Human Learning vs Machine Learning
12 July, 2021 by
Human Learning vs Machine Learning

Through some of our business system development projects and another type of work – STEAM education and training, I have the chance to explore how our brain remembers and how it is different from a machine learning system.

Let's start with today's AI image recognition system. A typical one is a supervised learning system: using some labeled data (for example, a set of sports car pictures and corresponding identification tags such as brand and model names) for learning. Machine learning systems need a little time to learn (much faster than human learning). In short, machine learning is a process of setting system parameters (technical term is equivalent to “hyperparameters”). Once the machine learning system has finished learning, it can recognize (predict) what a given new sports car picture will be.

For a machine learning system, it is effective because it learns and is set with some parameters to use. When we remove those learning materials from the system (i.e. the image files and labelled data used in the learning process), it will not affect its predictive ability. We can even replicate more ML systems that use the same parameter set.

Now we compare its learning process with humans. First of all, you can imagine a child memorizing through repeated learning. For example, a child listens to a song repeatedly, reads the same story book many times, or watches the same movie to remember some conversations. Can we separate or discard the memory of images but still recognize Thomas The Train? It seems that it will not happen in humans. We also cannot "transfer learning", what one person remembers (or has learnt), to another person.

People may think that computers can help us store a lot of knowledge, but memorizing is still a very important learning process to humans. The learning process not only helps us store memories in our brains, but is also a process that helps us build knowledge.

Humans can also use the memory of pictures, literary passages or musical melody as a basis to create something new of another knowledge domain – “Crossovers”. This has not yet happened in today's (or at least most) artificial intelligence systems.*

A teacher teaches a group of kids by singing, reading and what he or she does (以身作則). The knowledge we learn is useful for life until we lose some memories (that is about aging, accidental trauma, or disease that impairs memory.)

Will there be an artificial intelligence system that can learn from other machine systems without human involvement?

*Thinking topic: What the difference may be if the AI system uses Reinforcement Learning.

Share this post