Hello_sp.py producing identical SDR outputs despite different, noisy inputs

Engineering NuPIC
1

Hi

I’ve been having problems with my SDR outputs so decided to go back to the hello_sp.py demo to validate whether everything is working as it should.

Upon adding noise to the input array, one should expect the output SDR to be different. The demo states “Adding another 20% noise to the already modified input vector. The output SDR should differ considerably from that of the previous output”

However, despite verifying that the input array has indeed swapped bits when the addNoise method is called, the SDR outputs are identical.

Even when increasing noise by 50%, the SDRs share 78 of 81 active bits (almost identical). In this latter example, the two input arrays share 379 matching active bits of 529/531 active bits respectively (input arrays are size 1024), so I’d expect the SDR difference to be higher?

Anyone know why I might be experiencing this problem?

Thanks

2

I don’t understand why it is a problem. This is what you want. It means you have properly encoded semantics in the SDR. Adding noise should not affect the output SDRs.

3

This seems to center on how we define ‘considerable’ in the difference in SDR’s. Adding more noise should create more difference between SDR’s, though it takes a fair bit of noise to create any difference at all.

2 Likes
4

@momentum Do you understand why noise resistance in the SP is so important? It’s not a bug, it’s a feature! (A really important one)