Aug 27, 2017 · 1 min read
Bias may be considered like a systematic error, which can be corrected at the output rather than meddling with input data. As we can’t be sure how the inputs interact and are exactly transformed to the outputs (apart from generally knowing the algorithmic process of generating statistical models using back propagation with non-linear transfer functions), we will likely need to have some sort of “bias correction” layer(s). This may add complexity, but it would also be like “using different parts of the brain ” similar to making fast decisions and then refining them using a more nuanced but slower approach.
