Next: Using SNoW Up: The SNoW Architecture Previous: Threshold-Relative Updating Contents

Prediction

SNoW currently implements two different testing policies.

Winner-take-all: The target with the highest prediction confidence when testing a given example is selected as the final prediction for that example. Note that (in light of Section 4.3.5) this prediction confidence is actually a weighted sum of sigmoid activations of target nodes in a cloud that may contain multiple target nodes. Note also that if the user instantiates only one algorithm for each target, the prediction confidence is equal to the sigmoid activation of the single target node in the cloud.
Single-target: SNoW predicts either positive or negative if the single target node's activation is higher (actually, $\ge$ ) or lower than its threshold respectively. In this case, there is no weighted sum or sigmoid function involved.

Those are the policies that SNoW can handle automatically. The user may also choose to score examples externally, making use of the results of testing calculations that SNoW performs. For instance, target node activations and sigmoid activations can be reported. In addition, SNoW can automatically normalize target node activations with softmax and report the results. The softmax normalized activation for target on example is

$\displaystyle \frac{e^{\Omega_t(e)}}{\sum_{t' \in T} e^{\Omega_{t'}(e)}}$

where $\Omega_t(e)$ is the target node activation of on . See the -o command line parameter for more information.

Next: Using SNoW Up: The SNoW Architecture Previous: Threshold-Relative Updating Contents

Cognitive Computations 2004-08-20