[HN Gopher] Space is a latent sequence: A theory of the hippocampus
       ___________________________________________________________________
        
       Space is a latent sequence: A theory of the hippocampus
        
       Author : XzetaU8
       Score  : 54 points
       Date   : 2024-08-01 18:28 UTC (4 hours ago)
        
 (HTM) web link (www.science.org)
 (TXT) w3m dump (www.science.org)
        
       | smokel wrote:
       | Space is a latent sequence _to humans_. I wonder what it is to
       | God, or some other external observer.
        
         | protonfish wrote:
         | The paper is specifically about how space is represented in the
         | human brain. Data representing space can be encoded in many
         | different, but functionally equivalent, ways. Showing how it
         | could be encoded by neurons in the hippocampus is valuable to
         | understand our brain, but philosophically not significant.
        
           | ithkuil wrote:
           | How can understanding how the world works not be
           | philosophically significant?
           | 
           | How did we come to the point where we relegated philosophy to
           | be the study of only the things not connected with reality?
           | 
           | I'm fine with thinking about philosophy as a field that also
           | explores ideas that are not connected with reality, but it's
           | not only about those things.
        
             | meroes wrote:
             | They got sick of philosophers asking tough questions
             | probably.
        
         | 0xTJ wrote:
         | Philosophy isn't relevant here, this is talking about science
         | and theorizing about the function of real structures in our
         | brains.
        
         | exe34 wrote:
         | So like Narnia looking out of the cupboard?
        
       | fredgrott wrote:
       | If you read Wolfram's work you see him talking about this several
       | times...
        
       | observationist wrote:
       | This is coming from Dileep George and the Jeff Hawkins adjacent
       | theories around hierarchical temporal memory, cortical columns,
       | and much of the higher level theory around what it is that the
       | brain is doing in its smallest, repeated functional units. They
       | split some time back, but they're both quite rigorous and have
       | done exciting research. This paper goes over allocentric framing
       | of learning, allowing cortical networks to be built through
       | learning and thinking over time, driving a more complex and
       | nuanced model of the world to be built throughout the networks in
       | the brain, where the hippocampus seems to hold the long term
       | memory, with connections that feed back up to nearly every region
       | in the cortex. It's not complete, but it abstracts a level away
       | from phenomenological observations like the notion of Jennifer
       | Aniston cells, or mirror neurons, or things that occur as a
       | consequence of some underlying functionality. We're getting much
       | closer to a complete picture of the algorithms underlying human
       | intelligence, and those may unlock human level machine
       | intelligence, the understanding of consciousness, and all sorts
       | of great medicine and technology. Google made a good choice in
       | hiring Dileep George.
        
         | K0SM0S wrote:
         | Agreed.
         | 
         | Anecdotal (but deep) research led me to postulate that our
         | entire "inner world", for lack of a better word, is an emergent
         | construction based on a fundamentally _spatiotemporal encoding_
         | of the external world. This assumes that feeding and motility,
         | i.e., a geometric interpretation of the external world, is
         | among the first  'functions' of living organisms in the
         | evolutionary order. They subsequently became foundational for
         | neuronal systems when these appeared about 500 million years
         | ago.
         | 
         | The hypothesis was informed by language notably, where most
         | things are defined in spatial terms and concepts (temporal too,
         | though more rarely), as if physical experiences of the world
         | were the building blocks of thinking, really. A _" high"_
         | council, a _" sub"_ culture, a _" cover",_ an _" adjacent"_
         | concept, a _" bigger"_ love, a "convoluted" or "twisted" idea,
         | etc.
         | 
         | Representations in one's inner world are all about shape,
         | position, and movement of things in some abstract space of
         | sorts.
         | 
         | This is exactly how I'd use a 4D modeling engine to express a
         | more 'Turing-complete' language, a more comprehensive
         | experience (beyond movement: senses, intuitions, emotions,
         | thoughts, beliefs...): use its base elements as a generator set
         | to express more complex objects through composition in larger
         | and/or higher-dim space. Could nature, Evolution, have done
         | just that? Iteratively as it conferred survival advantages to
         | these genes? What would that look like for each layer of
         | development of neuronal--and later centralized "brain"--
         | systems?
         | 
         | Think as in geometric algebra, maybe; e.g., think how the
         | metric of a Clifford algebra may simply express valence or
         | modality, for those neuronal patterns to trigger the proper
         | neurotransmitters. In biological brains, we've already observed
         | neural graphs up to 11 dimensions (with a bimodal distribution
         | peak around ~2.5D and ~3.8D iirc... Interestingly for sure,
         | right within the spatiotemporal ballpark, seeing as we
         | experience the spatial world in 2.5D more than 3, unlike fishes
         | or birds).
         | 
         | Jeff Hawkins indeed strongly shaped my curiosity, notably in "A
         | Thousand Brains" and subsequent interviews. The paper here
         | immediately struck me as very salient to that part of my
         | philosophical and ML research--so kinda not too surprised
         | there's history there.
         | 
         | And I'm really going off on a tangent here, but I'm pretty sure
         | the "tokenization problem" (as expressed by e.g. Karpathy) may
         | eventually be better solved using a spatiotemporal
         | characterization of the world. Possibly much closer to real-
         | life language in biological brains, for the above reasons.
         | Video pretraining of truly multimodal models may constitute a
         | breakthrough in that regard, perhaps to synthesize or identify
         | the "ideal" text divisions, a better generator set for (any)
         | language.
        
       ___________________________________________________________________
       (page generated 2024-08-01 23:00 UTC)