Actions & Rewards:
How Mothers Are the Original Reinforcement Learning Experts
Parenthood is arguably the most difficult job a person will take on in their lifetime. The stakes are high, the pressure is immense and the hope being that all your time, dedication and sacrifice will result in the evolution of an independent and self-sufficient adult is the ultimate goal.
When most people look back on their childhood, one figure tends to stand out from the rest: moms. The women who (somehow) were able to take the place of so many others, yet remained an irreplaceable influence. When we discussed how we could honour them on Mother’s Day, we decided to focus on their role as teachers, rather than the more obvious emotional caregivers.
Believe it or not, but teaching a baby and training a machine with reinforcement learning isn’t as different as you might expect.
Brains & Computers
If we are going to imply that machines and children can learn in a similar way, we’re going to need to draw parallels between the two. Here’s where we begin: our brain, which is a network of neurons (and is centralized in our nervous system) processes information by sending electrochemical signals between synapses. The modern paradigm of AI, called deep learning, uses many layers of artificial neurons connected to each other. The pattern in which neurons are connected allows them to learn relationships in their environment.
There has been some research that implies that our brain is similar to a computer. Both understand the world by mapping relationships between inputs and outputs. There is an ability to take internal and external stimuli as input, process them and return an output. Decision making, problem solving and identifying patterns are all a result of this process and both the brain and computers can achieve this.
The belief is the more you can expose both of these entities to different things, the easier it will be for it to absorb information, train and ultimately, form intelligence.
Trial & Error
Now that we have the brain, the next step is to utilize it. It’s strange to think that at the very beginning of our existence we knew next to nothing. Somehow, with no skills or understanding of language, we were able to evolve into full-functioning human beings. A big part of our development happens through trial and error (with a little help from mom).
Reinforcement learning is a branch of Machine Learning inspired by behaviorist psychology. At its core, it is quite simple: an agent (ex: a child) is asked to take an action in a specific environment. This goal-oriented learning allows the agent to receive a reward once the action is completed correctly. Positive reinforcement strengthens a successful decision, while negative reinforcement does the opposite. Through this process, the hope is a person will eventually discover which actions will yield the maximum reward without being explicitly told what they need to do.
Machines vs Humans
Machines and humans are brought into this world in similar circumstances, since they have no input on what things actually are, therefore they must be taught. When your mother was teaching you the basics, she put labels on things in order to help you sort out the differences between various objects or people. Through repetition, human babies are able to learn, generalize these labels easily and apply them to new situations.
Machines are different because they need massive data sets to learn. When mom pointed at a dog, you were able to register the animal and recognize different breeds without a clear explanation. Machines, no matter how advanced, still need a lot more help and context. The data needs to have variety such as different angles and colors in order for it to slowly process that “dog” can look very different from one picture to the next. In order to speed up the process, the algorithm learns the same way a human does, being asked to take an action and given a virtual high-five when they achieve their goal.
Changing Tides?
This narrative seems to be changing thanks to experts at the Rice University and Baylor College of Medicine. Recently, they created a neural network that mimics the human visual cortex. As a result, they were able to train their model faster and with less annotated data.
At the end of the day, mother’s have more than reinforcement learning in their arsenal of parenting tools. They have an instinct that is impossible to replicate, the capability for love and devotion that is second to none, but mostly, thousands of years of knowledge that has been passed down from generation to generation.
Once we reach the point where learning and intelligence can be as simple as a mother teaching her child, the potential for AI will truly be limitless.
Interested in starting your AI journey? Contact us today.
