bore you restating the same points. Instead, I would like to approach the subject from a different angle and attempt to answer three fundamental questions that we should have in mind when thinking about memory in deep learning models:
a) What makes memory such a complex subject in deep learning systems?
b) Where can we draw inspiration about memory architectures?
c)What are the main techniques used to represent memories in deep learning models?
In order to effectively answer the first two questions, we should should look at both the biological and psychological theories of memory. That should take us to the two schools of thought that have influenced our knowledge about memory the most: neuroscience and cognitive psychology. Following that same trend of thought we are going to structure this essay in three main parts. The first part will explain the neuroscience theory of memory. The second part will approach memory from the perspective of cognitive psychology while the final segment will focus on how deep learning is drawing inspiration from those disciplines to incorporate memory into neural networks. So let’s start in the place where memories are created: the human brain.
The Neuroscience Theory of Memory
Understanding how memories are created and, sometimes, destroyed as well as the differences between long and short terms memory have been an important area of neuroscience research in the last decade. One of the iconic subjects that inspired that level of research was been known as patient HM.
Henry Gustav Molaison(HM) suffered an accident at the age of nine that caused him to experience convulsions regularly for the following years. In 1952, at the age of twenty-five, HM underwent a surgery to relieve his symptoms. The procedure was considered initially successful until the doctors discovered that they have accidentally cut part of HM’s hippocampus. as a result, HM was unable to retain new memories.
The idea of living without new memories is the analog of always living in the present. Trust me, I am not talking about the mindfulness way but in the way that you can’t relate to a recent event in the past or envision an event in the future. Patient HM went about his day only retaining information for a few minutes, greeting the same people and asking the same questions over and over again. The HM case was pivotal to help neuroscientists understand how memories are created, stored and recalled.
The modern neuroscience theory of memory involves three fundamental areas of the brain: the thalamus, the prefrontal; cortex and the hippocampus. The thalamus can b considered a router that processes sensory information(vision, touch, speech) and relays is to the sensory lobes of the brain for evaluation. The evaluated information eventually reaches the prefrontal cortex where it enters our consciousness forming short term memories. The information is also sent to the hippocampus which distributes different fragments to various cortices forming long term memories. One of the biggest challenges neuroscience today is to understand how those scattered fragments of memories can be reassembled into cohesive memory experiences. This is known in neuroscience as the “binding problem”.
The Binding Problem
Considered one of the most puzzling aspects of the neuroscience theory of memory, the binding problem challenges the concept of recreating memories from other sensory information. Take the experience of going to a concert with your loved one. Memories about the event will be broken down and stored across different regions of the brain. However, it will only take one experience such as listening to a melody of the same band or seeing your wife dancing to recall the entire memory of the concept. How is this possible?
One theory that solves the binding problem states that memory fragments are linked by electromagnetic vibrations that are constantly flowing through the brain. There vibrations create a temporal(not spatial) link between memory fragments allowing them to activate together as a cohesive memory.
The neuroscience theory of memory give us the foundation to understand some of the main components of an intelligent memory architecture. However, human memory is not only a by-product of the components of the brain but it is also deeply influenced by contextual circumstances. That will be the subject of the next post.