On Hawkins, Hierarchy, Homunculi and the Hippocampus

On Intelligence

As previously noted, Jeff Hawkins takes a very similar approach to Karl Friston’s ‘variational Free Energy’ with the concept of ‘hierarchical message passing’ playing a key role in both. In such hierarchical message passing, it is clear how levels in the middle of the hierarchy and at the bottom perform:

  • Those in the middle of the hierarchy interact with layers both above and below.
  • The layer at the bottom of the hierarchy interacts with the outside world.

But, what happens at the top? The highest level has nowhere to feed prediction errors to or receive predictions from. This is what a previous blog entry  asked.  A top level would be the ultimate decider of our actions, with nothing to defer to – a single centre of control. It is like the last ‘little man inside our heads, in the Homunculus Argument . It suggests that there is a single small region in the brain that is the coherent seat of the what Daniel Dennett derisively calls the ‘Cartesian Theatre’.


In order to avoid falling into this pit, I postulated that there was in fact no top; that this was just an artifact of the gross simplification in presenting a simple hierarchy, with each level communicating to one and only one region in the level above and precisely one region in the level below. But if they can communicates upwards and downwards with more than one region (as suggested in Friston’s diagram, right) it becomes possible for the hierarchy to loop back on itself, with regions feeding upwards to levels that were also below below it, and vice versa. The clean hierarchy therefore breaks down, but the basic concept of the operation of each block dealing with predictions and prediction errors remains intact.


Yet Felleman and Van Essen’s famous diagram of the visual processing of a Macaque monkey (right) puts the hippocampus at the top (marked ‘HC’). At the bottom of the hierarchy, signals from the Retinal Ganglion Cells (‘RGC’) within the eye pass through the Lateral Geniculate Nucleus (‘LGN’) in the Thalamus to the primary visual region, ‘V1’, and thence up to the Entorhinal Cortex (‘ER’) and finally the hippocampus.

And in ‘On Intelligence’ , ‘On Intelligence’ (2005) Hawkins explicitly maintains that the hippocampus is at the top of the hierarchy and provides a very nice account of learning and  relearning across levels of the hierarchy. (The bold emphasis is mine)…

“When you are born your cortex essentially doesn’t know anything.…you need outside instruction to help you decide which patterns belong together. …your brain slowly builds sequences of patterns that belong together.” (page 165)

“The unexpected result of the learning process is that, during repetitive learning, representations of objects move down the cortical hierarchy. During the early years of your life, your memories of the world first form in higher regions of cortex, but as you learn they are re-formed in lower and lower parts of the cortical hierarchyIt isn’t that the brain moves them; it has to relearn them over and over. (I am not suggesting that all memories start at the top. … It is the memory of sequences I am suggesting re-form lower and lower in the cortex …). … Consider how we learn to read. The first thing we learn is to recognize individual printed letters. This is a slow and difficult task requiring conscious effort. Then we move onto recognizing simple words … After years of practice, a person can read quickly. It isn’t just that we are faster; we are actually recognizing words and phrases as entities. … A young brain is slower to recognize inputs … because the memories used in these tasks are higher up the cortical hierarchy. Information has to flow all the way up and down, maybe with multiple passes, to resolve conflicts.” (pages 166-167)

“Experts and geniuses have brains that see structure of structure and patterns of patterns beyond what others do. You can become expert by practice.” (page 168)

“Under the neocortical sheet: basal ganglia, cerebellum and hippocampus. All three existed prior to the neocortex. …To some extent the neocortex has subsumed their original functions. For example, a human born without much of a cerebellum will have deficits in timing … but otherwise will be pretty normal. The hippocampus, however, is a different beast … it is essential for the formation of new memories. If you lose both halves of the Hippocampus … you lose the ability to form most new memories.” (pages 168-169)

“For many years …it didn’t make sense to me. The classic view of the hippocampus is that new memories are formed there, and later, over a period of days, weeks or months, these new memories are transferred to the neocortex.” (page 169)

“…the connections between the hippocampus and the neocortex suggest that the hippocampus is the top region of the neocortex. ….If a region doesn’t understand the current input, it passes it up the hierarchy until some higher region does. … when you get to the top of the cortical pyramid, what you have left is information that can’t be understood by previous experience.” (pages 170-171)

“It is these unexplained remainders, the new stuff, that enter the hippocampus and are stored there. This information won’t be stored forever. Either it will be transferred down into the cortex below or it will eventually be lost.” (page 171)

“I have seen so many plays in my life that I rarely do I see anything truly new. New plays fit into memories of past plays, and the information just doesn’t make it to the hippocampus. For my children, each play is more novel and does reach the hippocampus. … the more you know the less you remember. (page 171)


The Hippocampus

This entry was posted in Uncategorized and tagged , , , , , , , . Bookmark the permalink.

