Intro To Deep Learning: Taught by a 14-Year-Old
Jumping into the deep end of deep learning
“ I am an optimist and I believe that we can create AI for the good of the world. That it can work in harmony with us. We simply need to be aware of the dangers, identify them, employ the best possible practice and management, and prepare for its consequences well in advance.” -Stephen Hawking
Many of you may have started to hear the legendary tales of Artificial Intelligence curing cancer, or heard the frightening horror stories about intelligence robots taking over. Today, or whatever day you are reading this (hello future robot overlords), I will explain what Deep Learning is, why people are so afraid of it, and how it can be used.
As humans create new advances in S.T.E.M., one field seems to be rapidly progressing every day. Deep Learning, a subset of Machine Learning and AI, is advancing at an astonishing rate. Nearly every day, huge companies such as Google and Amazon figure out new ways to give computers the ability to detect things such as whether a person has cancer or being able to rapidly detect bone breaks in radiology scans. These computers can detect this with more precision and speed than the average doctor.
Artificial Intelligence vs Machine Learning vs Deep Learning: A brief history of the evolution of cognitive computation
Artificial Intelligence: Mimicking Human Intelligence
Contrary to what you would think, Artificial Intelligence has been around for a very long time. It wasn’t until the Dartmouth Summer Research Project on Artificial Intelligence in 1956 where the idea of machines exhibiting human intelligence seemed like a possible achievement. That conference marked the inauguration of endeavors into the field, and while standard computation was non-complex at the time, it still seemed like an important milestone. “What can Artificial Intelligence created in the 50s do?” you may ask. The most popular example is the chess app that nearly every computer comes preinstalled with. If you decide to play chess against a computer, you are really using the most basic form of Artificial Intelligence. The computer knows which move to make next based on hard-coded algorithms.
Machine Learning: A new take on Artificial Intelligence
With Machine Learning, a subset of Artificial Intelligence, engineers are able to feed massive amounts of relational data into a computer and gain educated predictions of future data. This flourished for things like understanding housing prices in an area, weather patterns, spam emails, and more. While machine learning was successful for the fields it was used in, it still was not vastly used due to its many constraints (better suited for small amounts of data).
Deep Learning: Demonstrating high cerebral activity
Deep Learning, a subset of Machine Learning, is giving a computer the ability to “learn” (progressively improving performance on a specific task) without being explicitly programmed on how to accomplish said task. Deep neural networks are able to take in data inputs, such as images and audio files, and learn from what is being labeled. This new technology has only started to be implemented recently due to new and improved understanding of the human nervous system as well as hardware improvements. Beforehand, the immense computational power required by these networks made it nearly impossible to train and evaluate models. With the rise of GPUs (graphics cards), the same key component used for Bitcoin mining, researchers were able to do more mathematical operations per second. Deep Learning is so new that many computers still are not powerful enough to run models; when I tried to teach a network the difference between playing cards, it would have taken 28 weeks to train on my laptop whereas my desktop with a GPU only took 15 minutes.
The neuroscience behind it
Our brains are able to comprehend what we experience by having neurons throughout our bodies that react when things happen. An example is that some neurons in our hands may fire off electrical pulses to our brains only when an object is hot, while other neurons may only fire when something is cold. Neurons are all over our body and inside our brain, each with a different task. The brain alone has approximately 86 billion neurons. This system of interconnected networks of neurons, feeding data to the brain developed into an incredible scientific breakthrough: Deep Learning.
How Deep Learning learns
Supervised Learning
Convolutional neural networks are the most common type of neural network using supervised learning. Supervised learning means that humans take non-corresponding input data and label the data to help the model learn. For image training, this would entail labeling objects within an image so that the computer is able to test itself with an “answer key.” This is what I did with the playing cards; I had to label 100 different photos to tell the computer where each type of card was within the given pictures.
Unsupervised Learning
Generative Adversarial Network
A GAN (generative adversarial network) is a type of unsupervised learning that is able to generate data from scratch from image training data. Nvidia created a GAN that generates a face; while the face in the photo looks real, it was never a human that ever existed. The GAN took images of humans and found generalizations between the way humans look.
Reinforcement Learning
One new method of training a neural network with unsupervised learning is called reinforcement learning. While most convolutional neural network models are trained with supervised learning (the model evaluates itself from given input data), the reinforcement learning model evaluates its performance based on given rewards. Currently, this form of training is being tested with video games where points scored in a game can give immediate satisfaction to the model, indicating improvement. This was used most recently with Google DeepMind’s AlphaGo, where it beat the 3-time European Champion 5–0 in the very complicated game, Go. The AlphaGo computer was able to learn without any instructions whatsoever.
