In the months leading up to my son’s birth earlier this year I was spending time increasing my knowledge and understanding of Machine Learning (ML). And then the baby arrived earlier than expected and all the extra learning activities were put on hold, replaced with sleepless nights, diaper changes, and all things baby. While watching my newborn son acclimate himself to this new world, it struck me that I was watching machine learning in real life.
Like the rest of us, my son was born with only a basic set of instincts that I refer to as his base-level operating system. We didn’t have to teach him how to breathe, cry, sleep, pee and poop, etc. Those functions, along with the ability to learn, came hardcoded. Everything else he’s learning in real-time. As I watched this happen I began to recognize patterns to his learning that resembled ML techniques in form of Supervised, Unsupervised, and Reinforced Learning. Allow me to explain…
Supervised Learning (SL)
Newborn babies have no idea where or how to get food. Their stomach signals their brain that it’s time for a fill up and the brain has no idea how to resolve this situation. So, baby hits the panic button and then starts to scream and cry. That’s when baby’s Mom or a caretaker guides the baby through the feeding process. This is a structured process with a targeted outcome. At the conclusion of feeding baby learns how to react to the feeling of hunger in order to get food and how to recognize and consume the food from the correct food source. The baby will use this lesson to be able to differentiate it’s food source from a similarly shaped object such as a pacifier/binky or finger. Trust me, I tried several times to trick my son with a binky to buy time until my wife was ready to feed him and he was not having it (probably an early sign that he’s a genius, just my unbiased opinion).
The way babies learn to feed is a real-life example of how we use supervised machine learning. To follow the example, we are the mother/caretaker, the algorithm is the baby, and the food, binky, and anything else the baby puts in it’s mouth (which is everything he/she can) is the data. The goal is to achieve a targeted and specific outcome. Therefore, we provide it with the right food source to teach it how to identify the data needed to achieve that targeted outcome. After the guided sessions the algorithm will have learned how to differentiate it’s food source from a binky, finger, toe, etc.
The best use cases for supervised machine learning are ones where there’s a specific targeted outcome or value you would like to predict from your data. Supervised ML is the most commonly used ML training style with a wide variety of use cases. Use case examples include sentiment analysis, predicting customer churn, predicting employee performance, and internet/email fraud detection.
Unsupervised Learning (UL)
When my son was born there were a number of people in the room other than my wife and I – doctors, nurses, a midwife, and, because he was born a few weeks early, a specialized team from the preemie NICU that, thankfully, was not needed. To my son this was all new data. His mind had to begin classifying shapes, sounds, colors, smells, things that moved, things that didn’t, etc. Over time his mind began to differentiate and classify all this new data and he began recognizing familiar shapes – namely my wife and me. When other people would hold him the expression on his face would remain unchanged even though they were smiling at him. When they would hand him back to my wife or myself he would look at our faces and smile. We did not teach him to smile or to react a specific way when he sees us. His brain observed a pattern of familiarity, classified my wife and my face as favorable, and determined that smiling is the appropriate reaction when he sees us. He is not able to put labels on my wife and me such as “Parents” or “Mom and Dad” but he has learned to distinguish us from other people and chose to acknowledge us by smiling.
An unsupervised algorithm (“baby”) learns through observation, recognizing patterns, categorizing and differentiating data to determine the best response/action without guided input. However, these algorithms cannot determine the correct label for the groups it creates – i.e. Mom, Dad, Parents, etc.
Unsupervised machine learning is supposed to uncover previously unknown patterns in data but, since you don’t know what the outcomes should be, there is no way to determine the accuracy. That makes real-world applications of this model difficult but there are still some use cases that continue to be pursued such as visual recognition and robotics.
Reinforced Learning (RL)
Our oldest child, version 1.0 if you will, was an easy baby. He started sleeping through the night in his crib at 7-weeks-old. The latest iteration, version 2.0, has been much more challenging. It took a lot of trial and error to figure out what worked and what did not. When he cried we would go through all the usual culprits – feeding him, changing his diaper, holding him, rocking him, putting him in his car seat, driving him around, etc. – and each time we got it wrong he would cry harder to let us know we got it wrong. When we finally found the correct option he stopped crying and returned to a satisfied and happy baby. Over time, we learned from past interactions and became more effective at predicting the correct choice the first time.
In this example our baby is training us through reinforced learning. My wife and I are the algorithm. We make a predictive decision about which actions will satisfy our crying baby based on our analysis of the information at hand and he responds with positive or negative reinforcement. It’s a trial and error process where we have learned from our errors to achieve the desired outcome.
Examples of how Reinforced Learning is being used include tutoring systems and personalized learning, learning treatment policies in the medical sciences, and Salesforce used deep RL for abstractive text summarization (a technique for automatically generating summaries from text based on content “abstracted” from some original text document). You can view RL in action in this short video from Google DeepMind, who created a reinforcement learning program that plays old Atari’s video games.
Conclusion
Our son is a genius that already understands machine learning techniques and is intentionally using them to boost his cognitive development. All kidding aside, I am not surprised to find commonalities in the way my baby is learning and techniques used to train machines. It’s logical that machines would be taught to learn in similar ways to how people learn. There’s no reason to reinvent that wheel. And, even though ML has been around for over 50 years it is still very much in it’s infancy. ML hit a growth spurt over the last ten years or so that has seen it’s adoption rates sky rocket but we are still just scratching the surface of what can be done with this technology.
As with all new technology there are concerns that need to be addressed. The biggest concern for ML is whether or not we can trust the models to be free from bias. Companies like IBM are working to provide better insight and explanability to ML models that will help build trust. As a parent, I also relate to bias concerns. I want my kids to make their own choices but it is also my responsibility to teach them essential functions regarding how to behave, interact with others, etc. that can overlap with their ability to make their own choices. As my kids grow and their cognitive abilities expand they will begin processing the lessons my wife and I teach them in ways we will not be able to predict. They will derive unintended conclusions from our teachings that will require us to correct their understanding of the salient points of the lessons or reteach them altogether. This represents the challenges facing data scientists as ML progresses and becomes more widely utilized. A recent example of this is issue is Amazon’s recruiting tool that began showing bias against women and, as a result, had to be abandoned by the company. Despite this concern I am confident that men and women much smarter than myself will resolve bias through model explanability and transparency. Parenting bias, on the other hand, will likely continue in perpetuity.