Digital images make a computer’s “eyesight” fantastic, but how well can software, even advanced neural networking programs, actually recognize and identify the contents of a photo? Not very well, currently, but a new paper published in the Public Library of Science may be changing that by modeling neural network functions after human visual abilities.
Computer Functions Modeled on the Human Brain
Artificial neural networking is a type of computer networking that models data processing and retrieval on our own organic neural network – the human brain. This project measured human abilities to recognize and identify images in small increments of time, and used the information gathered to postulate a lateral networking structure for artificial image recognition.
Interview with Co-Author Garrett Kenyon
Decoded Science had the opportunity to ask co-author Garrett T. Kenyon a few questions about this study:
Decoded Science: What do you consider to be the most important aspect of this project?
Kenyon: I believe the most important aspect of the project is the insight it provides into the possible role played by lateral connections between neurons in the cerebral cortex. Most neural network models, such as the one you cite below, are purely feed forward. There is an input layer, which receives the raw sensory input, usually several intermediate layers in which the input signals are remapped so as to allow more efficient classification of the sensory input, and finally a decision layer.
However, in the cerebral cortex, the majority of connections are lateral, which means they are doing something fundamentally different, since lateral connections do not seem to be involved in remapping sensory inputs. In our project, we test the hypothesis that lateral connections instead act to highlight the important aspects of the sensory input, in our case, smooth, continuous contours, which suppressing non-essential aspects of the sensory input, or clutter.
Pages: 1 2