FIG. 7.

Summary of the goodness of different single-unit response properties for supporting invariant object recognition tasks at the population level. In each subplot, the x axis shows the values of some identity-preserving transformation (e.g., object retinal position, or the presence of different distractor objects, see Fig. 5); the y axis shows the response of hypothetical single neurons to 3 different objects (red, blue, and green; red is the “preferred” object in all cases). Major y axis: single neurons can have a range of response sensitivity to a particular transformation X (e.g., for position, the receptive field size in response to the preferred object). Major x axis: neurons may also preserve or not preserve their object rank order. Among these single-unit response properties, rank-order preservation is much more predictive of the population's ability to support invariant recognition (assuming equal numbers of neurons in each population; see Fig. 6 and methods). For example, neurons can be largely insensitive to both position and clutter, yet form an inadequate population (see Fig. 6). Conversely, neurons can be highly sensitive to both position and clutter and still form a very good population representation. (*Note, large RF is bad for position-specific recognition, but potentially useful for generalization over position). The term invariant has been used to describe the idealized neuron in the top right plot, and the term tolerant to reflect the limited invariance of real neurons and real behavior.