A fundamental assumption in behavioral neuroscience is that animal activity while performing defined tasks can be vastly described by a finite set of stereotyped movements. Information about animal behavior can be then correlated with simultaneous recordings of neural activity, allowing us to understand how the brain encodes particular behaviors, what are the underlying neural circuits and how these circuits are modified during motor learning. However, classifying different types of movements can be a complex endeavor. On the one hand, the extent of animal activity recordings may be too large to be manually classified and such a classification may not be reproducible between subjects. On the other, heuristically created categories (e.g., walking, running, jumping) tend to ignore inherent information regarding intra- and inter-animal variability frequently found in unrestrained behavior. In this work, we used unsupervised machine learning techniques to classify different types of movements executed by mice performing a motor skill learning task known as accelerating rotarod. In particular, t-SNE maps were used to find intrinsic relationships between high-dimensional feature vectors within the frequency spectrum of mouse movements. In this way, we expect to elucidate the underlying feature structure, cluster them accordingly and identify these clusters with specific mouse movement patterns, improving our understanding of the dynamics of the learning process of a new motor skill.