A. The model architecture. Sensory input is received, and output generated, in three modality-specific sensory regions (black ovals) constituting the Spokes Layer. This layer is connected to an Intermediate Layer (green ovals) which in turn is connected to the Hub Layer (orange oval). A small number of sparse connections also directly connect the Spokes Layer to the Hub Layer. A context signal is received as input to three units forming a Control Layer (red oval). This context signal specifies the modality in which the sensory input is received and the modality in which the output should be generated. This Control Layer may be connected to all layers (full connectivity), or to the Spokes Layer (black arrow), Intermediate Layer (green arrow) or Hub Layer (orange arrow) only. B. The model environment. The 16 concepts are shown. Each concept is a row, and each feature is a column. If the feature is present the box is black, if absent the box is white. When control requirements are present, each concept can be presented in one of three modalities, and output required in one of three modalities, resulting in 144 different trials. Each modality has a surface similarity structure (orange, yellow or green boxes), which is orthogonal to a multimodal structure (red boxes) learnt through conceptual abstraction across modalities.

The effects of a maturational delay in learning control on conceptual abstraction, learning trajectory and training time. Control is added in full from the start, or after 1000, 2000, 3000, 4000 or 5000 epochs and the Control Layer has full connectivity to the rest of the model. Results are collapsed across the various protocols for gradually adding control. A. The effects of a maturational delay in control on the conceptual abstraction score. B. The effect of a maturational delay in control on the time taken to train the model. A&B. Bars signify the median value and the 25th and 75th percentile values. White asterisks signify a significant effect of the maturational delay in adding control in contrast to including control requirements from the start of training. Dashed lines allow easy comparison of each delay period to this baseline. C. The effects of a maturational delay in control on the evolution of the mean conceptual abstraction score across training. D. The effects of a maturational delay in control on the trajectory of the representation of the context signal, or control structure in the Hub Layer. C&D. The mean similarity of the representations in the Hub Layer to each structure is displayed after every 1000 epochs of training. Results for a model where control is never added is shown with a dashed line for comparison only as this model would never meet all core requirements of the human semantic system. A maturational delay results in greater representation of the conceptual than the contextual structure in the deep hub throughout training. In contrast, including control requirements throughout results in first learning the control structure in the deep hub before replacing it with the contextual structure. E&F. The effect of the developmental period without control on conceptual abstraction and training time are displayed for each of the different protocols and lengths for adding control (red = sigmoidal, blue = linear, lighter = control added over fewer epochs, darker = control added over more epochs). An additional instant protocol is shown for comparison (black). Error bars represent the standard error of the mean.

Visualising the model representations during learning across varying maturational delays in control. Control was present from the start of training or added around 1000, 2000, 3000, 4000 or 5000 epochs (dashed line). Control was added gradually using the sigmoidal protocol over 1000 epochs in the model with connectivity between the Control Layer and every other layer. Similarity matrices were constructed comparing activation in the Hub Layer across all the examples in each context. For comparison the target conceptual similarity matrix used to determine the conceptual abstraction score (i.e., the context-independent representation structure) is displayed, alongside the contextual similarity structure (reflecting the context signal only, which denotes the input and output modality of each trial) and the full structure (the combination of these contextual and conceptual structures). The Hub Layer specialises in the abstracted conceptual representation structure. Without control this functional specialisation is present in the hub from the start of training, whereas with control the context signal has a large initial effect on these deep representations.

Assessing the effects of a maturational delay in learning control on conceptual abstraction score and training time with control connections to the different layers. The Control Layer is connected to the Spokes Layer (black), Intermediate Layer (green) or Hub Layer (orange). Control is present from the start, or added gradually around 1000, 2000, 3000, 4000 or 5000 epochs (with a sigmoidal protocol over 1000 epochs). A. The effect on conceptual abstraction. B. The effect on training time. Bars signify the median value and the 25th and 75th percentile values. White asterisks signify a significant effect of the delay in adding control in contrast to the inclusion of control requirements from the start of training. Dashed lines allow easy comparison of each delay period to the baseline level where control is present from the start.

The emergence of semantic control after semantic representation in our meta-analysis of triadic comparison. Top panel: Examples of a match-to-sample task illustrating the two low control conditions (taxonomic vs. unrelated and thematic vs. unrelated concepts) and the two high control conditions (both including a taxonomic and a thematic match but differing in the instructional cue). The correct answers are highlighted in green. Bottom panel: Mean and bootstrapped confidence intervals indicating the proportion of times participants in different age groups chose the taxonomic (high proportion) or thematic (low proportion) target in each study type. Knowledge of taxonomic or thematic relationships is indicated by non-chance responding in the low control conditions (light grey and dark grey lines respectively, light or dark grey asterisks denote responding significantly better than chance, not all contrasts are possible). The ability to use the instructional cue to inform the choice of an appropriate match in high control conditions is indicated by the difference between the proportion of taxonomic versus thematic responding for the taxonomic (purple line) and thematic (orange line) cues. Significant context-sensitive effects indexing semantic control ability, are signified by the black asterisks.