講演寫的，沒啥水平，就是給AI小白總結一下， 當時自己想了解一下，就查了兩三星期的文獻， 當時還是啥都不懂，因為這個，我才開始學習AI和PYTHON等課程，現在還是初級階段，估計沒機會做AI項目了。
The History and Future of Artificial Intelligence
- The world’ first computer was built in 1943-1946, which helped US Army to break German code.
2. Soon after the time which computer became a reality, early Artificial Intelligence theories was published as well:
- McCulloch & Pitts: He described a model of artificial neurons which uses one and zeroes to do logical calculation in a paper “A logical calculus of the ideas immanent in nervous activity”.
- Alan Turing in 1947 introduced the Turing Test - compare the machine intelligence with average human. He also outlined machine learning, and reinforcement learning. These two later became the major methods of modern AI.
3. Period [1952-1969] is the age of early AI success. This created great expectations with encouraging result and a lot of excitement, like kids with new toys. Samples include:
- General Problem Solver. This program was designed from the start to imitate human problem-solving methods. Within the limited class of puzzles, it handles, it was similar to that in which humans approached the same problems.
- Geometry Theorem Prover, which was able to prove theorems that many students of mathematics would find quite tricky. You remember your middle school geometry questions, right?
- A series of programs for checkers that played at a strong amateur level. Checker (跳棋) is a simpler mind game.
- AI programming language called LISP is created and used for the next 30 years.
- Herbert Simon claims “Machines will be capable, within twenty years, of doing any work that a man can do.” in 1965, WOW – what a dream.
4. Period [1966-1973]. AI looked great until reality Hits. This is the first setback for AI. It quickly turned out to be winter of AI. For example:
- Herbert Simon stated that within 10 years a computer would be chess champion, and a significant mathematical theorem would be proved by machine. It did not happen on time. It take another 20 years to achieve this.
- (1957) Early machine translation efforts. It was thought that translate between Russian and English is easy and turned out the translation is funny, ridiculous and make no sense. The language translation was assumed to be the easiest achievement, but it failed.
- First AI winter arrived in late 70s because the early hope faded and did not match expectation and what we got were mostly useless.
- (1966), All U.S. government funding for academic translation projects was canceled.
- (1973) Professor Sir James Lighthill was asked by the UK Parliament to evaluate the state of AI research in the United Kingdom. He concluded that nothing being done in AI couldn't be done in other sciences. He specifically mentioned AI's were only suitable for solving "toy" versions of reality. British government ended support for AI research except two universities.
- It proved that early methodologies and computer power cannot achieve any practical intelligence. Human brain won the first round.
5. Period (1975-1985). People went back to reality and tried to solve specific problem to make some money. Examples:
- Knowledge-based systems called “expert system” started. Work on Expert systems increase drastically. These programs were created by programmers with the help of experts. They can perform some work usually carried out by the experts.
- Program to diagnose blood infections emerged. With about 450 rules, MYCIN, an expert system, was able to perform as well as some doctors, and considerably better than junior doctors.
- Many companies put AI to commercial use to replace simple human work.
- Nearly every major U.S. corporation had its own AI group and was either using or investigating expert systems.
- Even AI lost its luster, it becomes a reality nevertheless but in much smaller scope with low expectations.
6. Period [1987-early 1990s], Second AI winter - the dark age.
- The collapse of the Lisp machine market in 1987.
- By the early 90s, the earliest successful expert systems proved to be too expensive to maintain. They could not learn new knowledge. It is too complicated to program them. They were easy to make mistakes, a lot of bugs.
- By 1991, the impressive list of goals of the fifth generation computer project penned in 1981 had not been met and failed. Funding was cut in American projects as well.
7. Period [2001-present], AI revival - the availability of very large data sets started to show its strength. After AI was stuck for a long time, people started to employ new thinking and quickly found successes.
- AI is no longer isolated and it is merged with other science like voice recognition and image recognition.
