Hello 👋! Welcome to the first entry of my weekly series Learning Artificial Intelligence.
I am excited to start this journey and share what I learn with you. This week we will be covering the basics of Artificial Intelligence (AI). Let’s dive in!
The Beginning
Computers were invented by Charles Baggage to perform defined procedures or algorithms. External factors in technology have changed, but computers’ basic principles and utility remain the same.
Things begin to get a bit abstract if we do not know each step in a sequence to give to a computer, though. What happens then? Artificial Intelligence (AI) can help us solve this problem.
Weak vs Strong AI
I think it will be helpful to divide the AI we might see or hear about into two types. Weak AI is designed to perform a narrow task, such as facial recognition, while strong AI is designed to perform any intellectual task that a human can do.
Weak AI is far more common in our normal day-to-day. Some common examples include:
- Siri
- Spotify recommendations
- Customer service chatbots
Strong AI is a goal that the AI industry is striving towards. A strong AI could reason, understand, and learn just like a human.
Intelligence
What does intelligence mean to you? Critical or abstract thought? Superior awareness about yourself or the world around you? The ability to learn and adapt?
The irony is that the definition of Intelligence is a bit of a mystery itself. It’s hard to find a definition that can be set as a standard. No standardized test or metric system can be used to measure or quantify this concept in humans or any living organism.
Alan Turing created a way to potentially measure intelligence. His test involved having a human judge a conversation between a human and a machine. The machine was considered intelligent if the judge could not tell the difference between the two.
Thinking human
How would a computer think, act, or learn like you or I? Conceptually, this starts with how humans think and understand. How can you tell the difference between red and green? You don’t consciously think about this, you know. We learned this at some point, and a computer must also learn this.
The way computers learn is through models. A model is a representation of a system that is used to make predictions or decisions. The model is trained on data, and then used to predict new data.
There are two ways to approach this problem:
- Top-down Approach: Symbolic Reasoning:
- Try to model our reasoning process
- We follow thoughts when we reason, and try to formalize this process for a computer
- Knowledge representation and reasoning are the two main components
Imagine you see a doctor, and they discover that you have a fever. Once they know this, they know that inflammation is a possible cause. By reaching this conclusion, the doctor applied a large set of rules to a certain behavior that helped them reach a decision.
2 Bottom-up Approach: Neural Networks:
- Model off of a simple element found in the human brain, a neuron
- By constructing an artificial neural network in a computer, we can teach the computer to solve problems by giving examples for those problems and solutions
A way to think about this process is to imagine how you would teach a newborn baby things. A lot of this learning is done observationally, inadvertently being shown how to do things.
History
Artificial Intelligence may seem like a newer idea, and it is more in vogue than ever; however, its origin dates back almost 70 years. The term was first coined by John McCarthy in 1956. Back then, Top Down, or Symbolic Reasoning, was the predominant approach, as many successes stemmed from creating expert systems (computers that acted as experts in a particular field with problem domains). However, this system did not scale well because feeding these models updated data from experts in a field was quite cumbersome and clunky. Because of this, advancements in AI slowed quite a bit and ultimately led to what is commonly referred to as the ”AI Winter” in the 70s.
As technology advanced over a few decades, computer resources continually got cheaper. Data also became more regularly accessible, which helped stoke the fire of the neural network effort. Due to cheaper resources and greater data availability, neural networks began to advance rapidly, most often showcasing their ability to understand human speech and behavior.
In the modern-day, neural networks are used as an interchangeable term for AI because many of the AI successes you see or hear about are based on them.
To illustrate how past approaches evolved, we can look at a search algorithm that was developed to assist with playing chess.
This was based on search; the program estimated the moves that its opponent could take, taking into account the number of next moves. Once it conducted this search, it picked what it considered to be the optimal move based on the best position that could be achieved in a few moves. This program led to the alpha-beta pruning algorithm.
This approach performed better near the end of a chess match; however, when the chess canvas is larger, there is room to improve this algorithm by incorporating the learnings from human players.
As these experiments evolved, a method called case-based reasoning was employed. This directed the program to look for similar cases between the current games and past games that had been played.
If you played an AI in chess today and it beat you, chances are it beat you because neural networks and reinforcement learning are used. Now, programs can improve their game by playing themselves continually and, in turn, learning from their own mistakes.
Our modern-day speech AI assistants, such as Alexa or Siri, use hybrid systems composed of neural networks that transform our speech into text, attempt to diagnose our intent, and then execute reasoning or specific algorithms to perform the action we are asking for.
Recent History
Neural networks really started to grow exponentially about 15 years ago (2010); why? Large datasets became publicly available for these networks to train on. This has only gained momentum since then, as we are now seeing more advanced uses of large language models such as GPT-3 and BERT.
It’s exciting to see the evolution of neural networks over the past few years and the advancement of large language models as well. The future of AI is bright, and it will be interesting to see how it evolves in the coming years.
Even in the last 10 years, we started to commonly use Neural Networks for:
- Image recognition
- Speech recognition
- Machine translation
- Captioning images
These are all exciting and practical use cases and make our lives much easier. What will come next?