Assisted leg cycle exercise for wheelchair users with muscular dystrophy.

The ability to understand and generate geometric shapes has often been considered a distinctive feature of human cognition [–]. Here, we use this term “geometric shape” to refer both to non-figurative shapes that consist of simple lines and curves organized by spatial regularities, and to schematic figurative drawings depicting objects or scenes using such geometric elements (see for examples). While studies on non-human primates [,] and birds [], have demonstrated a capacity for abstract visual information processing, there is some preliminary evidence that these species recognize, generate, and manipulate geometric shapes in a way that is quite distinct from humans. Indeed, even after extensive training, such behaviors have not been observed in non-human animals, whereas human children, from a very early age, appear to engage with abstract figures effortlessly, even without explicit instruction [,,]. Recent work by Schmidbauer, Hahn, and Nieder [] has added nuance to this picture by suggesting that crows may be capable of using some abstract geometrical features, but this finding was criticized [] for using only a very small number of shapes and large differences that could be processed non-geometrically. In humans, the ability to perceive and produce abstract geometric regularities is observed across diverse cultural contexts, including populations with little formal education, such as indigenous groups in Namibia and the Amazon [,,], suggesting that it is not solely a product of Western schooling. Archaeological evidence further indicates that early Homo sapiens engaged in geometric mark-making at least 73,000 years ago [], and even earlier engravings, dating back 540,000 years, have been attributed to Homo erectus []. . Schematic depiction of the 3 different geometric tasks on which the neural networks were tested: 1, detection of outliers within quadrilateral shapes; 2, memory for geometry signs; 3, matching of geometric drawings with photos. In each case, we tested how different neural networks can model human behavior, and how they compare to symbolic models of shape perception. In line with the language of thought hypothesis [], Sablé-Meyer et al. [] suggested that humans encode geometric features symbolically. By symbolic, we refer to the representational format posited by the language of thought hypothesis: systems composed of discrete elements that can be composed into complex expressions according to a set of explicit rules, a grammar []. This proposal extends the language of thought hypothesis to geometric perception, treating shape cognition as one of several uniquely human abilities, alongside natural language, mathematics, and music, that may rest on a shared capacity for recursive, compositional representation [,]. The mental representation of a shape would be encoded as a “mental program” in this internal language of geometry: a compositional expression that draws the shape by combining a small set of primitives according to grammatical composition rules. Their research demonstrated that a model incorporating symbolic representations of an object’s features (e.g., parallelism, or right angle) provides a better account of human behavior than standard vision models [], including those with a high “brain score” such as CorNet [].