What Is AI?
Artificial intelligence (AI) is an interdisciplinary field that leverages mathematics and statistics, cognitive science, and computing to enable problem-solving based on vast and robust datasets with high-performance computers.
Both machine learning and deep learning are important sub-fields of AI. You can think of them as Russian dolls—all of deep learning fits into machine learning, and all of machine learning fits into artificial intelligence. What they have in common are algorithms that seek to create intelligent systems to help us make predictions, decisions, or classifications based on input data. Deep learning is the most advanced and mature type of artificial intelligence that most closely mimics human intelligence. And like the human brain, it can operate in rather mysterious ways.
Examples of AI:
- Self-driving cars
- Speech and facial recognition
- Digital personal assistants
- Chat bots and virtual customer service
- Recommendation engines
Rooted in computer science, the evolution of learning machines draws from multiple disciplines:
- computer science
- cognitive science
An umbrella term for computer programs that can make independent decisions based on supervised and unsupervised learning.
The most common type of artificial intelligence, where a machine makes independent decisions but still needs a human to guide it and correct its mistakes, not unlike a toddler or adolescent.
The most mature version of artificial intelligence, where the machine decides on its own whether its predictions are right or wrong based on artificial neural networks—it learns through its own method of computing, and the human sometimes doesn't know why or how the machine reaches a particular conclusion.
Self-driving cars and computer vision, such as facial recognition technology, rely on deep learning.
Is it AI or computer science?
Artificial intelligence draws from a wide range of disciplines, such as computer science, math, philosophy, biology, and cognitive science—psychology and neuroscience. The applications of artificial intelligence as a research method are even broader. Everything from infrastructure projects to construction and design, from drug development and medical diagnosis to the protection of endangered species, LSU researchers in every college are using AI.
How does it work?
Artificial intelligence, just like humans and animals, can learn by trial and error. By optimizing an objective function that penalizes mistakes, the machine can learn how to perform a task correctly and then use that knowledge (as experience) the next time it encounters the same problem. Machine learning algorithms break down data into small and abstract parts to understand information they have not seen before. Learning can also be supervised or unsupervised. In supervised learning, an algorithm is presented with labeled data that tells the program which classes the data items belong to. Presented with pictures of cats and dogs, for example, the machine would at least initially need help distinguishing between the two. Humans can help the machine by repeatedly telling it, “This is a cat,” “This is not a cat; this is a dog,” etc. The machine learns the features that are generic to the whole dataset (such as four legs, two ears, a tail) and specific to individual classes (cat-only features; dog-only features). The machine would learn that two ears and a tail isn’t enough data; to also look for class-specific features, such as whiskers. The main reason for the success of supervised learning is the availability of large volumes of labeled training data as well as great computing power. In the real world, acquiring large volumes of labeled training data can be quite expensive and impractical, however. Unsupervised learning is getting more and more attention, as it allows learning from unlabeled data.
In deep learning, layers of artificial neurons are stacked together (called a deep neural network) that learn representations that loosely resemble that of a human brain. Presented with a picture of a horse, the layers of the neural network learn increasingly sophisticated representations that are used as input to the next layer. The initial layer would learn the representation of lines and edges in the image. The next layer would then use these representations to learn the representation of simple shapes, and so on. To arrive at “animal” or even “horse” on its own, the machine must reason based on context, and go beyond drawing simple inferences.
Many current implementations of deep learning fall in the area of computer vision, making it possible for the machine to “see” the world in different ways. This is necessary for both facial recognition and self-driving cars. Analyzing spatial relationships, such as angle, distance, and depth, is key to the machine drawing the right conclusion. Thesame is true of light and shadow. When a self-driving car encounters a stop sign with a sticker on it that also stands in half-shadow from a nearby building—will it still “see” the stop sign? Perception remains one of the biggest challenges in deep learning.
Is AI safe? (The Terminator question)
Existing deep learning capabilities are known to be opaque when it comes to their decision-making processes. This can become problematic when inputs lack the information necessary for an accurate decision to be made. Such situations frequently arise when dealing with noisy data and is also something that could be exploited by adversaries and attackers.
Ideally, machine learning-based decision engines also provide an account of their own confidence in a particular decision. This makes it possible for humans to judge whether any particular outcome is trustworthy. LSU researcher Supratik Mukhopadhyay’s work on confidence-aware AI (in collaboration with Intuit AI) is addressing this safety problem in AI.
Why AI matters
AI, simply put, can help us make better, data-driven decisions by analyzing vast amounts of data in a fraction of the time it would take a human. Much of it boils down to pattern recognition, including patterns we might not think to look for, or know exist. From cutting-edge research on climate change to medical breakthroughs in cancer care to creating more energy-efficient buildings, AI will continue to play a key role in how we make decisions as well as predictions.