The geometry of neuronal representations during rule learning reveals complementary roles of cingulate cortex and putamen

An fMRI meta-analysis of the role of the striatum in everyday-life vs laboratory-developed habits

“…Indeed, the neurophysiological properties of the putamen are supported by a subpopulation of neurons that respond to sensory stimuli 56 , unifying actions sets for movement sequences 57, 58 or integrating elementary movement units such as individual finger moves 59 . Moreover, its activity is not solely dedicated to movement parameters, but also in the absence of motor plans 60 , increasing response magnitude to the reinforced choice 61 and for predicted well-learned and contextually-driven actions 62–67 . Hence, the diversity of sensory-related and high-order reinforcement neurons in the putamen support a pivotal role in context-rich scenarios at several levels (sensorimotor, predicted actions, object value, habitual action execution).…”

Anterior cingulate and putamen neurons flexibly learn whether a hot dog is a sandwich

“…Expectedly, the monkeys were better at some rules over others. Cohen et al (2021) took advantage of variability in the monkeys' performance by splitting the sessions into easy versus hard rules. This was done under the assumption that if the animals are learning more successfully during the easy sessions, greater learning-related changes in neural activity should exist.…”

“…While frontostriatal circuits are known to contribute to behavioral flexibility and category learning after extensive training (Miller and Cohen, 2001;Mansouri et al, 2020), the neural dynamics of individual neurons in cortex and striatum during the acquisition of newly learned rules remains largely unknown. In this issue of Neuron, Cohen and colleagues (Cohen et al, 2021) demonstrate a geometric way of representing the activity of individual neurons in the primate dorsal anterior cingulate cortex (dACC) and putamen that reveals the contributions of each region to category learning. By mapping neural activity into a multidimensional space where a rule is represented as a…”

Subspace alignment as a mechanism for binding

“…Mixed selectivity both offers a challenge for traditional interpretational strategies of physiological data and offers an opportunity for new solutions to classic problems in neural coding, such as the binding problem(s). In particular, mixed selectivity can bestow on populations the ability to use high dimensional geometry, which creates a ‘blessing of dimensionality’ and gives room to create specific subspaces for options, into which their component features can be combined, and which can be easily partitioned from other options (Fusi et al, 2016; Cohen et al, 2021; Bernardi et al, 2020; Tang et al, 2020; Parthasarathy et al, 2019; reviewed in Ebitz and Hayden, 2021 and Urai et al, 2021). Moreover, such combinatorial tricks are highly flexible, readily changeable on a moment to moment basic, and easily decoded.…”