The same point applies to many other computer-based object recognition systems, including deep convolutional neural networks, namely that they have no way of learning transform invariances from systematically transformed sets of training exemplars of objects.
An element r i, k) of R describes the degree that object k is suitable as the exemplar of object i comparing to other candidate exemplars for i.
Experiment 2 shows that VisNet produces single neurons with sparse representations, in which a single neuron can respond to many exemplars of one object and to no exemplars of many other objects (Figs. 8, 9).
Note that for the majority of objects even the large version was considerably smaller than real life exemplars of the object, since all of the morphs could be comfortably grasped by one hand.
The relatively poor performance of VisNet (which produces a biologically plausible output), and of HMAX when its performance is measured in a biologically plausible way, raises the point that training with a diverse sets of exemplars of an object class as in the Caltech dataset is not a very useful way to test object recognition networks of the type found in the brain.
The reason for this is that the exemplars of each object class in the CalTech-256 dataset are too discontinuous to provide a basis for learning transform-invariant object representations.
Fourteen schizophrenia patients and 10 healthy control subjects participated in an fMRI experiment to determine the functional status of the FFA by viewing a series of faces and exemplars of other object categories, while completing a low-level task designed to verify their engagement with the stimuli.
The observed decreases in activity in these regions replicate results on priming of different exemplars of single objects while further extending these results to similar complex visual scenes.
If view-tuned units were used to read the outputs of HMAX, then these units did have a more sparse representation, but again had some responses to all the exemplars of all objects as shown in the same Figures, and as noted for the first property, did not have view-invariant representations and so again required powerful decoding to read the VTUs to classify an image as a particular object.
In the category-match search displays, a second exemplar of the object was a distractor.
