A Brief Introduction to Artificial Intelligence

James Zhang
7 min readSep 20, 2020

“Robots will be able to do everything better than us…I am not sure exactly what to do about this. This is really the scariest problem to me.” — Elon Musk

You’d think that someone pushing for futuristic technology as much as Elon Musk would be head over heels for AI — but he’s not; he’s terrified.

From an individual perspective, the development of AI means that people who work normal jobs will be replaced by intelligent machines, but unemployment isn’t even the biggest potential problem. Our very human existence is.

You have to remember, machines are far more efficient than a human could possibly me. With their ability to analyze ridiculous amounts of data in a second, it begs two daunting questions.

  1. When will a robot be smart enough to consciously decide to eliminate the human species?
  2. When will an immoral person use AI to attack other humans?

Yikes. Hopefully, that never happens and we are just being cautiously pessimistic. I really do hope so. And until it happens, AI will continue to be a technological miracle.

Everywhere you look, Artificial Intelligence plays a massive role in your day to day lives. Your phone uses AI for voice recognition when it uses Siri. Your email uses AI classification methods to determine if an email is spam or non-spam. Your healthcare wearable devices — that uses deep learning regression analysis to predict if you will have a heart attack or stroke. 🤯 Pretty crazy, right?

I always looked at AI as some magical technology solely reserved for use in those cool Sci-Fi movies. But trust me on this one, it’s not magic 🧙‍♀️

If you have no idea what AI is, it’s essentially the science of making machines smart using algorithms. Before AI, humans would program algorithms that machines could follow. By definition, this would make robots equally as smart as humans, but never smarter. Here’s where AI comes into play.

Introducing: Artificial Intelligence

With Artificial Intelligence, we have the power to make machines smarter than any human in history. And you know what’s crazy? Anyone can get tap into this AI fountain of awesomeness 🔥, but let’s not get too excited. We have to start with the basics.

Machine learning is a subset of AI and specifically, it is the machine’s ability to learn. I mean, that’s why it’s called machine learning. The process of machine learning is building and testing a lot of networks; after they get tested, only the best networks remain and the rest are destroyed. More networks are made that are similar to the best but with small tweaks, and then the process repeats itself until eventually, one network remains and is perfect.

Machine learning is analogous to natural selection… except it’s artificial… haha, get it?! Anyways, if you look at all of your neural networks as a single species. Only the best of the species survive, they reproduce, and over time the entire species has the features that played an influence on the species’ survival. Machine learning is just like that but insanely quicker ⏩

Now, there are three main methods of machine learning:

  1. Unsupervised Learning is for dealing with raw, unlabeled data (which is most of the data in the world) and it does this by clustering the data into groups based on similarities.
  2. Supervised Learning is used when we are working with labeled data and we are trying to classify it into our target outputs.
  3. Semi-supervised learning combines a large amount of unlabeled data with a tiny bit of labeled data, essentially combining the best of both worlds. The goal is to use a small amount of labeled data to accurately predict the output of the unlabeled data.
supervised (classification) vs. unsupervised (clustering) vs. semi-supervised

By the way, I kinda lied :( there is a fourth type of special AI known as reinforcement learning… Essentially when there is a complex objective, it works by using a software agent whose goal is to receive as much reward as possible.

Reinforcement learning is most commonly used in the gaming industry, and here are some key terms to understand reinforcement learning:

  • Environment = the world that the agent is maneuvering in
  • Reward = the positive or negative feedback given to the agent based on its actions from the given environment
  • State = the current situation that the agent is in
  • Policy = the method to convert the agent’s input (the state) into the output (the action)
  • Value = the future reward that the agent will collect by doing a certain action in a state

The agent first must analyze its state and use the policy network to determine the action with the most value or reward attached. If the policy is bad, the agent will have trouble beating the game. Through trial and error and enough feedback, the network will become insanely good at moving the agent where it needs to go. You don’t believe me?

