Here is my conjecture : Humans, brains, neuro-cortex can predict only one step ahead.
It could be 1s ahead, or 1min or 1month …or…!
Could you guys come up with examples that contradict or confirm that ?
I don’t deny that it is possible to predict f.e. both 1s ahead and 1.5s ahead, but in this case it is still two times ONE step ahead.
The trick is that the two (or more predictions) are based on different CONTEXTS i.e. because it is a Memory instead of Computation we can have “infinite” contexts, but “time-series-like” prediction will require time for cmputation which is at a premium.
Ex. : You can’t imagine the steps to go from home to work simultaneously, but you can easily follow your path if you take it one at a time.
Think hard about it because if that is not the case then TM should be redesigned
It’s an interesting argument, especially because it also conjures up questions about attention.
But here are two seperate issues I see with the statement.
First, when a sequence of for example 5 steps gets represented in a union SDR after 3 steps, the two last steps have essentially been predicted simultaneously as proven by the representation of it in the SDR.
Second, there would be a problem if you look at what happens over different layers in a hierarchy simultaneously. Maybe at the lowest level, a prediction of 1/3 of a second ahead is happening, while, because of a sequence of union SDRs, there is a prediction of 1 second ahead one level higher.
But you can go through the steps mentally without actually doing it. Does that mean that you can somehow emulate the input sequence in a recursive way i.e. feeding the prediction from one step as input for the next step?
Could it be that the higher level prediction is a prediction of context for the lower level such that when the input arrives, the high level context prediction “collapses” into the appropriate sequence of predictions? For example, the 1 second prediction on the higher level would actually be represented by 3 recursive low level predictions each 1/3 of a second.
Hmm, I would say no. I do not think it is necessary for the lowest level to have predictions forced upon it for two reasons.
First, since the meaning is already represented higher up, there’s no clear point to forcing its complete detailed representation down below. The goal is interpretation, and acting upon that interpretation. Thus once meaning is represented, it can be used for action, for example, motor commands. There should be no cause to have the detailed representation anymore, unless you see some added value there that is lost on me.
In your view, would the “collapse” influence union SDRs, or actual columns and cells ?
Second, when going higher up the hierarchy, abstraction should take place. Abstract representations should not be able to predict detailed representations. If that was the case, there is no use for abstraction, because there would be none. For example, there are many ways the word “manifestation” can be exactly pronounced. Would having a representation of one ‘perfect, exact’, pronunciation be beneficial ?
To solve that problem, Mr Hawkins suggested (I’m liberally interpreting here) that unexpected signals travel up the hierarchy (does he mean that attention shifts downward?) until they hit a level of abstraction where they make sense, or grab our attention. Traveling up the hierarchy might be easier than going down.
I think imagination is a tough and fascinating problem, and I would like to speculate on this example. There are two mechanisms that might work together to help explain it. The first is our ability to not execute commands, to inhibit them. The second is our ability to predict stimuli according to our own behavior. In the example of imagining our walk home, the signals that allow us to predict what we would observe when we were actually walking home, might be let through, while the motor commands for doing it are inhibited. This way, the next step in the sequence can be executed imaginatively.
On the other hand, there is so much that we can imagine, and this explanation does not hold up for all types and things that we can see in our mind’s eye (or ear, or whatever), so there is probably a larger principle that allows for imagination and includes an explanation for this type of imagining.
There are many things we could think about, like attention serving as some kind of amplifier, lowering the threshold for firing or activating predicted cells. I read once about attention being manifested in the brain with a wave of signals in a region of the brain. Maybe instead of lowering the threshold, these waves give a ‘premium’ to cells in a region to fire, so sense can be derived. Since once a cell has fired, continuous input has less and less of an effect, it might be possible to go over different predictions, from most to least likely, by keeping up that attention on the region. The most likely interpretation would fire first, but die off quickly, less likely ones would only fire later, because they needed more input.
It’s a hard puzzle and the last word has not been said about it.
I am not sure. What I meant is that from a predicted union SDR, only a small subset would be selected. But I don’t know how that would actually happen, I assume feedback from the higher region into the lower region?
I do not think it is necessarily trailing. Part of a sequence should be enough for the union to become active, thus predicting more than one step (by implication in the representation of the full sequence).
True. That was not a constraint in the original statement. But if it is included, it does invalidate my second argument.
In my opinion, the first concern still stands, unless you can explain how and why the union would trail the sequence.
You seem to be assuming that the union-SDR represent distinctive steps ?
I seem to be assuming that it is more like flow, where union-SDR bits just go on and off over time (unless abrupt change), compared to the initial layer where the active bits are different almost every step. Thats why I used “trailing” to express it.
still thinking on it.! Don’t have yet full understanding what happens beside SP&TM, so I’m somewhat in the dark here.
