Computational mechanisms transforming visual codes into sparse representations

The human medial temporal lobe (MTL) is critical for declarative memory by forming sparse, concept-based neural codes. While sparse coding has long been considered the dominant representational scheme in the MTL, the mechanisms by which it emerges from preceding feature-based visual representations remain poorly understood. Here, we review recent evidence for a novel region-based feature code in the human MTL, in which neurons encode receptive fields within a continuous visual feature space-analogous to place cells in the hippocampus. This intermediate representational scheme suggests that traces of feature-based encoding persist in the MTL, bridging dense, visual feature-based coding in the ventral temporal cortex (VTC) and sparse, concept-based coding in the MTL. We propose a mechanistic computational framework in which VTC neurons encode feature axes and MTL neurons encode receptive fields within the VTC feature space defined by these axes. By linking two traditionally separate representational regimes-visual feature encoding and semantic abstraction-this framework provides a biologically grounded account of how the brain transforms perceptual inputs into conceptual knowledge and offers insights for artificial systems capable of abstraction.