In March 2016, Google’s DeepMind AlphaGo algorithm beat the Go world champion — a feat that many AI and CS experts didn’t expect to happen for at least another decade.

Specifically, the DeepMind algorithm initially mimicked amateurs who played Go. With reinforcement learning, AlphaGo played a past iteration of itself over 30 million times! As time went on, AlphaGo incrementally improved and learned to avoid actions that would lead to errors in the game, leading to a new and almost unbeatable version of the algorithm.

What are the pros and cons?

Each type of machine learning has its ups and downs, so before you get started you should know which model you will use.

Disclaimer: Even the world’s best make mistakes, and sometimes the only way to find out is by trying 🌱

Unsupervised learning is able to sift through and sort vast amounts of data. Moreover, it is capable of discovering patterns in the data that no human would ever be able to see. 📈 Sounds amazing, right? Well, it does because it is! There are some downsides, though. From one perspective of the data, it looks sorted well, but then you look at it from another perspective, and it looks like a mess. This is down to the fact that the data isn’t labeled, which takes us into supervised learning.

If you have answers already attached to your data, then supervised data is best for you. For each data input that comes in, there is a target 🎯 output that the model should produce meaning, that we can calculate the model’s accuracy and incrementally improve the model.

There are two methods of supervised learning: Classification and Regression. Classification is better for discrete problems such as facial expression recognition and determining if an email is spam or non-spam. Regression on the other hand is better for predicting future trends by establishing connections between present and future events. For example, predicting the cost of something in the future.

It sounds perfect, right? Well, that’s because it literally is. That’s a massive downside with supervised learning, functions can be overfitted, or too perfect.

Let’s say we have an algorithm that can classify if we have a light-colored piece of clothing or a dark-colored piece. After many iterations, the algorithm would have developed a function that is perfect in discerning if the clothing is light or dark. So what’s the problem?

Well, let’s say the data only contains yellows, whites, blacks, and reds; then the algorithm can only classify those colors. When a new, unseen data item comes in, the function has no idea what to do and just breaks. On the other hand, the unsupervised model would easily group it by similarity and there would be no problem.

Key lessons to take away from this example: The more data you have to feed into the algorithm, the more accurate it will be. The same thing applies to the real world for humans. More experience is directly related to better decisions. 🧠

Another problem with supervised learning is the lack (and cost) of labeled data, and one way to overcome that hump is with semi-supervised learning. Many machine-learning researchers have recently discovered that small bits of labeled data can drastically improve the classification accuracy if combined with a large amount of unlabeled data.

What are some of the current applications of machine learning?

Classification

  • Object detections
  • Bioinformatics (fingerprint and face recognition) 🙍‍♂

Regression

  • Health breakdowns (predicting strokes, heart attacks, diabetes glucose levels, etc)
  • Weather forecasting

Clustering

  • Pattern recognition (ex. in medical patients)
  • Anomaly prevention (detecting fraud)

Why should we be excited about AI going into the future?

I’ll just give you a few reasons.

  • Self-driving cars leading to better road safety
  • Better healthcare for assisting doctors by diagnosing diseases and better data analysis
  • More efficient manufacturing
  • Improved customer service
  • Less fraud in online transactions (in collaboration with blockchain)

Key Takeaways:

  1. AI is the science of making machines smart
  2. Machine Learning is a machine’s ability to learn through past experiences
  3. The three main types of ML are unsupervised, supervised, and semi-supervised and they are all useful in their own ways
  4. More data leads to better algorithmic functions
  5. AI is only in its infancy and it will continue to influence the world as we know it
  6. I’m super excited but also scared about AI. Uh oh.

Thanks for taking the time to read my article! It means a lot :)

Stay tuned for more articles that I will be writing in the coming weeks, and if you want, follow me for more or DM me 👏❤️

--

--

James Zhang

CS & Math at University of Maryland, College Park