50 SHADES OF IA
Encyclopedia Britannica defines Artificial Intelligence (AI) as "the ability of a computer or computer-controlled robot to perform tasks commonly associated with intelligent beings."
It's important to note that performing these tasks doesn't tell us how intelligent the computer is. It doesn't necessarily imply that the computer reasons or knows what it's doing. This is why we commonly distinguish AI into 3 sub-categories (3 major nuances):
- Narrow AI: performs specific tasks. All artificial intelligences today are narrow.
- General AI: performs as well as a human in perceptual tasks and is capable of reasoning
- Super AI: an AI as powerful as all humans combined
What do most ia's consist of TODAY?
They combine two things: abundant historical data and machine learning algorithms.
Machine learning is the process by which a computer is able to perform a specific task without using explicit instructions. It achieves this by relying on statistical models trained on historical data.
For example, a machine learning model designed to automatically detect an animal species will have learned from millions of animal photos that cats' ears are pointed. The model learns this information by taking the pixels of the image as input and finding that the pattern of pixels corresponding to the cat's ear is a good way of distinguishing cats from birds, dogs and other species. Another set of features will be taken into account by the model to distinguish cats from red pandas, for example.
There are many algorithms capable of detecting statistical patterns, which can be divided into two families:
- Shallow learning : e.g. decision trees, linear regressions
- Deep learning: deep neural networks with many hidden layers
For a semi-technical dive into surface learning and deep learning, you can read these two dedicated blog posts:
- 8 Machine Learning Algorithms Explained in Human Language
- 3 Deep Learning algorithms explained in Human Language
CE which ia not?
To set the record straight, here are a few things AI is not :
- A human brain by proxy. Neural networks are much simpler than the human brain
- A human-like robot capable of reasoning
- A know-it-all computer that can answer any question
- A program capable of predicting the future with 100% accuracy.
CONCLUSION
To sum up, today's AIs perform specific tasks, they are called restricted AIs. This can already be extremely useful for a wide range of industries and businesses.
Restricted AI is mainly based on deep learning, which is a sub-field of machine learning.
AIs can be combined together to create complex systems such as human-like robots. Such a robot can be equipped with the following models:
- A walking model (reinforcement learning)
- A model for detecting objects and people (deep learning)
- A face recognition model (deep learning)
- A reading model (deep learning)
- A game model (reinforcement learning)
Instead of a general AI, it's a combination of restricted AIs that makes robots capable.
Available data and computing resources are key factors in the expansion of AI applications worldwide. There are two major challenges for AI today:
- The short-term challenge is to democratize AI and deploy it on a large scale to help businesses and organizations take advantage of this incredible technology .
- The long-term challenge is to develop a generalist AI capable of reasoning by combining mathematical research in the fields of deep learning, graph theory and symbolic AI.
