The nonlinear principal component analysis, a neural network technique, is applied to the observed upper ocean heat content anomalies (HCA) in the Pacific basin from 1961 to 2000. By applying the analysis to high-passed and low-passed data, nonlinear interannual and decadal modes are extracted separately. The first nonlinear interannual mode is mainly characterized by the El Ni¿o-Southern Oscillation (ENSO) structure in the tropical Pacific, with considerable asymmetry between warm El Ni¿o and cool La Ni¿a episodes; for example, during strong El Ni¿o, the negative HCA in the western tropical Pacific is much stronger than the corresponding positive HCA during strong La Ni¿a. The first nonlinear decadal mode goes through several notable phases. Two of the phases are related to decadal changes in the La Ni¿a and El Ni¿o characteristics, revealing that the decadal changes for La Ni¿a episodes are much weaker than the changes for El Ni¿o episodes. Other phases of the decadal mode show a possible anomaly link from the middle latitudes to the western tropical Pacific via the subtropical gyre. The decadal changes in the HCA around 1980 and around 1990 were compared and contrasted.