Errors Favor Intelligent Beings

Zhang Junping

Tsinghua University Press

July 2019

55.00 （CNY）

Brief introduction：

This book starts from the analysis of humans’ perceptive function and introduces anatomical knowledge and basic principles of humans’ vision， auditory sensation， tactile sense and somatosensation with vivid examples. It then moves onto human psychology， beginning with humans’ emotions， memories， and dreams， then to inspiration and illusion. In this process， the author also discusses the problems and treatment principles that computers will encounter when dealing with human perception， while also introducing the mathematical background of the subject under discussion. The author effortlessly cruises among the disciplines of life， computing， mathematics， physics， ect.， giving readers a dizzying tour of interdisciplinary sciences.

Zhang Junping

Zhang Junping is a professor and doctoral supervisor at the School of Computer Science and Technology， Fudan University. His main research fields include AI， machine learning， image processing， biometric authentication and intelligent transportation.

Machines designed to demonstrate AI may make mistakes， either due to the amount of data used for learning not being enough to cover the data or sample space required to solve problems， or in that the machine learning model is too complex to generate an effective prediction for new unknown samples or data – commonly known as over-fitting; either the model itself designed based on AI is incapable， resulting in insufficient ability to characterize samples， or the limited hardware condition results in the failure of related tasks. No matter what the mistake， there is always a reason for it.

However， many mistakes made by agents， especially by human beings， are due to a lack of clear explanation. Human beings are prone to making mistakes， and wrong judgments often caused by poor vision， auditory sense， distance， cognition， emotion; and even the very basis of human development， that is one’s genes， may lead to mistakes as well. Why could such an error-prone agent dominate the earth over all other species？ What can these mistakes bring us？ Understanding these mistakes may provide useful clues for rethinking the development direction of AI in the future.

Let’s start with our visual fallibility， which is often called “optical illusion”.

Let’s begin with optical imaging. The first problem is visual inversion which has not been fully understood yet， but is the most fundamental issue. The pinhole imaging principle （FIG. 1.1） tells us that an object’s image is inverted after it is mapped onto the retina through the pupil – something like the convex lens. For a machine， the inverted image can be restored to normal by optical transformation. But the agent does not seem to have the ability of optical transformation as the visual information obtained from the retina will be sent to the brain through the optic nerves. Cone cells and surrounding rod cells in the central part of human’s retina （commonly known as the fovea） are mainly responsible for detecting light intensity， color and motion and do not seem to have the ability to invert images automatically.

What if the eyes failed to automatically revert inverted images into normal and saw the world upside down？ Jin Yong mentioned this in his martial art novel The Legend of Condor Heroes. The West Poison Ouyang Feng transposed the meridians of his body by virtue of his strong inner strength to practice the 1 Jiuyinzhenjing –Jiuyinjiajing that he got from Huang Rong. As a result， he was possessed by Devil and became a man who walked with his hands and ate with his feet. People in martial arts circles all thought his kungfu was over from then on. Surprisingly， after a period of time， he seemed to have got used to this upside-down world. He refined his kungfu， learned a new set of martial arts， and won the first place in the second sword competition at Mount Hua.

Of course， this is just fiction. But historically， some scientists did attempt something similar. In 1897， the American psychologist George Stratton published his paper “Vision without inversion of the retinal image”. In his paper， he describes the experiment on retinal inversion in detail （FIG. 1.2）. In this experiment， he wore a convex lens and completely covered one of his eyes. For the first four days， the images reverted to normal by convex lens was still inverted in his vision. As a result， it was very difficult and awkward for him to act as normal as the images were upside-down while he went about his daily routine. He found himself reaching up for something on the floor and down to reach for something on the shelf. However， after the fifth day， his vision became normal， as if the optic nerves had already adapted and his movements were once again in tune with the world. But when he took off the convex lens again， he found that the world was actually upside-down， and it would take him some time to adapt to it again. When the experiment was performed on the other eye， the result remained the same. Therefore， he concluded that human’s retina does not revert the image; instead， the inverted images are achieved at the back end of the optic nerves. In other words， the inverted signals perceived by retina are transmitted through the optic nerves to the visual center of the cerebral cortex， where they are automatically inverted. This is also the consensus achieved in academic community at present.

In fact， there is a simple way to check the vision on retina is inverted. As you read here， you may place your hand under your lower eyelid and gently push it up. You can see a blurry block of words and pictures going down， not up. For people with strong visual ability， they may see a dark spot during the process. This is precisely caused by optical visual inversion.