The pattern of light that strikes the eyes is processed and re-represented via patterns of neural activity in a “deep” series of six interconnected cortical brain areas called the ventral visual stream. Visual neuroscience research has revealed that these patterns of neural activity underlie our ability to recognize objects and their relationships in the world. Recent advances have enabled neuroscientists to build ever more precise models of this complex visual processing. Currently, the best such models are particular deep artificial neural network (ANN) models in which each brain area has a corresponding model layer and each brain neuron has a corresponding model neuron. Such models are quite good at predicting the responses of brain neurons, but their contribution to an understanding of primate visual processing remains controversial.

These ANN models have at least two potential limitations. First, because they aim to be high-fidelity computerized copies of the brain, the total set of computations performed by these models is difficult for humans to comprehend in detail. In that sense, each model seems like a “black box,” and it is unclear what form of understanding has been achieved. Second, the generalization ability of these models has been questioned because they have only been tested on visual stimuli that are similar to those used to “teach” the models. Our goal was to assess both of these potential limitations through nonhuman primate neurophysiology experiments in a mid-level visual brain area. We sought to answer two questions: (i) Despite these ANN models’ opacity to simple “understanding,” is the knowledge embedded in them already useful for a potential application (i.e., neural activity control)? (ii) Do these models accurately predict brain responses to novel images?

We conducted several closed-loop neurophysiology experiments: After matching model neurons to each of the recorded brain neural sites, we used the model to synthesize entirely novel “controller” images based on the model’s implicit knowledge of how the ventral visual stream works. We then presented those images to each subject to test the model’s ability to control the subject’s neurons. In one test, we asked the model to try to control each brain neuron so strongly as to activate it beyond its typically observed maximal activation level. We found that the model-generated synthetic stimuli successfully drove 68% of neural sites beyond their naturally observed activation levels (chance level is 1%). In an even more stringent test, the model revealed that it is capable of selectively controlling an entire neural subpopulation, activating a particular neuron while simultaneously inactivating the other recorded neurons (76% success rate; chance is 1%).

Next, we used these non-natural synthetic controller images to ask whether the model’s ability to predict the brain responses would hold up for these highly novel images. We found that the model was indeed quite accurate, predicting 54% of the image-evoked patterns of brain response (chance level is 0%), but it is clearly not yet perfect.

Even though the nonlinear computations of deep ANN models of visual processing are difficult to accurately summarize in a few words, they nonetheless provide a shareable way to embed collective knowledge of visual processing, and they can be refined by new knowledge. Our results demonstrate that the currently embedded knowledge already has potential application value (neural control) and that these models can partially generalize outside the world in which they “grew up.” Our results also show that these models are not yet …