Deep learning network for South African birdsong and its wider implications on the concept of intelligence

We develop an operational convolutional neural network, that distinguishes 170 different South African birdsongs. The dataset consists of 3228 recordings extracted from Xeno-Canto, a user-generated database. Most recordings are however low quality, with multiple birdsongs a nd background noises. This problem is solved by pre-training using the larger and higher quality BirdCLEF2016 dataset (24607 recordings, containing however no South African birds). The fact that one makes a quite complicated recognition on the basis of a limited and noisy training set reveals a key ingredient about recognition and, “by a leap of faith”, about intelligence: the pre-existing architecture has to identify the relevant constitutive features.