In the era of modernity, understanding human brain mechanism has been one of the most intriguing topics for scientists from different aspects. Physicians and psychologists use this knowledge in the prevention and treatment of human physical and mental illness. Besides, artificial intelligence scientist tries to mimic the brain mechanism for solving the real-world problems. The first artificial neural network was introduced in the middle of 20th century, and from then on, the concept and structure have been improved over the years.
Todays, the most famous structure of artificial neural network is deep neural networks (DNN) also known as deep learning introduced by Geoffrey Everest Hinton. Deep learning is designed in a layered structure to connect millions of artificial neurons which are close to the number of neurons in human’s brain. Nowadays, the deep learning powered applications are beating the human intelligence in various subjects, from image recognition to gaming to problem-solving, one by one.
How DNN or deep learning works? Similar to its biological counterpart, it learns from what we feed them. As a simplified example, if we want to teach a DNN to detect cats in any given picture; we need to provide several hundred images from a variety of cat breeds from different angles in different positions and background. The DNN analyses the training images to learn the detail and get ready to detect the presence or absence of a cat in the given picture. If the aim is to teach our ten favorite animals, in the same way, we just need enough training images containing desired animals to feed the network.
The process of training seems pretty meaningful and straightforward, also similar to the way that the human brain learns things. However, there are two significant points that we can learn from deep learning.
First, the trained DNN tries to generalize the concept to be able to detect any similar instances rapidly. When we train the network to identify the cat, it learns the main facial and body characteristics of it such as the shape of ears, the position of eyes, nose, and mouse, number of feet, the tail and the pattern of colours on the skin. However, it may happen that the DNN recognizes a tiger as a cat because it is the most similar thing that the DNN has learned. It means, what a DNN understands is just based on what it learned and doesn't rely on the reality.
Does it sound familiar to you? Our brains, as an advanced network of neurons, see and interpret things based on what it has learned from previous observations and experiences. Here, as an example, let’s explain a funny memory from my friend about her three-year-old boy. One day, the son was watching the TV, suddenly he shouts excitedly, Mom, Mom, come and see... a big cat caught a sheep. And He asked confused, Mom, why does a cat catch a sheep? Naturally, we expect something like this scene of Shaun, the sheep, was cast on the screen.
But my friend explained that it was actually a tiger try to catch a buffalo calf.
We, all, have similar experiences with children which the concepts are mismatched to the instances. The reason is known to us: lack of training. The solution, also, is pretty straightforward: showing some sample images and movies from the right concept and in case of my friend's son, visiting a zoo may clear all ambiguities.
But what about us as adults. Fortunately, we have no problem in recognizing objects and animals but as an adult what about the higher level of reality? To how extend are we sure that our perception of what happens around us and people’s behaviours are exact reality and are not our imperfect interpretations. How we can improve ourselves.
Second, the accuracy of the trained DNN highly depends on training images. It means if the training images are not comprehensive enough, the learning will not end up with the desired accuracy. In case of our cat-detector network, if the majority of images contains the front face of the cat, it is very likely for the trained network to cannot recognize a cat when only half-face is visible. The solution is pretty clear; continue training on what missed until achieving the desired results.
In the case of the human race, human babies follow the same pattern; when they fail, they try to gain more information and learn what is missed and try again until they master the case. But, adults usually forget the process of learning. When they fail, some think this failure is because of their personality, luck, fate and anything else but the lack of knowledge and lack of training.
Fortunately, we have a developed reasoning ability which helps us to challenge ourselves and to correct the wrong training. The point is to use this magic ability. Maybe it is time to think, what we feed our brain with? Do we need any change to its diet?