Accidental Memory

There is memory. And then there is memory.  It can refer loosely to the ability to remember the present as it becomes the past.  It can refer precisely to temporal filling in based on executive attention pattern retrieval.  It can be an unfortunate misnomer that means something else entirely.  What exactly is memory?

Psychological memory entails encoding, storage, and retrieval.  The Atkinson-Shiffrin (1968) model breaks memory into sensory, short-term, and long-term forms.  Sensory memory provides a “buffer” for incoming sensory stimuli that lasts in the range of seconds.  It includes vision (about 1 second buffer) and hearing (about 20 second buffer) and presumably also touch, taste, and olfactory.  The concept is that people may sometimes visually glance at a scene, wait a second, then do a double-take as they belatedly process the image from the buffer.  Or a person can be so engrossed in a movie that they may not register a verbal greeting from their spouse until moments after the fact, whereupon said spouse’s very long term memory exhibits itself.  Rehearsing or repeating a memory stores it in the short-term storage, and continued rehearsal encodes it in long-term storage.  Capacity, where it has meaning, is on the order of a half-dozen “chunks.”

Baddeley and Hitch (1974) focused on a working memory that actively processes stimuli into a meaningful pattern.   Again the emphasis is on acting on the material memory with regards to comprehension, integration, and learning.  The identified brain portions primarily include the frontal cortex and executive processing region.  A memory is nothing unless translated through the lens of actions, rehearsal, or processing.

Further subdivisions in memory follow from neuroscience principles referring to procedural and episodic memories.  Procedural relates to primarily motor or muscle memory, such as remembering how to walk or how to properly perform a tennis backhand smash.  This form typically requires much repetition and rehearsal.  Episodic memory relates primarily to event-based scene memory, such as where we were walking when a scary vicious Dobermann burst out of the neighbor’s yard.  This form typically does not require repetition at all – who in their right mind would need to be reminded multiple times of the location to avoid when walking?

Computer memory provides a safely operable and concrete mechanism in contrast.   Data enters via a peripheral and may be transferred via the data bus into long term storage (a hard drive, today measured in Terrabytes), short-term storage (RAM, today measured in Gigabytes) and really short term working memory (registers and cache, today measured in Megabytes).  The data bits of memory are passive, behaving as targets of the instruction set.  They are read, transferred, added to, and placed wherever the code tells the hardware to do.  It is important to note that computer memory uses the nomenclature of psychological memory (e.g. short term, long term, fast, slow) but functions more like a filing system that serves a completely different purpose.  A computer is, after all, a compute-er – something that computes, adds, moves, and stores numbers. 

The crucial point to note is that these different usages and disciplines force the expert user to define and think about their respective memory in fundamentally different ways.  The Atkinson-Shiffrin model follower, for example will think of memory in terms of the different senses and durations.  If visual memory is different from audio memory, then vision is necessarily a different phenomenon than audition.  They have different mechanisms, speak different languages, and perform non-overlapping tasks.  The same follows with short term and long term memory – there must necessarily be different parts of the brain to store different forms of memory.  The challenge becomes one of identifying and classifying what goes where. 

Baddeley and Hitch followers are forced to focus on the active working term.  If memory needs work to operate, what is the work?  What is the intent?  What is the person attending to at the time of the event in question occurs so as to create a memory?  The challenge becomes one of exploring the person’s intent.

The neuroscience follower with respect to episodic vs procedural is forced to examine which neural model fits and replicates which form of memory.  Clearly deep learning neural networks fit the procedural form by requiring repetition to stably organize the patterns into a distributed network while minimizing crosstalk.  What, however, can best fit an episodic memory function?  Further, since both forms are simply sides of the same coin, how do these memory forms fit into a unified brain?  Physics has the challenge of seeking the Grand Unified Theory and ultimately uniting the rules of the very small subatomic with the rules of the very large and gravity.  Neuroscience will eventually need to address how to fit episodic and procedural into the same neuronal wetware.  

