寫在前面
這學(xué)期選了一門課叫science communication and outreach, 主要目的是讓我們這些理工科生明白如何讓大眾接受和認知相對晦澀的科學(xué)知識。最近一次作業(yè)需要些一篇400-600的popular science article，在此順便先構(gòu)思寫好內(nèi)容。本來想借總結(jié)回顧下知識點，但是由于字數(shù)限制在了，也很多內(nèi)容無法展示了。在此之前，總結(jié)下寫作技巧，這樣可以在寫作的時候順便參考。
- Entertainment value has to be more pronounced
- Define the audience, scope of the paper, format
- Popular science papers often address the ‘wider issues’, from philosophy to ethics. Emphasis the higher goals of even small step of science.
- A good title typically contains a verb.
- A lead text stimulates the readers to continue and promises some ‘reward’ at the end.
- use various methods to draw readers attention such as use different levels of text, divide the text into small paragraphs, use bold italic characters.
- Use quotations from scientist.
- Include informative simple data if necessary.
- Simplify the statement, delete useless information, avoid jargon.
- Create an easily travelled story line, e.g. there’s a situation at the beginning and another at the end, so there must be something happened in between.
- Use beautiful and engaging illustration without didactic value.
正文
AI doesn't deserve this name, at least not now
Heated debates about the future of AI leads to concern among public. People with positive attitude would say it can change the way we live, while others hold a negative viewpoint, being afraid of the life in series 'Westworld' would come true.
I.
The history of artificial intelligence is not superior to the origin of other disciplines, which we can be simply summarized as an interdisciplinary development. If you really know the process, then you could have rational awareness rather than being influenced by the some engineering applications of AI. The fundamental of AI contains three main part: mathematics, psychology and computer science.
II. Cybernetics (control theory)
This is the description of cybernetics defined by the pioneer. Literally speaking, people want simulate the behavior of animals to make item to be self-governed. Dating back to 18th century, James Watt's steam engine was equipped with a 'governor', which can be used for controlling the speed of engine.Cybernetics is the scientific study of control and communication in the animal and the machine -- Norbert Wiener
III. Psychology
When an axon of cell A is near enough to excite cell B and repeatedly or persistently takes part in firing it, some growth process or metabolic change takes place in one or both cells such that A's efficiency, as one of the cells firing B, is increased. ---Donald O. Hebb
This complicated statement from famous psychologist assumes a way our brain work, that is if keeping stimulating two relevant neurons, there relationship will become firm, then the stimulation of neuron A would active neuron B.
Based on previous two concepts, it's easy to consider, what if we have a number of neurons, combining the core idea of cybernetics 'feedback' (a way for self-government), then this neuron network can be learned by itself.
This idea first proposed by another psychologist Frank Rosenblatt, called perceptron, which introduced 'weights' between neurons (as the line showed in image below). Neuron network calculated by certain weights from different knowledge can be used for corresponding application such as distinguishing dog or cat. Latter, Rosenblatt's theory of perceptron was simulated on an IBM 704 computer at Cornell Aeronautical Laboratory in 1957.
But how to obtain the weights for different application becomes the biggest problem. In 1986, another notable psychologist David Rumelhart reinvented 'backpropagation'. Based on the concept of feedback (computing the error between input and system output), backpropogation is a method to compute the error function (called cost function) and use differentiation to find a direction to minimize the error, which leads a boom of neural network universal approximators.
IV. Connectionism
Psychologists played significant role in realm of AI ascribe to the vital importance of cognitive science. Furthermore, concept of connectionism speeded up the process of AI. It says mental phenomena can be described by interconnected networks of simple and often uniform units. Later in 1980s, the core idea parallel distributed processing becomes popular ascribe to the release of book by Rumelhart and McClelland. Generally speaking, it simulated the method of image processing in our brains, using hierarchical features, which is also the idea of deep neural network.
Image above shows the mechanism of hierarchical features, simply speaking, lower level features represents some basic symbol such as line, and higher level features becomes to the combination of line such as complex shape, in a similar vein, combination of shape can be more abstract item e.g. eyes, ear.
V. Computer science
Implementation of neural network is strongly influenced by the development of computer. As the performance of computer increase, training complex neural network model becomes easier. It's worth mentioning that GPU is naturally suitable for deep network owe to the concept connectionism, because GPU has a lot of parallel units. This is the reason why applications of AI re-boomed from 2010.
VI. Problem: Specificity and universality
Let's go back to the question mentioned in the opening, despite that AI has many applications and is a powerful means in some fields such as natural language processing and computer vision, but it just because scientist found the suitability of previous methods and algorithm, or modifying theories to deal with certain questions, to sum up, no universality. As a human, we can learn painting or playing chess with one single brain, but machines still can't now. I hope we would only use the word intelligence once something deserves it. Just like we can perceive different emotion in the malposed item in Guernica while computers even have problem for recognition.
