On Fri, Nov 15, 2013 at 2:03 AM, Fergal Byrne
<[email protected]>wrote:

> Hi Mark,
>
> As I explained, time is often a datum in its own right. When learning the
> cycles of power consumption in hotgym, the time of day and day of the week
> are the most important determiners of power usage - the gym is closed at
> night and there are peaks before and after the business day as people go to
> the gym before and after work. Time in this sense is data just like any
> other kind.
>

Yes, and I argue this can be achieved as well by a sequence
ooXXXooooooXXXXX__________ , representing morning and evening queue, and
empty at night. I'm not saying exp time is wrong, I mean it isn't
biologically correct and (we should test if) it's possible to replace it
with implicit time.


>
> In NuPIC, you don't do anything between inputs. The inputs decide the
> timesteps - one per record, and NuPIC runs only when you call it with a new
> record. NuPIC learns sequences, but there's no "between" between any two
> timesteps.
>

What do you actually mean by timestep? To achieve learning, you need to
boost connections between events at the (nearly) same times, and inhibit
distant (meant on time scale) ones.  So if you touch redhot plate, and pain
arrives 1hr later, nothing is learned from this unfortunate event.  If it
hurts 1-2 secs after, you learn not to touch that again. To achive this, I
imagine there's a penalty for every "time step" (unless boosted by
co-occuring events). timestep is minimal time your sensors can distinguish
(1/24th sec for eye, 1/100 sec for neuron), a refresh rate. At each step,
connections that are both ON are boosted, other are inhibited.

>From the programming side, I said it's a hack, so I sure could decrease
permanences before computing new input.

Thanks, M.

>
> Regards,
>
> Fergal Byrne
>
> --
Marek Otahal :o)
_______________________________________________
nupic mailing list
[email protected]
http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org

Reply via email to