FIG. 6.

Rank-order preservation of single units, not sensitivity to transformations, predicts population performance on invariant recognition tasks. A: combined average performance on the invariant recognition tasks across all the simulated populations when they are sorted by their single-unit rank-order preservation or single-unit sensitivity to position or clutter. The degree of sensitivity was a measure of a neuron's average response reduction from its preferred stimulus, (see methods). Each dot on the plot is the performance from 1 population. The performance of simulated IT and V1 populations is highlighted in the plots on the position-invariant recognition task. The more clutter sensitive populations appear to perform slightly better than the less clutter sensitive populations because LIN, AVG, and DIV all qualified as clutter sensitive when sorted by their clutter sensitivity. B: combined average performance on the recognition task in clutter when rank-order preserved populations were further sorted based on their single-unit separability. A joint tuning is strictly separable (i.e., independent) if it is the product of the tuning along individual stimulus dimensions. Some rank-order preserved populations could have nonindependent tunings (e.g., CCI, DIV). C: IT and V1 population performance on the position-invariant recognition task (single object) as a function of unit number. Error bars in B and C indicate SDs.