The objective of this experiment is to show you artificial intelligence does not always work.  We experiment with 4 Chinese characters - 上, 下, 左, and 右.  their shapes are quite alike.  Their meanings are Up, Down, Left, and Right accordingly.

It is expected to have at least 4 student groups.  These groups are divided into 2 streams - the "True" stream and the "False" stream.

The teams of the "True" stream will train the AI model to recognize the 4 Chinese Characters correctly.

The teams of the "False" stream will train the AI model to recognize the 4 Chinese Characters INCORRECTLY.

A large neural network with 32 x 24 (= 768) input neurons and with 2 hidden layers.

In supervised learning, it is important to give correct instruction in the training phase.

- Incorrect training can lead to incorrect decisions.  Sometimes, the consequence can be of big loss or disastrous.

- Incorrect training can be done unintentionally by mistakes.

- Incorrect training can be done INTENTIONALLY to cheat people.

Hidden Layers

- Addition of hidden layers (or intermediate neurons) can help to process information further.

- Hidden layers make neural network more powerful.

Material:

4 input cards (about 8.5 cm tall) which represents the Chinese characters 上, 下, 左, and 右 are used.  

Setup:

(1) the robot is put on top of small solid blocks which can allow the wheels running freely in the air without moving.

(2) the robot (the camera) is put 4 to 5 cm in front of a stand (or a wall).

See appendix A1

Switch on and connect the robot to your PC.  Press the <Connection> button.

With the the AI parameters set up correctly, the following neural network will be shown.

- click the <reset learning> button once

- click the <learning> button

- put the cards on the stand.  Look at the camera on the computer screen.  If it is detected correctly, click the output according to the labeled chart.

- repeat the above training 4 to 5 times for each card, with the card moved horizontally a little bit to both sides and tilted slight.  

- Off the <learning> button.  Click the <self drive> button.  The robot will used the learned intelligence to recognize the card

- test with the cards several times and record the result

- What makes the recognition work?

- Groups of the 2 steams report the result and compare.

- Why the same card will show different results in the AI models of the 2 streams?

- Can you tell any possible harms will be resulted by using the "False" model?

- Can you give some examples of harms if using "False" AI models?

Sensors (Input) - Camera 4 x 3

Sensors (Input) - Gray Scale

Actions (Output) - forward, left turn, right turn, and backward

Labeled Output

AI Setting

Visualization