Do machines learn like kids or like dogs?
Understanding the similarities between machine learning and animal training shed lights on the possibilities and limitations of AI
Machine learning is not the same thing as human learning. I find that the distinction often gets lost in discussions about artificial intelligence. Most machine learning professionals don’t know anything about wetware learning, and most educational psychologists and pedagogy professionals don’t know anything about how machines learn. I think that a sad state of affairs. Even a high-level understanding of the difference between human and machine learning can clarify a lot of common misconceptions.
First of all, we need to understand that humans learn in many different ways. The topic is very vast, but let’s list some of the essential learning types.
Play is a behavior which helps humans of any age learn things that improve performance in similar future situations. It’s also a common learning method for many mammals and birds, but it’s unheard of in the machine realm.
Active learning is the way good students take control of their learning experience. They might self-direct themselves toward topics of interest, read books, or watch informative videos and maybe practice their newfound knowledge with others or with formal exercises. That is all standard A student behavior. Not even AlphaGo on its best day can choose to learn another game.
Episodic learning is the type of learning that happens someone that has been bitten by a dog start fearing dogs. My dog bites my phone regularly, but Apple hasn’t taught Siri to be afraid of dogs, so my iPhone stays fearless.
Habituation is a type of non-associative learning by which humans — and animals — create habits. My computer certainly looks like it’s developing a blue screen habit. I’ve been told that it’s all in my mind and that computers don’t have habits.
Enculturation is the learning process by which people learn the values and behaviors of their cultures by being immersed in it. Software programs have no opportunities to develop a culture of their own.
There is one type of learning that is common to both animals and machine, or at least similar enough. Associative learning is the way people learn an association between two stimuli. The classic example is Pavlov’s experience of conditioning his dogs. He used to ring a bell before feeding his dogs. Over time, the dogs started to salivate when the bell rang, even if the food didn’t come. What makes the approach work is the use of a reward or reinforcement, in that case, dog food. Using operant conditioning, I’ve trained my dog to come when called, sit on command, and wait for her food. Using the same approach, my cat has learned to jump on the dinner table and try to steal our food when we are eating. Poor training can give unexpected learning results.
There are a few different methods to make machines learn, but most are related to associative learning. The model learns to associate data points to particular outcomes. A cost function serves as a reward/punishment mechanism.
For example, in the case of a neural network trained with gradient descent, the model is tasked with learning a representation mapping from input data points to output labels. For each data point, the learning algorithm updates the model’s parameters in the direction that best approximate the desired output label. At each iteration, the model “knows” that it’s getting better because the cost function is decreasing.
That process is very similar to the way animal trainers work: at first, when the command is issued, give rewards for behaviors that are approximately right. Then give better rewards when the response gets closer to the ideal, and less or worse rewards when behaviors are diverging from the perfect state.
Machine learning software can be trained to do almost anything in that way, but unsurprisingly with limited “real” intelligence. Think of the machines as having the learning abilities of a dog, or a small child. Would you let a dog manage your marketing strategy without supervision? Your pricing strategy? Client support? You probably shouldn’t.
However, consider the following: if you want to do bomb checks in an airport, would you prefer:
A. A team of humans
B. A team of dogs
C. A team of dogs with human handlers.
In the case of bomb checks, the performance of mixed dogs and human teams is an order of magnitude superior to the results of either humans or dogs alone. Similarly, the multi-species teams are great for finding victims after a natural disaster, finding truffles in the woods, hunting animals, and many other situations. Machine learning AI software, as of 2019, has similar limitations to our canine friends. The machines can significantly improve performance at many tasks when skilled human handlers are involved. However, the abilities of software to act autonomously are still considerably limited.
What do you think? Does the way software learns limit its capabilities? Could software be improved by using other learning strategies?