Speech recognition is one such case. We thought that by breaking down the incoming speech into phonemes we would be able to get words from the bottom up. That never worked too well, whereas statistical speech recognition has been much more effective.