„Nature meets Technology: From lions, smart farming and artificial intelligence in language“
Workshop 2:
“Inside the Mind of AI: Teach Your Own AI to Recognize Handwritten Digits”
Ines Neji (ZUKOnnect Fellow from Tunisia / Department of Computer and Information Science
In the project “Inside the Mind of AI,” Dr. Ines Neji explained everything about the topic of Handwritten Digit Recognition.
First, she asked us what AI is and then showed us where AI plays a role in our everyday lives: AI is when computers can think, learn, and make decisions, almost like people. This only works when we provide them with the necessary information so that the AI can recognize this information everywhere. Machines learn by looking at examples, just as we do. They make many mistakes and learn from them. After many examples and mistakes, AI knows exactly how the digit 3 looks, for example.
Dr. Neji also demonstrated how difficult it is for AI to learn handwritten digits because of our handwriting. Some people write small, others messily, and it always looks a bit different. AI has to recognize each one of them, which takes a significant amount of time. We often use AI in our daily lives: when we unlock our phones with Face ID or when Google Maps finds the best route.
Dr. Neji also explained what MNIST is: it is a dataset where every digit from 0 to 9 is saved as black-and-white images. The AI receives a total of 70,000 pictures: 60,000 for training and 10,000 for testing whether the AI knows every digit. The AI gets more examples if it makes more mistakes.
The next big question was how AI makes decisions: many neurons in an input layer process the 28×28 pixel images. These images go to another layer, the hidden layer. In this layer, there are fewer neurons than in the input layer. Here, the “robot brain” analyzes the patterns. After that, the patterns go to the third and final layer: the output layer. In this layer, there are again fewer neurons than before. They decide which digit they can “see” in the pattern.
To demonstrate how this works, Dr. Neji showed us a program called Colab Notebook. With this, we trained a neural network on MNIST and observed how the AI improved and learned on its own. Then, we drew a digit in poor handwriting, and the AI tried to recognize which digit we had in mind.
AI in the real world is much more advanced and experienced. It is also designed for specific areas: as an assistive driver to help cars park or even drive themselves, like Tesla, or to help doctors check X-rays and scans to find diseases faster. It is also used to suggest movies, music, or games we might like, based on our algorithms. With the digits, it can even do our homework.
To sum up, we learned how we use AI in our daily lives and how it improves through mistakes and pattern recognition. Some AI decisions can also be unfair, as it learns from biased data, so we always need to verify the answers.
