Output Encoding

• a single output unit, assigning 0.2,0.4,0.6,0.8 as being left, right, up, and straight
• OR have 4 output nodes and choose the highest-valued output as the prediction (1-of-n output encoding) - more degrees of freedom to the network for representing the target function
• the difference between the highest valued output and the second highest valued output gives a measure of the confidence in the network prediction
• what should the target values be?
• could use etc., but we use - sigmoid units can't produce 0 and 1 exactly with finite weights so gradient descent will force the weights to grow without bound