Bernal S D_2010.pdf - University of Plymouth

2.1. OBJECT RECOGNITION
2.1.2.1 HMAX / 'The Standard Model'
In 1999, Riesenhuber and Poggio presented a landmark paper describing the fundamental ele­
ments of object recognition models in the visual .system (Riesenhuberand Poggio 1999). These
principles were exemplified in a computalional model, HMAX, also known as the .-iiandani
model. It was labelled standard as it attempts to consolidate in a single model many of the
widely accepted facts and observations in the primate visual system (more specifically, the ven­
tral path). It has subsequently been employed to simulate other phenomena such as attention
(Walther and Koch 2007). biological motion (Giese and Poggio 2003) and learning using spike-
lime dependent plasticity (STDP) (Masquelier and Thorpe 2010)- it is also the backbone of the
architecture used for the model in this thesis, and for that reason il will be described in greater
detail in this section.
The HMAX model attempts to reproduce activity and functionality observed along the ventral
visual pathway, comprising areas VI, V2, V4 and IT. Tlte model is based upon widely accepted
basic principles such as the hierarchical arrangement of these areas, with a progressive increase
in receptive field size and complexity of preferred stimuli, as well as a gradual build-up of
invariance lo position and scale as we move further up the hierarchy. These concepts have been
described in Section 2.1.
Several versions of the model have been published, although they all share the same underlying
structure. It usually comprises three different levels representing V1. V2/V4 and FT. which are
subdivided into two layers, simple and complex. Figure 2.4 shows a schematic representation
oflhcHMAX model including Ihc different types of units and operations, and the mapping onto
the visual cortex.
Two operations are performed in aliemaling layers of the hierarchy: the invariancc operation,
which occurs between layers of the same level (e.g. from SI to CI); and the selectivity operalion
implemented between layers of different levels (e.g. from CI lo S2}.
Invariance is implemented by applying the max function over a set of afferenis selective lo the
same feature but with slightly different positions and sizes. Thus, the respimse of a complex
16