Each day there are new articles being published about Artificial Intelligence (“AI”) and its impact on society, but what exactly does it mean to use AI? In more recent years, AI has been associated with pieces of software that analyzes a user’s input which outputs a response that resembles the way humans talk (think ChatGPT). However, this is only one flavor of AI. This blog post will look at what it means to use AI and different subsets and implementations of AI.
Artificial Intelligence:
As you may have realized, the term AI can mean a ton of different things, but AI’s true goal is to replicate human intelligence and problem-solving abilities[1]. As in the ChatGPT example, the program is attempting to reply to human prompts in a way that a human would respond to that prompt. To do that, the algorithm has been trained on billions of documents, websites, and computer code to predict how a human would respond to the input prompt. Using this databank of documents, the algorithm predicts how a human would respond based on historical responses.
These algorithms have been around for quite a while. I remember being in 7th grade and a cartoon paperclip would show up mid document to say “It looks like you are writing a resume”. This was a very early form of a Large Language Model algorithm, but for Clippy to be able to guess I was writing a resume was very impressive at the time. Now, it seems like everything is getting labeled as “Having AI” or “Built with AI”. I even saw a golf club that was marketed as using AI, but I am certain no amount of AI technology could fix my slice.
AI is used to describe a whole host of techniques used to either derive human-like responses or estimations from a user input. Up until this point, we have looked at text based outputs from these algorithms, but what about other outputs? Facial recognition software compares an input image to a database of images and finds the highest probable match to the input. Another algorithm could estimate the number of individuals who would buy a certain product in the next month for an automated inventory system. All of these examples have different outputs, each of which can be solved by AI.
Machine Learning:
As we saw from the Large Language Model example, there is a group of algorithms that use an aggregation of documents and other medium to create a suggested output to a prompt. Machine Learning is a more nuanced form of AI that uses historical data to estimate or project something based on historical input. A good example of this would be estimating the number of hits a baseball player will get in a certain game. A data scientist may look at data that contains historical batting hits and outs for individual players, the pitcher they are going against, the stadium they are playing at, weather forecasts, and any other data that they find relevant. This data will be used to train a machine learning algorithm to estimate the number of hits that player will get on that day, at that location. There are several common machine learning algorithms including:
Generalized Linear Models
Random Forests
Neural Networks
K-Means Clustering (Check out our past blog on this!)
Generalized Additive Models
Gradient Boosting Algorithm
What does it mean when we say a model is learning? There are two types of learning in the machine learning world: supervised and unsupervised learning. Supervised learning has a clear objective such as estimating the number of hits a player will have during a given baseball game whereas unsupervised learning mostly looks to group certain input. For example, say you have 100 baseball players, and you want to group them by into four groups by the number of hits they had last year, the number of games they missed due to injury, and salary. You could use a clustering algorithm such as K-Means to group these players.
In general, the output to a machine learning model will be either an estimated value like number of hits in a baseball game, the probability of experiencing a tornado, or a way to group certain types of datapoints into like sets. To summarize, Machine Learning uses historical data to create estimates.
Predictive Analytics:
I’m a big believer in transparency, so if I said I knew what predictive analytics was before writing this blog, I would have been lying. However, after a quick Google search, it seems like I’m not the only one who isn’t 100% certain about the difference between Machine Learning and Predictive Analytics. My research came across numerous explanations that ranged from Machine Learning and Predictive Analytics being the same thing, to them having a brief overlap, to them not necessarily being the same with Predictive Analytics being an implementation of Machine Learning algorithms. Instead of giving a definitive answer, I’ll give you my thoughts on what I think Predictive Analytics is and where it fits in this story.
My understanding, and the way that I have seen the phrase used in practice, is Predictive Analytics is an implementation of Machine Learning and/or Artificial Intelligence algorithms to estimate an outcome based on an input. The first example that comes to mind as I write this is a Spotify radio station picking the next song on my playlist. It uses the songs that I have liked and runs an algorithm that produces the probability of me liking a song. It then can select the song with the highest probability for me to listen to next. The algorithm uses past listening data and user inputs to predict a song I will like.
Final Thoughts:
I think the best way to wrap this up is with an image.
Machine Learning is a subset of Artificial Intelligence and once the algorithm is used on unseen or production data to make decisions, the practitioner is using Predictive Analytics.
If any of these algorithms or concepts are something that you or your company want to explore more, reach out to me and we will be more than happy to help your company implement the Predictive Analytics solutions your company needs.
Comments