Neural representation of action symbols in primate frontal cortex

Main Understanding the mechanisms of intelligence requires an explanation for generalization, especially to situations or problems that considerably differ from those previously encountered. For example, if asked to draw an animal that does not exist, children can generalize from previous experience to produce an imaginary animal, such as a dog-like creature with six legs, three camel humps and three pig tails 8 . An influential hypothesis for this ability is that such generalization depends on an internal representation of discrete units (symbols) that can be recombined into composite representations in a process called compositional generalization 1 , 2 , 3 , 4 , 5 , 6 , 7 . Symbols enable the combinatorial derivation of numerous possible novel representations from a few reused components (for example, animal = 1 torso + 8 arms + 4 legs). This hypothesis is not restricted to concepts explicitly represented as symbol systems in language, computer programs and mathematics but may also be broadly applicable to abilities that are not superficially symbolic 3 , 5 , 7 . In humans, these include geometry 5 , handwriting 9 , drawing 10 , dancing 11 , musicianship 5 and speech 12 . In nonhuman animals, such abilities may include reasoning (logical 7 , spatial 13 , physical 7 , numerical 6 and social 14 ), object manipulation and tool use 15 , artificial grammar learning 16 and communication 17 . Despite behaviour implicating the existence of symbolic representations, we lack definitive evidence for whether and how symbols are implemented in neural activity. Furthermore, it is uncertain how symbols reconcile with other mechanistic theories of cognition, including those based on distributed processing in neural networks 18 , 19 , dynamical systems 20 and cognitive maps 21 , 22 . Given that symbols are discrete representational units that are recombined, a neural population that represents symbols should exhibit at least three essential properties: (1) invariance, (2) categorical structure and (3) recombination. Invariance means that activity is independent of variables irrelevant to the task goal. Categorical structure means that there is expression of one distinct activity pattern per symbol and a bias towards these discrete patterns even with continuous variations in task parameters. Recombination implies that the activity pattern of a symbol should reoccur in all contexts in which it is composed with other symbols. Neural recordings during cognitive tasks have revealed a diversity of invariant representations, including of rules 23 , 24 , actions 25 , 26 , sequences 27 , 28 , numerical concepts 6 , perceptual categories 29 , cognitive maps 22 and other high-level concepts 30 . Moreover, specific brain regions have been associated with such representations, including the prefrontal cortex (PFC) 23 , 24 , 27 , 28 , 29 and the medial temporal lobe 22 , 23 , 30 . However, it is unclear whether these activity patterns exhibit the other properties expected of symbols. First, with a few exceptions 29 , previous studies rarely assessed categorical structure by testing systematically whether activity varies discretely with continuous variation in task parameters. Second, evidence for recombination in activity is also rare, with a notable exception being hippocampal neurons that encode novel spatial paths that seem to reuse parts of experienced paths 31 , 32 . However, whether these continuous paths reflect a recombination of discrete components is unclear. Third, the tasks in these hippocampal and other studies of invariant representations generally lack tests of compositional generalization. Consequently, it is unclear whether and how identified activity patterns support compositional processes. Thus, we still lack evidence for a neural representation of symbols. That is, activity that jointly exhibits invariance, categorical structure and recombination in the behavioural setting of compositional generalization. To search for such a representation, we developed a task that involved symbol-based compositional generalization implemented in macaque monkeys (Fig. 1a,b ). This task includes the generation of novel, goal-directed action sequences, an ability that is thought to often involve the recombination of discrete units of motor behaviour, or action symbols, into sequences 9 , 10 , 11 , 12 , 15 , 33 . For example, imitating a dance may depend on symbolic representations of dance poses 11 . Action symbols are also essential to various computational models of action sequencing, including in handwriting 9 , drawing 10 , object manipulation 34 and tool use 35 , and may be related to movement segments identified in naturalistic animal behaviours 15 , 36 , 37 . Thus, a task that requires compositional generalization in action sequencing may be ideal for studying the neural basis of action symbols. Here we establish such a task and then, through behavioural and neural analyses, identify a neural representation of action symbols in the ventral premotor cortex (PMv).

Neural representation of action symbols in primate frontal cortex. Main Understanding the mechanisms of intelligence requires an explanation for generalization, especially to situations or problems that considerably differ from those previously encountered. For example, if asked to draw an animal that does not exist, children can generalize from previous experience to produce an imaginary animal, such as a dog-like creature with six legs, three camel humps and three pig tails 8 . An influential hypothesis for this ability is that such generalization depends on an internal representation of discrete units (symbols) that can be recombined into composite representations in a process called compositional generalization 1 , 2 , 3 , 4 , 5 , 6 , 7 . Symbols enable the combinatorial derivation of numerous possible novel representations from a few...