The article introduces new tools from probability theory to analyze learning algorithms. These tools help create better estimates of how well algorithms will perform and suggest new ways to measure the complexity of learning tasks. This can lead to the development of more effective learning algorithms.