Updated: Oct 13, 2019
A while back in March this year I had the pleasure of sharing some knowledge on neural reinforcement learning from my studies at UCL with my colleagues at IBM. I started the talk with defining some basic concepts: what is intelligence, what is learning, what types of learning exist? When you think about these basic questions more deeply before trying to relate this to how you would operationalize that in a learning computer system, it already gives you some insight into how complex these basic answers can be. People have different understandings and definitions of what true intelligence is and how they would go about describing what types of intelligence or learning there are.
To define learning, I went with the Oxford dictionary’s definition of learning: “the acquisition of knowledge or skills through study, experience, or being taught”. Now coincidentally when I read that definition, I thought the three aspects of learning as defined here can be directly related to the major three distinctions we make in machine learning (a big subdomain within the realm of artificial intelligence that studies how machines can be taught how to learn – my definition of it). It almost seemed too good to be true...
When we have acquisition of knowledge or skills through 1. Study, 2. Experience or 3. Being taught and 1. unsupervised, 2. reinforcement and 3. supervised learning as three main ways of machine learning, you can see what I’m getting at.
The way humans learn by studying books and observation could be seen as a largely unsupervised, independent and “intrinsic” way of learning. For instance, you pick up a book about human anatomy and from processing the mush of text and descriptive pictures in there you gradually build up your internal understanding of the human body.
When I think of learning through experience, I think mainly of the “trial-and-error” type of learning. We constantly come across new situations in life where we try something and evaluate the result of that afterwards. Some may say that is exactly what makes life a b*tch sometimes. This is the same principle we apply in reinforcement learning: the computer agent gets input information X about the environment it is in, tries action A, evaluates outcome R and goes to a new environment – and so on until after sufficient numbers of trial and error it learns what the optimal action is to take in which situation. (A slight difference though is that it does not feel like a hassle to a computer playing millions of games continuously to eventually beat world champions in two-player games.)
Learning from being taught may be the easiest to imagine in a machine. Like a teacher that shows you what is true and false about the world (according to the knowledge s/he has at that point in time) and a parent or caregiver that tells you what is right or wrong (notice this is entirely different from true or false), a machine can be “told” through programming what is what, too. From the few examples that it gets, it learns to identify the characteristics and patterns that form a certain entity that you try to teach it about.
Opposed to unsupervised learning, you may think of supervised learning as “showing someone/the computer what it needs to recognize”, whereas unsupervised learning would be akin to “letting someone/the computer figure out a pattern themselves". A powerful statistical method that has been around for more than a century by now called “principal components analysis” could therefore be regarded as an unsupervised learning method. For anyone who has used this in their own studies: isn’t that great to know you were applying a "primitive" form of machine learning “already”?
Apart from the technical intricacies of current reinvigorated AI research developments, I love thinking about what it would really mean to have artificial general intelligence (AGI) and moreover, the potential applications of AI for people and planet. There are so many questions you can derive from the similarities I just described between human and machine learning. For instance, is a very intelligent person we know especially good at “unsupervised” learning? Would the same principle hold for designing AGI? Could there be other forms of learning that we may have overlooked? What other cognitive capabilities make up an intelligent person that we want to see reflected in an “intelligent computer”?
As you may tell from this very first finished blog post, AI connects so many different fields of studies. From lexicography, linguistics and philosophy to psychology, neuroscience and computer science. With such a broad palette of topics I want to discuss, I hope to bring you some new insights with each new post. Yet most of all, challenge your thoughts on hypes (such as AI) and other worldly matter as well.
Please feel free to share what you think in the comment section below, subscribe to this website and/or share this post on social media. :)
Until the next one (aiming for biweekly weekend reads)!