Machine learning techniques have been often applied to perform analysis and visualization of complex data sets, and in many cases they provided better results than conventional statistical methods. Our work explored the potential of such an approach in a fish biology study aimed at analyzing shape development and its relationships with prey selection. Lateral outlines of larval stages of gilthead sea bream (Sparus aurata, n=362) and dusky grouper (Epinephelus marginatus, n=199) were recorded and decomposed by means of Elliptic Fourier Analysis (EFA). Gut contents of the same specimens were also collected and analyzed. A Self Organizing Map (SOM) was then trained using the EFA coefficients as a proxy for actual fish shapes. As fish shapes are obviously autocorrelated in time and nearest neighboring units in the SOM correspond to very similar shapes, ontogenetical sequences were very clearly mapped onto the SOM. Shape visualization was obtained by transforming the SOM weights, which can be regarded as EFA coefficients for virtual fish shapes, back into fish outlines. This ontogenetical SOM also allowed visualizing relationships between shape variability and fish age. Abundance of food items in gut contents was then mapped onto the SOM units, showing changes in diet throughout ontogeny and relating prey selection to fish shape. In conclusion, this approach was very effective in summarizing and relating to each other several morphometrics and ecological features, which could not have been taken simultaneously into account by conventional statistical methods. Moreover, while the latter are mostly based on linear paradigms, SOMs are certainly more effective in handling complex non-linear relationships.