Food literacy interventions to improve sustainability in multicultural diets: a scoping review.

Two seemingly discrepant accounts propose that primate prefrontal cortex (PFC) neural activity should track either low-dimensional or high-dimensional representations of the environment. Traditionally, it has been proposed that PFC cells are tuned adaptively to task-relevant information, leading to low-dimensional neural activity. This results in the population displaying structured selectivity patterns, as commonly observed after training on a cognitive task (Fig. , low dimensional). A contrasting hypothesis suggests that the PFC may rely on high-dimensional, nonlinearly mixed representations of task features to support complex cognition (Fig. , high dimensional). According to this notion, the PFC serves as a nonlinear kernel such that when a low-dimensional input is projected onto it, dimensionality expands, and a wide repertoire of responses can be generated. Learning can reduce or expand neural dimensionality, changing how many linear decoding axes can be implemented on neural firing rates (discriminability). , High-dimensional representations enable high discriminability. A high-dimensional regime allows the strong separation of all task features using three possible readout axes (left), whereas a low-dimensional representation only allows task-relevant features to be strongly separated (right). , Each neuron can be represented as a point in the three-dimensional selectivity space spanned by color, shape and XOR (their interaction). In the random model, selectivity is distributed according to a spherical Gaussian distribution in this space (‘Generative models’); the covariance matrix is computed across the selectivity coefficients; zero-mean Gaussian noise () was added to each selectivity coefficient to illustrate measurement bias under finite sampling. , Analogous to , but for the minimal model; neurons are strongly selective only for the XOR (interaction between color and shape), as this is the only feature that is necessary to solve the task. c, color; s, shape. Recently, it has been proposed that the PFC can transition between high-dimensional and low-dimensional representations during learning to accommodate the changing demands of the environment. For example, early in learning, high-dimensional representations may allow flexible exploration of all possible input–output mappings (‘contingencies’) to discriminate which task rules are currently relevant. This is because a high-dimensional representation allows for a high number of linearly separable task features (Fig. ). Conversely, once an animal has learnt that only one set of contingencies is relevant, a low-dimensional representation may be used to encode task-relevant features more robustly. Moreover, these low-dimensional representations may enable generalization to new contexts, as aligning new with old representations is likely easier when fewer dimensions must be considered. In other words, different stages of learning impose different demands on the neural population. Learning could thus shape neural dimensionality and progressively push neural activity toward different solutions along the trade-off between discriminability and generalizability, that is, from a high-dimensional regime toward a low-dimensional regime.