Overview Of The Study On Ai Optimization Of Images To Stimulate Monkey Neurons
According to Harvard Medical School, over the past 50 years, researchers have known for a fact that within the brain’s visual system, there are neurons that respond more to certain images than others, which is a feature that is vital for humans’ ability to be familiar with, recognize, understand, and interpret their surrounding visuals as well. However, up to this day, it is still unclear what these neurons are responding to. On May 2, a study published in the journal Cell was conducted to figure out what visuals specific neurons react to. The study was conducted among primates by utilizing artificial intelligence.
An Advantageous AI Approach
There are previous studies and experiments that have been conducted in a similar effort to measure neuronal preferences, however, the vast majority of such have only used real images. Real images which are stated to carry an inherent bias as they were only limited to stimuli available in the real world and only to images chosen by researchers to show. Contrary to such previous work, the recent study led by a team of researchers from the Blavatnik Institute at Harvard Medical School took a different approach as Will Xiao, a graduate student at Harvard University, developed a computer program in the laboratory of Gabriel Kreiman at Children's Hospital that was designed to be a form of responsive artificial intelligence to create self-adjusting images that are based on the data of neural responses gathered from the monkeys.
As stated in The Scientist, Carlos Ponce, then a post-doctoral fellow in the laboratory of senior author Margaret Livingstone at Harvard Medical School and now a faculty member at Washington University in St. Louis, explains the method to be a “GAN”. He says that these programs are neural networks of algorithms that could receive visuals and reproduce it in their own internal abstractions. Based on sample images uploaded to the program, they are also able to create their own infinite various realistic and abstract images, he added.
In Science Daily, he further stated that the approach has a big advantage due to the fact that the algorithm allows the building of the neuron’s preferred images from scratch using a tool that could “create anything in the world to creating things that do not even exist.” The algorithm, dubbed as “XDREAM”, takes advantage of a neuron’s firing rate as a base for the evolution of a novel, synthetic image. The algorithm then randomly generated a total of 40 images that were first shown to the primates and singled out 10 images that induced the most activity among the neurons. The images were then put through a genetic algorithm that recombined the images to produce 30 similar images along with the previous 10. The cycle was repeated for over one to three hours, 250 times until the animal was no longer interested in the images. Researchers also included a control group of naturalistic images such as those of people, places, etc.
Results of the Study
According to The Scientist, the team of researchers implanted microelectrode arrays in an area slightly above and behind the ears called the inferior temporal cortices of six macaque monkeys to record the activities of the neurons that are stimulated by visuals and, over the course of a few hours, images were shown to them in 100 millisecond blips. The images first appeared to be noise-like, or a random textural pattern in greyscale, but they gradually morphed into shapes that could be recognizable in the monkeys’ environment such as familiar faces wearing surgical scrubs and the food hopper in their room.
For instance, a neuron in one of the primates had generated images which consistently appeared to be a body of a monkey with a red splotch near its neck. Later, the researchers realized that the monkey was beside another monkey who wore a red collar. Every final image that has been produced does always look recognizable as Xiao reports that one involved a small and black square while the other, involved an amorphous black shape with orange below. Margaret Livingstone, the Takeda Professor of Neurobiology at HMS and senior investigator, pointed out that over separate runs, the results were consistent and although specific neurons had likely evolved images that were not identical, they were “remarkably similar”.
Further Human Goals
The authors of the study are stated to believe that they perceive the brain to learn relevant features of its world and that it is adapting to the environment along with “encoding ecologically significant information in unpredictable ways.”
The authors believe that the technology could also be applied to any other neurons in the brain be it on sensory, auditory or hippocampal neurons. They also suggest that more efficient artificial intelligence could further investigate how the brain works as they indicated that it is important to understand which neural networks behave safely to further human goals.
