Neural Coding, Flash Memory Stores and Symbolic Representation

Neuroinformatics emphasizes the application of analytical tools and computational models to neuroscience data, typically using serial computers.  But a particularly powerful parallel computer, the human brain, analyzes and models data about the world, both recently acquired data and information encoded across evolutionary time.  A massive stream of newly stored data, our Flash Memory records, are consciously accessible, providing a window into a symbolic operation of the brain whereby chronologically-organized information is written into brain networks.   Introspection reveals that these day-long Daily Memory Records (DMRs) can be accessed for several days, after which an excerpt comes to reside in neocortex as enduring episodic memory.   There are myriad and profound open questions about this process, but from a computational perspective, it is important to understand the size and nature of this memory store, and how the stored items contribute to ongoing neocortical operations.  After introducing symbolic aspects of these DMRs, we consider new DMR data in the context of linguistic and non-linguistic representations of the world, with emphasis on economy of neuronal computation.  While neuroinformatics traditionally operates on data extracted from the brain (e.g. connectomes, gene expression atlases, cortical network simulations), a neuroinformatic-style investigation of internal brain operations and representations may at some point prove complementary.

Learning and the associated processing of symbols are often viewed as de novo processes (as in e.g. many machine-learning algorithms), but evolution has compacted a great deal of world-knowledge into the genetic code (Baum, 2004), including e.g. stored information about predators and prey items (Tinbergen, 1969; McElligott and O’Malley, 2005).  In the human lineage, an innate ability for symbol processing has been proposed in terms of a universal grammar and innate recognition of human speech phonemes.  While the extent to which human symbolic skills differ fundamentally from chimps is fiercely debated (Penn et al., 2008), we are certainly not dogging it.  For humans, the concept “dog” not only categorizes many items to a single 3-letter symbol, but it also entails a one-to-many calculus, where the letter sequence d-o-g can elicit recall of a huge number of memories (first pet, dog-fighting stories, etc.).  Upon seeing a dog in the street, we instantly encode the dog as a compact symbol into Flash memory, presumably without sufficient time for a vast subconscious interrogation of every memory record of ours that utilizes the concept/symbol of dog.  While the word “dog” is (at least) a compact linguistic “tag”, the Flash memory system appears to use a distinct symbolic system.

Conscious recall of our daily memory log is not an obvious place to look for the computational details of neocortical symbol processing, and certainly consciousness and introspection are fraught with “illusion” (as e.g. our visual sense of a proportionate 3D visual space), but our ability to retrieve new memory elements (bits of information) during DMR recall requires that every new bit has some physical residence within the brain.  The DMR store is interesting because it is a one-trial (zero effort) record of our sequential experiences.  Flash activation of coupled oscillators and/or silent synapses has been proposed as a mechanism for this effortless encoding (Gioioso and O’Malley, 2009), but regardless of mechanism, these DMRs capture a large quantity of linked symbols.  The quantity of symbols encoded in any given time-epoch depends upon the extent of our prior, related experiences (Chase and Simon, 1   973) indicating that our extant cortical memory stores facilitate this process—perhaps with cortical prediction (Hawkins, 2004) playing a major role.  While working memory can typically store 7 to 9 random items, DMRs are vastly greater.  Currently, a database is being constructed of normal subject DMRs and this will be extended to clinical populations, eventually using analytical tools to categorize and quantify the diverse elements that make up DMRs.

To my knowledge, no one has yet attempted to quantify the total amount of information in an entire day-long memory record, which is admittedly problematic, given the fuzzy relationship between memory records and weight-adjusted synaptic bits of information.  But DMRs do store the gist of events in the physical world, in agreement with Brady et al. (2008) who documented an almost limitless record of objects viewed.  By our collecting DMR-fragments from a large number of student subjects, it appears that DMR traces are dominated by sequential actions and observations as well as thoughts.  But specific linguistic (word) sequences are generally not recorded (excepting particularly poignant phrases).   This implies that we have 2 parallel symbolic encoding schemes: a (more right-brained?) time-and-space scheme, which operates in concert with a second linguistic system.  But only the older, non-linguistic system is capable of flash writing of day-long records.  These records are likely important because evolution trends towards optimization and economy, and because the DMR items reflect some conjunction of novelty and/or salience (trivial as they might seem).  Neocortex stores these day-long sequences of symbols, perhaps by compressing the information in some way.  But, the degree to which a neuronal symbol (invariant representation or “neural word”) can be compact, while embodied in a 10,000-fold connected network, distributed across some unknown number of neurons, and realized in a fast-paced neurodynamic environment, is rather unclear.  While this “economy of storage” is the most speculative aspect of the DMR symbolic system, any valid accounting of human brain operations (from fMRI to computational models) must be consistent with our easily-accessed Flash memories.