Appendix 4 : Lecture Notes

( There are hypertext links in the orginal document which is available at http://www.gsia.cmu.edu/bb26/70-451/lectures/lecture14.html ) Management Information Systems Lecture 14: July 19, 1994 Knowledge-Based Systems Artificial intelligence (AI) has been defined as the capability of computer systems to provide a level of performance that reflects human-like intelligence. Examples include systems that can read handwriting, provide assistance for a doctor diagnosing patients, and performing other tasks that are commonly understood to require intelligence. These systems are also called knowledge-based systems. In this lecture we will consider a variety of different AI systems, some available commercially, and others still in development. 1. Natural Language Systems 2. Robotics Systems 3. Vision Systems 4. Neural Networks

Natural Language Systems The most common way of interacting with a computer system is by typing commands into a keyboard, or perhaps selecting menu items with a mouse. While this is effective for some tasks it is not the same way that we interact with other people. Natural language systems are systems that attempt to enable users to interact with computer systems using written, typed or speech communications in a natural language such as English. Speech Systems One important type of natural language system is speech recognition systems. Some of the problems that speech recognition systems must overcome are: Different volumes (soft voices) Mumbling and monotone voices Different Accents and Emphasis Most existing speech recognition systems have the following limitations: Limited Vocabulary Must be trained for individual users The need for near perfect, noiseless conditions However, even with these constraints, commercial speech recognition systems can be effectively used in environments where users are not able to use conventional keyboard-based data entry. One example is a system developed by Zales Jewelry stores. In order to increase the productivity of jewelry appraisers Zales created a system that accepted speech input so that the appraisers hands would be free to handle the jewelry. Another important type of speech system is speech synthesis systems. Currently the belief is that speech generation technology has reached a limit. The speech sounds vaguely human, but it is still easily distinguishable as artificial. Translation Systems One of the goals of natural language system research has been to create computer systems that can translate items from one language to another. However, as people began to build translation systems it became apparent that translation was a very difficult task. One example from an early translation system was a case where the sentence "The spirit is willing but the flesh is weak" was translated from English to Russian, and then retranslated from Russian to English. The result was the sentence "The vodka is good but the meat is rotten". Currently, there are systems that provide assistance for human translators. Also there are prototype systems that can translate structured technical documents. Carnegie Group Inc. is one company that provides this type of system. Character Recognition and 'Reading' Systems An example was given in class that went as follows: Hide the marker board from the students, and write the following phrase: "Help, I've fallen and I can't get up". Erase some portion of each letter in the phrase. Then reveal the marker board with the modified phrase to the students and ask them to try and read the phrase. (Most students are able to identify the phrase with a little concentration.)

The purpose of this example is to illustrate how good humans are at recognizing items based on clues from the context. From the "Help, I've fallen and I can't get up" example it is useful to point out that if you saw the modified letter 'H' alone you probably wouldn't have been able to identify it. However, given the context of the word "Help" and the phrase as a you most are able to identify the letter as being an 'H'.

The challenge with many AI recognition systems is to enable the computer to use the context to identify objects.

One example of a recognition system that is currently available commercially is a Optical Character Recognition (OCR) systems. These systems, used in collaboration with scanners, can be used to convert paper documents into computer files suitable for storage and searching.

Another area where 'recognition' systems are becoming relevant is in the realm of information filtering. It is important to realize that in order to screen information it is necessary to 'understand' the content of the documents that are being screened. An example of this might be a filter that can scan the subject lines of e-mail messages and flag certain messages for further investigation. This type of system also must use some context to 'understand' the messages.

Another type of recognition system that has recently become commercially available is handwriting recognition systems. The most famous (or infamous) example of a handwriting recognition system is the Apple Newton. Interestingly enough these systems have many of the same limitations as the speech recognition systems described above.

Robotics Systems

In order to demonstrate some of the limitations and challenges of robotics the following demonstration was used:

At the start of the lecture ask the students to write a set of instructions for making a peanut butter and jelly sandwich. Collect the instructions from the students. Then tell the students that you will be using the directions to actually make some sandwiches. (Using some peanut butter, jelly, bread, and knifes that you have brought to class) Then, following the instructions exactly, demonstrate the difficulty of giving instructions to a being that has no common sense.

Interesting things that can occur during the demonstration are:

If the instructions don't specifically say to open the bag of bread, the jar of peanut butter, or the jar of jelly then the 'robot' should have a difficult time.

If the instructions specify that the the bread should be picked up, the knife should be picked up, etc, but never specify that the objects should be put down the 'robot' will begin to have problems.

The purpose of this example is to illustrate that information about spatial relationships and other commonsensical information is not known to a robot.

Common applications of robots include:

Stationary robots : Factory automation robots are an example of this. Mobile robots : Creating mobile robots that can endure harsh environments, drive in routine situations, or perform repetitive tasks over a large area. 1. NavLab - A driving 'robot' 2. Ambler, Dante and Dante II - Walking robots for hazardous conditions Vision Systems

The challenges and importance of vision systems were demonstrated with the following two demonstrations:

Demonstration 1 : Context and Vision Systems

Drop a small pile of rubber band (1 dozen or so) in front of one of the students. Ask the student to count the rubber band, without touching them. After the student completes that task, ask another student to lift one rubber band at a time using a stick, without disturbing the other rubber.

After this demonstration a discussion can be structured around what knowledge/techniques the students used to identify and manipulate the individual rubber bands. This can be used to demonstrate the importance of context for vision systems.

Demonstration 2 : Two-Dimensional Representation of Three-Dimensional Objects

Show the students the following picture:

[Image]

Ask the students what they see. Most answers will be things like: the a top view of a box that is projecting out of the board or a inside view of a box embedded in the board. At this point tell them that the picture actually is a single square with an X in it and another square on top of it (as shown below).

[Image]

This demonstration is intended to illustrate the interesting ability of many of us to use certain context cues to see 2-d representations as 3-d objects. This is an important skill for a driving robot to have - If it sees a square with an X and another square then it is likely to hit the first box-like object it encounters.

Another application of vision systems is monitoring images, such as satellite images for changes or monitoring manufactured parts for flaws.

Neural Networks

To demonstrate the value of neural networks (or neural nets) the following demonstration was used.

Write the following sentences on the board:

Madam, I'm Adam.
A man, a plan, a canal - Panama
Never odd or even.
Step on no Pets

Then ask the students if they can identify what the sentences have in common. After a few minutes also write the string of digits "11011011" on the board and say that the sentences are also similar to this string of digits.

The sentences are all palindromes or sentences (or strings of digits) that read the same from right-to-left and left-to-right. This is an example where there is an underlying pattern that is disguised by the context.

Essentially neural nets are software systems that attempt to identify patterns such as the palindrome pattern in spite of a distracting context.

An example of how this pattern recognition capability of neural nets is used in a business environment is its use by catalog retailers to identify potential customers. Using demographic information and records of past purchases retailers can use neural nets to identify customers who are most likely to purchase items (i.e. match a certain 'pattern') and this information can be used to target mailings and advertising.

Brian Butler / bb26@andrew.cmu.edu