KAIST Prof. Se Bum Baeks team clarifies the principle of spontaneous generation of advanced cognitive function

(From left) Professor Se Bum Baek Gwang Soo Kim Integrated Masters and Doctorate Course Dr. Jae Sun Jang [Photo provided = KAIST]

The professor Se Bum Baeks research team the division of Bio and Brain Engineering at KAIST announced on the 4th that higher visual perceptual function can generate spontaneously in a neural network that has never gone through the learning process.

The outcome of this study is entirely contrary to the existing common sense that sufficient data learning is necessary to generate higher perceptual functions in neural networks and make fundamental questions of currently used AI realization means.

Moreover the results of the research team not only give theories that can account for the occurrence of innate perceptual functions observed in the brain of various species but also the principle of the generation and evolution of perceptual intelligence which is one of best fundamental question in neuroscience research. It presents a absolutely new perspective from the past.

Through artificial neural network simulation that mimics the visual neural network of the brain the research team voluntarily creates a quantity selectivity that selectively respond to a specific number even when a neural network initialized so that all link weights are set randomly and is not trained at all. Moreover it seemed that the quantity-selective units made spontaneously follow the same main characteristics such as *Weber-Fechner law that quantity-selective neurons found in actual animal brains are seen.

☞ Weber-Fechner law : Psychophysical law that indicates the relative relationship between stimulus and sensation. It is known as a fundamental principle in perceptual biology that the amount of perceptible stimulus intensity change is exponentially proportional to the current intensity.

This research which was co-participated by Dr. Gwang Soo kim of the division of physics at KAIST and Dr. Jae Seon Jang of the division of Bio and Brain Engineering as co-first author was published in the online sister magazine Science Advances of the international journal Science on January 1st. done. (Article name: Visual number sense in untrained deep neural networks)

Research about cognitive intelligence in neural networks is one of the core research issues in both brain and cognitive science and artificial intelligence. Interestingly distinct from artificial neural networks which commonly have to go through a learning process by way of numerous data input to make cognitive functions it has been known that animal brains have innate cognitive intelligence that carry out various cognitive functions right after birth.

These dissimilarities are supposed to play a key role in understanding the principles of the generation and evolution of biological intelligence and give core clues presenting the differences from currently developed artificial intelligence however Nothing is plainly known yet about how these cognitive functions occur spontaneously.

Consequently the research team forecasted that the simple physical structural characteristics that appear in the initial state of the neural network that have not undergone learning could generate various cognitive functions. To confirm this a deep neural network simulation study was conducted and it was found out that even in a neural network in which all the connection weights are initialized at random neural network units that selectively respond strongly to a specific quantity are generated spontaneously if only a hierarchical structure and a random feed forward connection are formed.

These neural network units present characteristic similar to the main characteristic of the quantity-selective neurons found in the real brain. This outcome proposes that the innate number selectivity found in the early days life in the biotic brain may also be originated by the same basics.

These outcomes presents that fundamental cognitive functions already exist at the time when the first structure of the neural network is equipped and can be controled through miscellaneous learning afterwards. It is evaluated as a discovery that provides very important clues about nature vs. nurture.

The outcomes of the research team show that a more in-depth study of brain functions that come about typically and spontaneously is needed Escaping from the conservative view that most brain functions occur following learning and training. On the other hand it is shown that a biological brain-based theory that can present the fundamentals of artificial intelligence implementation that is entirely different from the current artificial intelligence application skills.

Professor Se Beom Baek mentioned It is a crucial study that enabled important insights by applying the ideas gained through the brain neural network research to artificial neural network research and using the outcomes to find brain scientific principles. It is expected to bring a turning point in our understanding of the origins of cognitive intelligence one of the most important questions.

Figure 1. A photo showing voluntary quantity selectivity through a random hierarchical link structure in a deep neural network [Data provided = KAIST]]

Figure 2. Pictures showing the sense of numbers observed in various animal behaviors [Data provided = KAIST]

Figure 3. Quantity-selective neural network unit that generates spontaneously in a randomly initialized deep neural network [Data provided = KAIST]

Website: : https://www.wip-news.com/news/articleView.html?idxno=3995