Computer memory practitioners are forced to approach the issue as one of filing efficiency.  How does one organize the memory data store into a database?  How many data fields/features are there and how much space should one allocate for each?  How does one make accessing the memory data faster and more efficient, such as with the cache or with parallel partitioning?  The challenge becomes one of computational complexity and efficient algorithmic design.  

Perhaps everyone knows the answer – more or less – but nobody can explain it.  Perhaps every different discipline teaches us to view memory through its idiosyncratic lens for its own purposes.  Like a mosaic, every definition of memory is canonically correct, but locally limited and blind to the alternatives.  Each is a game where players willingly participate for the fun factor but unwittingly adapt themselves to the accidental idiosyncrasies of each.  Each definition of memory is an accident.

Think of a simplified card game form, “Deuces,” where the four players each have a pool of 24 cards.  At each turn, the player is faced with a decision point and must select from their pool of cards to submit a decision hand consisting of one or more cards.  Similarly, the memory game played would be played like this: each of us has, say, 100 billion memories.  At each moment in time we are faced with a decision point and must select from among the pool of 100 billion memories to inform a decision action.  As in “Deuces,” the devil is in the details.

How does one store 100 billion memories in a computer memory model and how does one subsequently find and extract one single memory file?  The computer scientists and engineers have an answer for that one, and in O(log(n)) time or less.  But what is the size of a memory?  Is it a named video file or audio file?  Is it a second, a minute, a book, or a book summary?  Is it a feeling?  Is it a story?  How much space does the memory of a spouse take?  How about the extended family?  In-laws?  How does one find, extract, and combine multiple memories?  The answer to these is that no discussion of memory can be complete without a psychological discussion on selective and executive attention.  The mosaic only begins to make sense from afar, by crossing disciplines into a bigger picture.

Following this line of exploration, perhaps we need a new definition of memory – a definition that breaks from our legacy definitions of the past and attempts to consolidate the meaning of memory anew.  The legacy concept of memory is one of passively recording events for future recall, as with written documents; it almost avoids the obvious follow on question of, “for what and by whom?”  A new definition of memory would be one of actively imprinting our habit of actions to form our behaviors.  It is a temporal filling in that connects our past, present, and future to constrain our decision action space.  Our past actions guide our future actions.  Every action in the past is training for the future and present.  It is a fully recursive definition by necessity.

There is a wonderful scene in the movie, “Catch Me If You Can,” where Leonardo DiCaprio’s character slips by unnoticed to escape legions of police agents looking for him.  He slips by in plain sight wearing his fake, “Look at me!” pilot’s uniform because he surrounded himself with beautiful flight attendants.  All the agents were too busy choosing to gaze at the beautiful flight attendants to notice their quarry walking casually not 10 feet away in full view.  None of the agents can recall seeing Leonardo DiCaprio’s character.  What does this mean?  It means memory is a choice; the agents chose to remember the beautiful attendants but not their work.  Their past decisions – to look at beautiful girls – guided and constrained their current decisions – to look at beautiful girls – and formed their subsequent memory – of looking at beautiful girls.  They will then talk about the beautiful girls afterwards.  This causes them to assign a higher priority (i.e. priming, or holding on the surface short term memory in legacy parlance) to beautiful girls such that beautiful girls command a higher salience, which in turn causes their eyes to inevitably saccade towards beautiful girls.  They cannot remember that which they did not even see.   “Sorry, sir.  I was not paying attention.”

Attention and memory are not just linked; attention and memory may be the same.  There cannot be memory without selective attention to focus our senses.  There cannot be attention without the temporal filling in and tracking effect of memory to guide our executive attention.  This definition approach would have profound effects on modeling episodic and procedural memory –as they may then become the same phenomenon on different scales.  This definition would have profound effects on even defining intelligence.  The effects may go on and cascade throughout all forms of decision making.  If so, then the legacy definitions of memory are indeed accidents, but happy ones that can lead us to an even bigger, happy accident of memory.