Feature extraction is a crucial part of many MIR tasks. In this work, we present a system that can automatically extract relevant features from audio for a given task. The feature extraction system consists of a Deep Belief Network (DBN) on Discrete Fourier Transforms (DFTs) of the audio. We then use the activations of the trained network as inputs for a non-linear Support Vector Machine (SVM) classifier. In particular, we learned the features to solve the task of genre recognition. The learned features perform significantly better than MFCCs. Moreover, we obtain a classification accuracy of 84.3 % on the Tzanetakis dataset, which compares favorably against state-of-the-art genre classifiers using frame-based features. We also applied these same features to the task of auto-tagging. The autotaggers trained with our features performed better than those that were trained with timbral and temporal features. 1.