The prefrontal cortex is thought to be engaged in diverse cognitive tasks such as working memory, decision making, response inhibition, category learning, time estimation, and many more. There is consensus that prefrontal cortex neurons exhibit mixed selectivity, i.e., a single neuron’s activity generally contributes in several of such tasks. It is therefore likely that such distributed coding could be based on the dynamical organization of prefrontal ensembles across tasks. To test this key hypothesis of the Research Unit across species and developmental stages, we propose to develop and extend required data analytical tools that allow both to identify recurring population patterns (ensembles) and to relate them to task parameters. Specifically, we intend to connect patterns to behavioral parameters by adopting recent developments (so called adversarial attacks) from deep learning research, which allow to identify classification boundaries in high dimensional feature spaces, i.e., the neuronal ensembles that are most informative about the behavior. The methods developed and validated in this project will be applied in collaboration with the experimental laboratories of this Research Unit to a) identify neuronal ensembles that are most informative about certain task b) explore to which extent these ensembles are already present as intrinsic patterns before engagement in tasks, c) how the representations change across development and species, and d) which neurons contribute to the ensembles.