Multi-layer neural networks
We have seen that if we are prepared to wait, using an SNN can produce, at least in some cases, better results than any of the previous algorithms. For a lot of problems, adding extra hidden layers can produce better results than networks with just an input layer and an output layer. Will this help with our current task?
SNNs compute very similar information to that calculated by Naïve Bayes and SVMs. The links between input and output nodes carry information about how strongly the input nodes (that is, words) are correlated to the output nodes (that is, emotions) and how the biases roughly carry information about how likely the given output is. The following tables show the links between several common words and emotions after training on the CARER dataset:
anger |
fear |
joy |
love |
... |