Introduction to Artificial Intelligence.
The content of the articles in this series of AI is retrieved from different sources. Access the link/s at the end.
In this article, you’ll learn:
- A summary of AI’s history.
- Unsupervised learning.
Understanding AI
Artificial Intelligence is on your Instagram feed that decided which ads to show you, which insurance you get, if you meet the requirements of your job or college applications… etc.
The term was coined in 1956 by John McCarthy after the Darmouth Conference, which lasted 8 weeks. Another name you should know is Marvin Minsky, who co-founded the AI Lab at MIT and also participated in this conference.
How much data and computer power does AI need? Most sorts of AI do not have senses, they are programs. We need to give these plenty of data and label it, as well as a powerful computer to make sense of it.
Do you know what the AI winter? until 2010, the field was frozen. However, certain events gave place to the AI Revolution, which happened thanks to 2 big developments:
- Increase in computer power and processing capabilities
Darthmouth Conference in 1956, the IBM 7090, as big as a whole room, could do 20K operations, which was focused on helping the US Airforce with its missile balistic system. The power depends on the number of transistors. Now, we can fit many more transistors in a much smaller space.
Our iPhones can do 600B operations per second.Impressive? Well, a modern supercomputer can do around 30 quadrillion operations per second. Yes, you read right, 30 quadrilllion!
These computers started to mimic certain brain functions in 2005.
2. The internet and social media
When you click on the screen, tag someone, post on any platform, buy something on Amazon, call Uber, apply for a loan… This information ends up in a data system. We are generating an incredibly amount of data.
However, how is this being used…? I guess you’ll have to keep reading this series for that!
Reinforcement, supervised, and unsupervised learning.
As humans, computers are not created with knowledge. These are 3 ways in which computers learn:
- REINFORCEMENT LEARNING
This takes place when the computer learns in an environment through feedback from its decisions. As you learned to walk; you fell down many times but then stood up and tried again, right?
2. UNSUPERVISED LEARNING
This process allows computers without traning labels. This is possible through clustering or grouping.
3. SUPERVISED LEARNING
Process of learning through training labels.
A supervisor points out the mistakes in the computers’ process of learning.
AI needs data and computer power as we discussed in the previous article.
After training the model, it should be able to do this grouping or clustering by itself. For example, this happens on your email account when it classifies different messages as spam, or when FB recognizes your face and suggests a tag.
Does this sound mind-blowing? Well, the same processes happen in your body too since this was inspired by our nervous system. In it we can find neurons, in which we have the cell body, axon, and dendrites. The axon is separated through other neurons by a sinapse. When the electric signal is passed from one axon to the cell body, then this is passed through another axon, to another neuron; and so on.
In 1958, Frank Rosenbell scientist wanted to create an artificial neuron to classify triangles and those figures different from triangles.
The PERCEPTRON was the machine he created to achieve this goal. Different shapes will be shown to a camera connected to it and it would recognise the pixels and send different electric signals to the Perceptron. The machine would add up all the signals it gathered from it and if the total charge was above its threshold, it would send an artificial signal to turn on a light. This meant that it was a triangle. At first these were random guesses so they used yes/no buttons. When the ‘no’ button was pushed, the amount off charge would be adjusted.
The artificial neuron receives inputs multiplied by different weights. The threshold is represented by the bias, which can be adjusted. AEach input has different weights and multiplied and added together, you get a function, a step function that can just equal a 0 or 1 — false or true. If the sum is less than bias, the neuron’s output is 0. If the sum is greater than the bias, then output is =1.
The different parameters are used to calculate a decision boundary that separates object A from object B. All those objects that do not correspond to object A — which is what we want the machine to identify — will be classified as a data point on the opposite side of the decision boundary that represents object B.
The Update Rule
The Perceptron can just be compeltely run or wrong. If object A is identified as B by the machine, then the update rule will measure the value of that signal, add that to the weight, and the weight will change.
When object A is identified as object A, then the value of the signal will make the weight = 0 and the decision will be considered as right.
It learns from failure thanks to our labelling, where we tell the machine what’s right or wrong.
After training the computer with enough data to make it understand what is A and what’s B, we can visualize the results in a confusion matrix — a table — or in a graph with a decision boundary.
To measure accuracy, we get the number of times it was right for both objects, add it together, and divide it by 100.
‘Recall tells you how much your program can find of the thing you’re looking for’ — Jabrils.
Sometimes the computer struggles to distinguish objects. For example, if you train a model with 2 parametres: diameter and mass so that it differentiates donuts from bagels, if you show a bigger donut than the average, it might confuse it with a bagel. ‘Figuring out what criteria to use is the key to most AI challenges’. — Jabrils.
More inputs are better for accuracy, but it requires more processing power and time.
So, what does an ideal AI system look like?
It’s small
It’s powerful
It has perfect precision
It has perfect recall
Unfortunately, nothing is perfect, so make sure you prioritize what’s most important so that you can achieve your goal.
For example, to make sure we get the most important emails, we need a high recall.
Make sure to read the next posts to know how to solve more complex problems.
Most of this information has been retrieved from the YouTube Channel CrashCourse taught by Jabrils. Access it here: https://youtu.be/GvYYFloV0aA
You can follow me on LinkedIn here: https://www.linkedin.com/in/nur-younis-aa79a9183/
