hmax

Hmax (Hierarchical MAX) is a computational model of the visual ventral stream that has been employed in neuroscience and computer‑vision research to simulate the feed‑forward processing of object recognition in the primate brain. The model was first introduced by R. Riesenhuber and T. Poggio in the late 1990s and is based on a hierarchy of alternating layers that perform simple‑cell filtering (S‑layers) and complex‑cell pooling (C‑layers). S‑layers apply Gabor‑like filters to extract oriented edge information from an input image, while C‑layers compute a max‑operation over local neighborhoods, providing invariance to position, scale and minor variations. By stacking several S‑C pairs, the architecture progressively builds increasingly complex feature representations that become tolerant to transformations, mirroring the functional properties observed in visual cortical areas V1, V2, V4 and the inferotemporal cortex.

Hmax was originally implemented as a biologically inspired alternative to purely engineered features, and it demonstrated competitive performance on early object‑recognition benchmarks. Its simplicity, interpretability, and close correspondence to physiological data have made it a popular testbed for investigating the relationship between neural coding and machine perception. Variants of the model have incorporated learnable filters, stochastic pooling, and deep learning components, extending its applicability to modern vision tasks while preserving the core max‑pooling mechanism.

Critics note that the original Hmax lacks the depth and representational capacity of contemporary deep convolutional neural networks, and that its hand‑crafted filters limit adaptability to diverse visual domains. Nonetheless, Hmax remains a seminal reference point in the study of hierarchical processing, providing insight into how successive layers of selectivity and invariance can emerge from relatively simple operations.