The morality of Artificial Intelligence
These new scientific breakthroughs give us a glimpse into the future of society, one where cars drive for us, a medical analysis is much cheaper and more accurate, and our lives are organized by our robotic personal assistants (like Siri, Alexa, or Google Home). Many prevalent executives and researchers fear what could come if we misuse these new technologies. Being that Deep Learning is a new subject, we cannot determine the limits of its usage; training methods like reinforcement learning could lead to computers becoming more independent from humans as we let them figure out the world on their own. However good these possibilities sound, it leads to serious unanswered questions regarding things like the rights of artificial beings or the possibility of the end of society as we know it today.
Per my quote at the top of the page by Stephen Hawking, he summarizes the products of misuse of these new technologies due to poor decisions from early adopters. While at this point we are unable to be sure as to what the future will bring from artificial intelligence, we must, as ridiculous as it sounds, recall movies we may have watched where robots invade and beware of the dangers AI could bring.
Deep Learning may not yet be ready for large scale commercial applications mainly because of a fatal flaw: Deep Learning training data is what allows it to give predictions, not morals. As proven by an Amazon trial in October of 2018, Amazon scrapped an AI initiative after they realized it was sexist as a result of bad data. Amazon tried to make the process of hiring easier on Amazon’s part by creating an AI that can help to decide whether an applicant is a qualified candidate as an employee. This AI had trained off information about the past hiring history of Amazon employees and rejected applications. Those doing the hiring of applicants were often sexist as less qualified males were hired in the past over more qualified females. This serves as a perfect example of how bad training data can cause poor models. Another example of poor training data would be an early test of autonomous cars in which a computer was unable to detect a stop sign if there was a post-it note on the sign, whereas most humans would still know that it is a stop sign.
The status of Deep Learning in 2019
While there are many pros to AI, there are many drawbacks that may not seem threatening without deeper thought. In reality, the likelihood that AI does become out of control and end the human race as we know it is very slim, for now at least. The bigger threat to society is job loss; one job that is not thought over often is truck driving. Right now in the United States, there are over 3.5 million truck drivers. Volvo and Uber are already testing out their fully autonomous trucks that have the capability to take a driver out of work. As of January 30th of 2019, Amazon was already testing fully autonomous trucks on Interstate 10. With even further thought into the loss of 3.5 million truck driver jobs, many small truck stops would also go out of business if there was no need for drivers to stop and take a break.
Deep Learning is actually the 4th industrial revolution and may prove to be beneficial to employment rates. Succeeding previous revolutions caused by inventions like the steam engine, electricity and the combustion engine, and the internet, Deep Learning will hopefully employ more jobs than the number of lost jobs. Many people feared job loss when the internet became widespread, but it created more jobs than there were previously; some examples of jobs that were created because of the internet include social media manager, data scientist, Uber driver, and more — these jobs did not even exist 15 years ago.
Innovation with assistance from Deep Learning is not stopping soon as an executive order was just signed to put sizeable resources towards turning the U.S. into the world leader in AI rather than China. “In July 2017, [China] unveiled a plan to become the world leader in AI aiming to create an industry worth $150 billion to its economy by 2030, and two Chinese cities promised to invest $7 billion in the effort,” The New York Times reported.
It is daunting to cope with the fact that Deep Learning is already being implemented into today’s society, but it isn’t stopping any time soon. A team at Google is using AI to help doctors prevent blindness in diabetics for patients in India where doctors cannot analyze patients quick enough on their own. In 2011, IBM’s Watson computer won ‘Jeopardy!’ with more than 7 times the amount of money Ken Jennings, the famous contestant that won 74 games in a row, received. Ultimately, the competition between the games played showed Watson as the clear victor as expressed by Ken Jennings, “I, for one, welcome our new computer overlords.” IBM’s continuous quest to defeat human beings in games of wit continued recently on February 12th of 2019 when IBM debated against Harish Natarajan, who is often considered the best debater in the world. While IBM gave many persuasive arguments for why preschool should be subsidized, IBM’s debater failed to connect to the judges on an emotional level due to its monotonous voice. If only IBM took some tips from Google’s Tacotron natural audio synthesizer. Don’t feel too bad for IBM’s loss… Google’s DeepMind just created an AI that defeated the best human StarCraft players.
Conclusion
While it may seem hard to believe at first that many inconceivable ideas that TV shows thought of are already coming to market, it shouldn’t seem scary. Deep Learning is just getting started; many ideas once thought to be unsolvable are now becoming plausible inventions. I hope this article has not made you shy away from Deep Learning, but embrace what good can come and hopefully inspire you to continue on to learn more about making neural networks.
Dedication
I never would have found my passion for Deep Learning without the help of my school, Pine Crest School, and the folks down at Florida Atlantic University’s Machine Perception and Cognitive Robotics Lab. I remember how excited I was to learn what Deep Learning is when Dr. William Hahn and Dr. Elan Barenholtz spoke to Pine Crest at an innovation conference. This journey has been an incredible experience and I look forward to furthering my research into AI and Deep Learning.
“The true sign of intelligence is not knowledge but imagination.” -Albert Einstein
