According to the official website of the Massachusetts Institute of Technology, October 16th，from the National Science Foundation、The National Institute of Health Research and other institutions funded and recently published an article published in "Nature -Neuroscience" revealing the different "non -degeneration" between biological and artificial neural networks，It is found that the image that humans think that it is completely irrelevant can be classified as the same image by calculating the model。

Human sensory system bet365 best casino games is very good at identifying the objects you see or the vocabulary you hear，Even if the object is upside down，or this vocabulary is the voice that has never been heard。When this classification of human sensory system，It will ignore features that have nothing to do with the core identity of the object，For example, light and shadow、Observation angle, etc.，Even if the objects are not too important, the differences are considered the same，This is the so -called "non -degeneration"。

In recent years，The deep neural network trained by researchers can analyze millions of external factors that can cause people's reactions or affect human behavior，mainly sound or image，Learn to distinguish its common features，so as to be able bet365 Play online games to classify the target vocabulary or objects in general。

The deep neural network model of simulated human sensory system is very good at dealing with the transformation of uncompansence characterization，No matter what the dog's fur is，can correctly identify the dog's image，Or no matter the voice of the speaker，Can correctly identify a word。But，A new study found by researchers from MIT Institute of Technology，These deep neural network models also have the same reaction to images or vocabulary that are not similar to targets。

If it is required to simulate the deep neural network model of human hearing or vision to generate a image (such as a bear) or a word，and put it in the same category with a specific natural element (such as the picture of Bet365 app download the natural bears)，The considerable part of the content they generate is irregular for human observer。These models have established their own unique "non -degeneration"，They will make the same reaction to stimuli with completely different characteristics。Most images and sound generated in this way are not like the examples initially given by researchers，More messy pixels and noise in human eyes。Not only that，Different deep neural network models also have their own unique non -degeneration，When the researchers show the data given by one model to another model，The latter cannot recognize like a human observer。

Researchers also found，Through confrontation training，Including add some images that are changed slightly in the training Bet365 app download data，The generated model data is easier to recognize for humans。Targeted model improvement improves human recognition of model data，but did not eliminate the overall difference between humans and models。

Josh H. McDermott, an associate professor at the Department of Brain and Cognitive Sciences at MIT，These findings can help evaluate the imitation of the computing model to the human sensory perception system，Improve the existing model。(Yang Lanlan/Compilation)