This study investigates the use of Multi-Layer Neural Network (MLNN) to estimate HC, CO, and NO concentration in the exhaust gas of Spark Ignition (SI) engine. The training and validation data for developing MLNN are generated at various operating points on an experimental test rig consisting of four stroke single cylinder SI engine coupled with an eddy current dynamometer. The inputs to MLNN are variables deduced from averaged cylinder pressure over 120 cycles. Cylinder pressure variables are chosen as input because they represent the combustion process more closely than externally measured variables, such as Air Fuel Ratio (AFR), manifold pressure, spark advance, etc. Further, engine-out emissions are very much sensitive to the combustion process going inside the cylinder. Initially, ten numbers of variables are derived from averaged pressure profiles. After that, the dimensionality of inputs is reduced to three using principal component analysis. The concentrations of emissions predicted by MLNN are found to be quite close to experimental values, with relative error for HC, CO, and NO to be 5.11, 5.12, and 3.50%, respectively.