Regarding biomedical imaging, DNNs can be used for anomaly classification \cite{plis2013deep, hua2015computer, suk2013deep}, segmentation \cite{li2015cross}, recognition \cite{xu2015stacked} and brain decoding \cite{van2010neural, koyamada2015deep}. Concerning biomedical signal processing we similarly have anomaly classification \cite{huanhuan2014classification, wulsin2011modeling, turner2014deep, zhao2014deep}, brain decoding and emotion classification \cite{freudenburg2011real, an2014deep, li2013affective, jia2014novel, zheng2015revealing}. There is aslso research on patients suffering from schizophrenia \cite{plis2013deep} and Alzheimer’s disease \cite{taigman2014deepface}. One area that can lead to progress is figuring out how to encode raw data to be fed to the networks, so that DNNs can learn features on its own, instead of being provided by the …show more content…
On one hand, deep learning requires large amounts of data \cite{bengio2009learning} and that can provide some challenges: data can have too much noise and that can sometimes very easily lead to wrong conclusions; there are a lot of very rare diseases, which means the sample size will be small and hard to gather knowledge from; some data can even be restricted to parts of the scientific and research communities due to various reasons. Solutions to these problems mostly consist of pre-training \cite{lopez2013insight, erhan2010does, pan2010survey, deng2009imagenet} and advanced training methods \cite{lee2015boosted,