How would you interpret Multiple active cells in a minicolumn if you are using it as OUPTUT of TM ?
In a sense you have to generate a SDR. Normally let say I have 2000 minicols I will generate 2000bit SDR where in the column I have 1 I will set this bit in the SDR to 1 … but what if I have more than 1 bit active.
I know this is not expected behavior. I’m asking how will interpret it !?
any ideas …
If a mini column has more than 1 cell active (without bursting), it basically means that feature was correctly predicted – but is ambiguous about which sequence it recognizes itself within.
So let’s say you had a TM which learned from the sequence:
If you then fed in --> B,C, 2 sets of cells should activate for C – those from the A,B path and those from the X,B path – since it doesn’t have enough context to rule out either one. In this case the system would predict both D and Y.
However if you fed in --> A,B,C then only one cell per C-column should activate, since it has the context to narrow down which path from B it is within, and thus predict only D.
So this behavior actually should be expected in systems which have recurring subsequences.
for this to happen D and Y have to overlap, right ?
What do you exactly mean by “overlap”, can you elaborate?
However, it is partial overlap.
Do you mean the SDR of D and Y should (partially) overlap? Isn’t it irrelevant, no?
By @sheiser1’s example, it’s the minicolumns of C that have multiple active cells.
Maybe I’m missing something.
the SDR’s of the two symbols, have to partially overlap
If the HTM network predicts D and Y at the same, and by coincidence x% of the active mini-columns are overlapped, you may have mini-columns with two predictive cells.
They could overlap by chance but I don’t think that’s required. You would have two predicitive/active cells in C no matter if they overlap in the example.
But come to think of it, you would have two predicitve cells in the intersect of D and Y.
@sheiser1 said multiple active cells in C so I guess I didn’t think of D and Y that well. I’m sorry.
For this categorical example there’d be no overlap in the columns representing D & Y – but they’d both be predicted because they’d both be connected to cells in C-columns. So in a different sense they do overlap, in time, in terms of which values precede them.
In this case you could of the connections as:
D’ cells link to C’, which link to B’, which link to A’
Y’ cells link to C’’, which link to B’’, which link to X’
With these distal segments in place:
Inputing --> B,C to the TM would activate B’ and B’’, predicting C’ and C’’, which would then predict both D’ and Y’.
Inputting --> A,B,C to the TM would activate A’, B’ and C’, which would then predict only D’