7

Both SPA (Eliasmith et al.) and predictive coding (Friston, Clark, Rao, et al.) seem to have a lot of explanatory power.

My understanding of SPA is probably rudimentary, but Eliasmith and others seem to be able to explain and simulate higher-order psychological constructs (e.g., emotions, concepts) impressively well. Similarly, predictive coding (e.g., Clark, 2013)--if correct--is a very convincing account of the fundamental computations of the brain. And there's a lot of emerging evidence to support it.

So I'm curious whether these frameworks can be integrated. Would Spaun perform better by implementing some sort of predictive coding algorithm(s)? Or does SPA make this unnecessary? These questions may be poor due to my ignorance about the complexities of SPA and NEF, but any insight would be great.

Seanny123
  • 8,853
  • 3
  • 25
  • 61
mrt
  • 4,388
  • 1
  • 15
  • 36
  • I know a lot about the NEF/SPA and I'm trying to read the predictive coding paper you linked. However, would you like to create a summary of the key points of the paper, so that I can answer the question more quickly and to your greater satisfaction? – Seanny123 Mar 06 '16 at 23:34
  • @Seanny123 Summarizing predictive coding isn't an easy feat! ;) But a super basic outline is: (1) your brain is constantly launching predictions about incoming sensory input, (2) correct predictions "explain away" sensory input, and (3) incorrect predictions result in prediction error, which flows up to modify your future predictions (i.e., learning). Under this account, attention corresponds to modulating the gain of prediction error units. So PC runs counter to the traditional stimulus-organism-response model, where we assume the brain is "inactive" until a stimulus causes it to react. – mrt Mar 06 '16 at 23:49
  • @Seanny123 Instead the brain is constantly anticipating incoming sensory input, both from the world (exteroceptive) and the body (interoceptive). My favorite hypothesis generated by this model is that the affective feelings we have (of pleasantness and arousal) are mostly interoceptive predictions and not incoming interoceptive input. This is because our interoceptive predictions are explaining away that information (unless there's prediction error). So your feelings are largely generated by predictions about sensory input and not the sensory input itself! – mrt Mar 06 '16 at 23:52

1 Answers1

2

To begin to answer this question, we must first unpack the concepts in their current context.

The NEF makes no prediction about how error is propagated in the brain. It explains how to do computation using vectors in spiking neural networks. Also, it defines how error signals can be used to change how the signal is encoded (take in) and decoded (sent out) from the neural population, as well as perform probabilistic computations. For further information on this, see PES, BCM and VOja, as well as Chapter 9 of the NEF textbook.

The SPA also makes no claims how error is propagated throughout the brain. It only claims that compression must occur for scalable and grounded knowledge representation and defines some operations for accomplishing this.

Predictive coding appears to be (based off the paper you linked) about defining how the error is propagating. Specifically, errors go up with higher level impressions adjusting how errors go up the hierarchy based on predictions/impressions.

As hinted at previously, this is somewhat separate from the NEF/SPA, however this does not imply they can't be integrated, nor does it mean that there doesn't already exist some models that fulfill the predictive coding criteria other than simply unifying perception and action by using similar representations (I think the DeWolf's REACH control hierarchy might fall under this category). It just means that as of now, there have been no explicit relations between the two modelling philosophies. To create a stronger conclusion, I would need to understand the implementation details of Predictive Coding to see if they are biologically plausible or if they can't be formulated as vector based operations within the SPA. This seems to be the position that my colleagues share (they actually have an comment article appended to the one you linked called "God, the devil, and the details: Fleshing out the predictive processing framework") which is that more work needs to be done before a position can be taken.

Seanny123
  • 8,853
  • 3
  • 25
  • 61
  • Interesting! I was in the middle of reading Eliasmith's book when I asked this question (and I still have to finish it one of these days… ;)), but I'm definitely ignorant about the details of his work (and of computational neuroscience in general). But I think your response, as it's currently written, definitely makes clear the fundamental differences between the frameworks--which is very helpful. – mrt Mar 07 '16 at 00:43