- People realized that use specific algorithm to program AI is a dead end. It is to say that people develop AI just by recreating their knowledge though programming cannot work no matter how hard we try. We can never have enough people to program unlimited scenarios. Most problems are too complicated to program anyway.
- Massive data gives significant improvement of quality with the same algorithm. System can be trained with data to do things and their behavior is reinforced and improved by additional data.
- a self-learning through massive data proves to be a better approach than coding every expert system action.
8. Period [2010-present], Deep-Learning Revolution and new AI age has arrived.
- deep Learning
Deep learning become a very powerful method, where neural network is formed and propagated through multiple layers of neurons. Each layer transforms the raw image to a more meaningful concept. Therefore, the information is inputted as raw image or voice and come out as information.
Take a look at this diagram, The first column of circles represents neurons of the input images. They are composed as pixels which make no sense to the computer, even the image may make sense to human eyes. Then the computer decomposes the image and extract information to translate pixels into the second column of neurons. Now it makes some sense such as auto-parts shows up. Then in the third layer, these parts are put together into object such as cars and other object. The analysis tells the computer how many car or other objects are in the image and at what position. With many similar images to compose large amount of data, the computer can be trained to pick up any objects in the image more reliable and faster than human can do.
- Due to deep learning with large data, AI again become a fast developing and explosive technology. People later developed voice recognition, image analysis, automatic driving cars. AI can find a particular picture from millions of pictures, diagnose medical photo, etc.
- The essence of new AI is that the computer really does not know what it is doing. It is simply extract what data tells them and learn correctly to give result by imitating. It groups similar objects together even it does not know what it is.
- The Reinforced Learning is also very powerful. A computer makes random moves at beginning and people award good moves and punish bad moves. Eventually it starts to do things better and better. There is a computer games which simulates human body parts putting together. The programmer asks the robot to move. The first moves were funny and the robot fell on every move. Every fall will get a bad score and it start to avoid doing the same move next time. Gradually the robot started to crawl and then walk. It falls less and less and eventually it can run gracefully like a professional runner in hours. So in case there are large amount of trial and data available, the system can be trained to do optimized moves.
9. Now -. AI started to grow exponentially. Each month someone shows something excitingly new. Just some recent cases include:
- Speech, images, video recognition is quickly maturing. Now even smell recognition is evolving.
- Computer can create simple tasks classification, localization, semantic segmentation, style transfer, image generation (artistic).
- Text reading like a human is becoming widespread.
- Google’s Alpha Go beat world’s best players.
- Self-driving cars started to mature. Some industry self-driving trucks can do better than human in mining industry.
- Detect various things, animal, bridge, cars, facial recognition, age, expression, etc., Mostly better than human. AI can see a picture and tell what it sees and translate image to words. This lays the foundation for machine to sense the environment, which is vital for human less machines.
- Drone technology is quickly used throughout various industry.
The near future is quickly approaching – and It is happening as we speak.
- Computer reasoning and rationalization imitates human logic. It has great potential for strategy and planning.
- Expert system – give guidance as expert without programmed by expert input. Instead the knowledge comes from deep learning.
- Planning – decision making
- Most importantly, swarm intelligence organizes thousands of individual AI units and behave like an organic unit，similar to an army:
- Swarm robotics: coordination of multi-robot systems could be used in disaster rescue missions, distributed sensing, military applications, cooperative environment monitoring, surveillance and asset management.
- Real-world optimization problems: ant-based routing in telecommunication networks, vehicle routing problems, traveling salesman problem, warehouse planning, job scheduling optimization, etc.
- The 5G network and internet of things (IoT) will empower swarm intelligence. Many devices communicate through internet to cover large areas which single device cannot do. The large bandwidth and small latency of 5G makes the large swarm AI network possible.
Now Complex AI has become a reality. Very soon, the world will be dominated by AI. Human will not be replaced by machines, but we need to do something which do not require speed and repetitive work. People have to work smarter.
