Intelligent Sensor Data Pre-processing Using Continuous Restricted Boltzmann Machine

The objective of the project is to finda solution to pre-process noisy signalfor sensors in
Lab-on-a-Chip (LOC) and System-on-Chip (SOC) technologies. This solution must be
able to process continuous-time, analogue sensor signals directly. It must also be
amenable to hardware implementation, with low power consumption. This solution is
found in the Continuous Restricted Boltzmann Machine (CRBM), which is a type of
Artificial Neural Network which exhibits probabilistic and stochastic behavior. CRBM
utilizes continuous stochastic neurons, where Gaussian noise is applied to the inputofthe
neurons. The noise inputs cause neurons to have continuous-valued, probabilistic
outputs. The use ofstochastic neurons in CRBMgives it modelingflexibility that is useful
with real data. The training algorithm of CRBM requires only addition c;nd
multiplication, which is computationally inexpensive in hardware and software. The
ability ofCRBM to model any given data set is shown by training the CRBM on various
data sets reflecting real-world data. In this study, CRBM is shown to be suitable to be
implemented in LOC andSOC applications aforementioned.