4 Responses to On Hawkins, Hierarchy, Homunculi and the Hippocampus

  1. headbirths says:

    Felleman and Van Essen diagram in layers from top to bottom, left to right are:
    1.HC = Hippocampus
    2.ER = Entorhinal Cortex
    3.46 / 35 / 36 / TF / TH
    4.STPa / AITd / AITv
    5.7b / 7a / FEF / STPp / CITd / CITv
    6.VIP= Ventral Intraparietal / LIP = lateral intraparietal / MSTd = medial superior temporal (dorsal) / MSTI = medial superior temporal / FST = floor of superior temporal / PITd (dorsal) / PITv (ventral)
    7.DP / VOT = ventral occipitotemporal
    8.MDP = medial dorsal parietal / MIP = medial intraparietal / PO = parietal-occipital / MT = Middle Temporal (V5) / V4t / V4
    9.PIP = posterior intraparietal / V3A
    10.V3 / VP = ventral posterior
    11.M / V2 / P-B / P-I
    12.M / V1 / P-B / P-I
    13.M / P (LGN = Lateral Geniculate Nucleus in the Thalamus)
    14.M / P (RGC = Retinal Ganglion Cells)

  2. Don’t you see a contradiction here: cognitive hierarchy is supposed to select for novelty & generality at the same time? Isn’t it obvious that they’re mutually exclusive: general (pattern) is what repeats, & novel is what doesn’t? There’s nothing more novel than random noise.
    I think the hierarchy must select for additive generality: patterns recognized on lower levels, but not expected by higher levels. Which means that hippocampus is not the top, it’s just another relatively primitive mediator, like thalamus. Anyway, the brain is a kludge, I’d rather stay on conceptual level: http://www.cognitivealgorithm.info .

    • headbirths says:

      Hi Boris,

      Thanks for the link to your website – I’d like to give it due attention sometime.
      I agree with you: the hierarchy must select for additive generality. I’m not clear what you mean by ‘selection for novelty’ in Hawkins’s work. I think you are taking exception to Hawkins’s statement
      “It is these unexplained remainders, the new stuff, that enter the hippocampus and are stored there.”
      As though all the novelty gets stored there, in a ‘short term memory’ FIFO.

      As I see it:
      1. Each level tries to pattern-match as far as possible within its limited capacity and input sources.
      2. Levels higher up the cortex have inputs from a wider range of input sources so have the potential to find patterns not possible to lower levels.
      But patterns matched at the lower levels will not be ‘visible’ to higher levels;
      the higher levels only ‘see’ the noise that the lower levels failed to match.
      3. A ‘top’ level can only pattern-match in a similar fashion to the other levels.
      Any novelty/noise in patterns input at this level that is not matched will remain unpredictable.
      4. However, the top level may behave *slightly* differently.
      The hippocampus is physiologically different from normal cortical regions.
      But it would not be radically different.

      Now, to accommodate Hawkins’s ideas…

      5. It is not the case that representations of objects actually “move down the cortical hierarchy”.
      It is that the lower levels require more repetitive stimulus for them to learn the pattern.
      (The information to all levels is always coming in from sense input at the bottom of the hierarchy.)
      They are less ‘volatile’/’dynamic’/’impressionable’ than upper levels.
      I do not know the appropriate terminology here. In a simple filter Y[t] = (1-k)Y[t-1]+k.X[t], a high k, nearing 1, is ‘volatile’/’impressionable’ and a low k, near 0, is not.
      Over time, the higher levels ‘forget’ previous short-term adaptation as it has been replaced by more recent adaptation.
      6. There needn’t be a radical increase in k from lower levels up to higher. But I don’t of any physiological explanation of varying k.
      7. Contrary to 5, there *could* be a mechanism which allows lower levels to be adapted on the basis of higher level adaptation, i.e. something that takes place in a ‘sleep’ mode.
      8. Anything we say about the brain is a gross simplification but to say that the hippocampus is at the top still seems to go a bit too far. I’m prepared to be persuaded that this is a useful conceptual generalization though.

      So its a case of varying degrees of selection for generality rather than selection for generality and/or novelty.

      All speculation on my part from a rather low knowledge base.

      • Hi Andreas,

        >I’m not clear what you mean by ‘selection for novelty’ in Hawkins’s work. I think you are taking exception to Hawkins’s statement “It is these unexplained remainders, the new stuff, that enter the hippocampus and are stored there.”

        That, & many other statements to the effect that only novel inputs, those that don’t match on a given level, are forwarded to the next level. Which you agree with.
        This precludes selection for generality, – the only way you know that inputs are generally applicable is if they *do* match prior experience on the way up. So, you select for noise, which makes no sense to me.

        I think the confusion here is between generality / pattern discovered on a given level, which is “positive” for forwarding thus-compressed inputs upward, & negative or inhibitory feedback that a discovered pattern sends to a source level. It is negative because target (higher) level can now predict the source level, which means that its outputs are less “additive” to the former. But this feedback is location-specific, rather than output-specific. Relatively stronger output patterns are still selected over the weaker ones, it’s just that common threshold is raised for the whole location.

        So, I agree with you & Van Essen (& disagree with Hawkins) that this feedback is modulatory rather that representative.

        Hippocampus (primitive “reptilian” cortex) fits into this scheme because it represents locations. Value of any prediction is precision of “what” * precision of “where”, one is useless without another. There is a dorsal “where” pathway in neocortex itself, but I think hippocampus is privileged because it’s densely interconnected with amygdala & other “emotional” areas.
        Basically, an animal only cares about things / memories conditioned by their association with emotionally charged places, those previously or currently close to itself. And we are still animals, believe it or not.

        None of which has anything to do with cortical hierarchy, from primary to association areas: http://www.scholarpedia.org/article/Cortical_memory.

        Again, our brain is a mess, on all levels, from the synapse up.
        So, my approach is strictly functional, neuroscience is only a general inspiration: http://www.cognitivealgorithm.info.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s