A very simple-minded way of trying to identify what a particular neuron in the upper layers is doing is to find the 50 positions in the database that make it produce the highest activation values. If the neuron is in one of the convolutional layers, you get a full 19x19 image of activation values, which would let you figure out what particular local pattern it seems to be detecting. If the neuron is in a fully-connected layer at the end, you only get one overall value, but you could still try to compute the gradient of its activation with respect to all the inputs, and that would tell you something about what parts of the board led to this activation being high. I think this would be a fun exercise, and you'll probably be able to understand something about at least some of the neurons.
Álvaro. On Thu, Mar 31, 2016 at 9:55 AM, Michael Markefka < michael.marke...@gmail.com> wrote: > Then again DNNs also manage feature extraction on unlabeled data with > increasing levels of abstraction towards upper layers. Perhaps one > could apply such a specifically trained DNN to artificial board > situations that emphasize specific concepts and examine the network's > activation, trying to map activation patterns to human Go concepts. > > Still hard work, and questionable payoff, but just wanted to pitch > that in as idea. > > > > However, if someone was to do all the dirty work setting up all the > > infrastructure, hunt down the training data and then financially > facilitate > > the thousands of hours of human work and the tens to hundreds of > thousands > > of hours of automated learning work, I would become substantially more > > interested...and think a high quality desired outcome remains a low > > probability. > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
