[mythtv] Thoughts on new commercial detection option

[Kyle Schlansker]

> Do you have any ideas on which ANN architecture to use? 

That is an open question.  I've worked with many architecture's, but
the one that has proved most useful (for my applications) has simply
been your standard 3-layer feed-forward network.

> I guess that a ordinary feed-forward packprop with a sliding window will miss the
> long-distance relationships (such as how long a commercial break usually
> is), but may be accurate enough. I imagine that one use a seperate ANN
> for each channel, and a way to share ANNs with each other. 

This completely depends on what you give the ANN for inputs.  Since
commercial flagging is not real-time, we have the advantage of being
able to calculate useful statistics for inputs to an ANN.

I see you'r reasoning for having a separate ANN for each channel, but
why not just make the channel number an input to a single ANN?  ANN
are really good at picking up on things like that.  Of course, to
begin with I'd suggest just picking a single channel to work with and
see if the idea is even feasible.

I've written a program that uses Particle Swarm Optimization to select
inputs (from a database of inputs we think might be useful) for an ANN
as well as train it.  I'm in the progress of extending it to also
adapt the network architecture.

>Another challenge is to integrate the ANN into the GUI, the ANN needs user
> feedback to be able to learn.

Yes and no.  If we can soundly say that all user input will be valid,
then this is true. What you don't want is to have some ignorant user
feeding invalid inputs to the ANN (it can be devastating in the
learning process). However, if there is a way we can pre-train the ANN
then no learning will be necessary in the final code, and no user
input would be required.

-- Kyle S.


On Fri, 10 Dec 2004 11:17:46 +0100, Simon Lundell <simon at ida.his.se> wrote:
> 
> 
> > > Hi!
> > >
> > > Would it be possible to dump this data to a file? I would really
> like
> > > to test how an ANN performs on this kind of data.
> 
> > I'm extremely interested in this as well.  Over the last few months
> I've had many ideas for using ANN to flag
> > commercials.  I've done a lot of ANN programming so I would be more
> than willing to contribute code if anyone is
> > interested in following up on this idea.
> 
> Do you have any ideas on which ANN architecture to use? I guess that a
> ordinary feed-forward packprop with a sliding window will miss the
> long-distance relationships (such as how long a commercial break usually
> is), but may be accurate enough. I imagine that one use a seperate ANN
> for each channel, and a way to share ANNs with each other. Another
> challange is to integrate the ANN into the GUI, the ANN needs user
> feedback to be able to learn.
> 
> //Simon
> 
>