According to the Science Daily, in the past few decades, neuroscientists have been working hard to design computer networks that can simulate the precise and rapid completion of human brains, such as identifying objects. Prior to this, no computer model could match the ability of the humanoid brain to recognize visual objects after a short while. Now, the latest research by neuroscientists at the Massachusetts Institute of Technology has found that one of the latest generations of so-called "deep neural networks" can match the brains of primates.
Scientists successfully build computer networks to simulate human brain recognition objects
Since these networks are based on neuroscientists' current understanding of how the brain recognizes objects, the success of the latest network suggests that neuroscientists have a more precise grasp of the basic principles of object recognition, studying the brains and recognition of senior authors, the Massachusetts Institute of Technology. James DiCarlo, dean of the College of Science and Neuroscience, said. The study was published in the December 11 issue of the journal Public Library of Science and Computer Biology.
“These models are able to predict neural responses and object distances in the neural population space, suggesting that these models encompass our current best understanding of the mysterious part of the brain,†said Di Carlo, a member of the MIT McGovern Brain Institute. .
A better understanding of how primate brains work will facilitate the development of artificial intelligence and a new way to repair visual dysfunction one day, research lead author Charles Cadieu, postdoctoral fellow at the McGovern Brain Institute. Added. Other co-authors of the article include graduate students Ha Hong and Diego Ardila, research scientist Daniel Yamins, former MIT graduate student Nicolas Pinto, and former MIT Undergraduate student Ethan Solomon, and researcher Najib Majaj.
Inspired by the brain
As early as the 1970s, scientists began to build neural networks, hoping to simulate the brain's ability to process visual information, recognize speech, and understand language. For vision-based neural networks, scientists are inspired by the hierarchical representation of brain visual information. As the visual input sequentially enters the primary visual cortex and the underarm (IT) cortex from the retina, the visual input is processed at each level and becomes more and more clear until the object is finally determined.
To simulate this process, neural network designers created multiple computational layers in the computer model. Each layer performs a mathematical operation, such as a linear point product. At each level, the representation of visual objects becomes more and more complex, and irrelevant information, such as the position or movement of objects, is discarded.
“Each individual element is generally a simple mathematical expression,†said Cardie. “When you combine millions of such mathematical expressions, you can transform the original signal into a very suitable representation of the object through complex transformations.†In this study, the researchers measured the brain for the first time. Object recognition ability. The research led by Hong and Ma Jiajie implanted an electrode array in the inferior temporal cortex and V4 region, part of the visual system that connects the infraorbital cortex. This allows them to observe the neurological performance of the animal as it is seen, that is, the number of neurons that respond.
The researchers then compared these neurological manifestations with the neural representations produced by the deep neural network, which contains the digital matrix produced by each of the computational elements in the system. Each picture produces a different array of numbers. The accuracy of this model is determined by whether it can organize similar objects into similar clusters in the neural representation.
"Through each such computational transformation, through each layer of the network, specific objects or pictures will gradually approach, and other objects will be further and further away," said Cardier. The most appropriate network was developed by researchers at New York University, which classifies objects and macaques.
More powerful processing power
This successful neural network recently discovered depends on two important factors. One of them is a major leap in computer processing power. Researchers have been using graphics processing units (GPUs), a small chip that high-performance processing of the huge visual content required for video games. The second factor is that researchers can now use and import algorithms into large data sets to "train" them. These data sets contain millions of images, each of which is annotated by people from different levels of authentication. For example, a picture of a dog can be annotated as an animal, canine, domestic dog or dog breed.
Initially, neural networks were not good at identifying these images, but as they saw more and more pictures, and after they found themselves wrong, they gradually improved their calculations until they finally identified the objects more accurately. Cardier said the researchers did not know what caused these neural networks to distinguish between different objects.
“This is both an advantage and a disadvantage,†said Cardier. "The advantage is that we don't need to know anything about these objects. But one big drawback is that it's hard to monitor these networks and investigate the internals. Now people find these neural networks very reliable and they will try to understand how they work internally."
DiCarlo's lab is currently trying to produce models that mimic other aspects of visual processing, including tracking motion and recognizing three-dimensional forms. They also hope to build a model that includes feedback projections from the human visual system. The current network only models "forward" projections from the retina to the subcortex, and there are up to 10 times more connections from the subcortex to other parts of the system.
Threaded Insert Magnets,Threaded Bushing Magnet,Threaded Bushing Magnets,Thread Rods Insert Magnets
Ningbo Shine Magnetic Technology Co., Ltd. , https://www.shutteringmagnetic.com