Human learning & intelligence
I recently finished reading Prof. Song-Chun Zhu’s 60K+ character article on Artificial Intelligence (AI) over the holidays. One part of Prof. Zhu’s essay that caught my attention was Section 4, where he discusses the cognitive framework of AI research.
In the prior section, Prof. Zhu gives an extended example of a common crow that not only learned how to take advantage of passing cars to crack open nuts for the bird to eat, but also intentionally dropping the nut on pedestrian crosswalks so that the bird won’t be run over by vehicles as it swoops down to eat the nut. Prof. Zhu postulates that the crow’s intelligence is a direct product of both its physical environment and its will to survive. He calls these 物理环境客观的现实与因果链条 (objective reality of the physical environment and the causal chain) and 智能物种与生俱来的任务与价值链条 (inherent task of intelligent species and value chain). An agent that exhibits these two conditions will be able to live autonomously when placed in a given environment and social group. In fact, the agent will learn the following: “认识世界、利用世界、改造世界”.
Loosely translated, it means:
“To know the world, to use the world, and to transform the world.”
When I read this conclusion, I was struck by the elegance of his intelligence equation. With all the hype around machine learning, deep learning, and AI today, we often forget how amazing human intelligence is, and how far away our current progress is in AI when it comes to modeling our own intelligence.
Machine learning and deep learning continues to follow a big data, small task paradigm: feed terabytes of data to a learning algorithm to generate a model that can accomplish a small, specific task very well (ex. recognizing human faces). Human intelligence, on the other hand, becomes increasingly abstract and philosophical as we age and grow. Modeling human intelligence in an artificial agent requires much more than a big data, small task approach; Prof. Zhu’s idea that an agent should be able to know, use, and transform the world closely models machine intelligence after human intelligence. I hope to outline the beauty of our learning process and our philosophy on life, to further illustrate that human intelligence is truly a beautiful, non-trivial phenomenon.
I want to introduce some insights made by 马未都 (Ma WeiDu), the founder and current curator of China’s 观复 (Guan Fu) museum. He spoke on a TV show about his experience with learning, which was truly fascinating! His ideas parallel the observations Prof. Zhu made on modeling intelligence, so I will try to unify the two.
At 12:41 in the video, he makes the following observations about the 3 stages of learning in life:
“Learning in a lifetime has 3 stages: Ages 5–15: 诵读 — Memorization. Ages 15–25: 学贯 — Interdisciplinary understanding. Ages 25–35+: 渉猎 — Learn everything.”
Let’s dive into these 3 stages of learning:
诵读 (Memorization) - When we attended elementary school through middle school and early parts of high school, we were learning ideas and acquiring knowledge of various subjects independently: Math, Geography, and English had no relations. We were expected to learn, but we were not expected to understand. However, all the multiplication tables, world maps, and grammar rules built the foundations for the second stage of learning. This first stage allows you to know the world.
学贯 (Interdisciplinary understanding) - As we go through high school, college, and potentially graduate school, we experience the critical stage of connecting all the siloed knowledge we acquire and observing that there exists relationships between subjects previously thought of as disparate. For example, engineering students are expected to take at least one ethics class in undergrad, learning about laws, philosophy, and moral consequences of working in the field. My engineering ethics class consisted of a lot of reading, and a lot of writing — almost similar to an English class. On the other hand, I took a History of Electronic Dance Music course (also with plenty of reading and writing) that inspired me to code this web-based interactive essay about how The Chainsmokers prepare their DJ set (I majored in Computer Science). In college, I took advantage of the opportunities to learn by combining various disciplines instead of studying them in isolation. With interdisciplinary understanding, you can use the world to help you learn.
渉猎 (Learn everything) - In this stage, we begin to recognize the rewards of learning new things. Generally, college graduates are more open-minded about learning new ideas than when they entered college, and they start to have a yearning for lifelong learning. Read subjects you are not familiar with. Talk with someone outside of your discipline field. While you may argue with people on ideas you disagree with, you also have the ability to step back a little, and understand the world from their point of view. Embracing this third stage of learning is instrumental for building the emotional intelligence critical to being a valuable contributor to our society today— empathy and sympathy. These two qualities allow you to influence others and transform the world.
It is clear that as we progress through each learning stage, we are driven by different factors to learn: memorization allowed us to attempt singular tasks such as solving arithmetic equations, but lifelong learning allowed us to attempt complicated, multi-layered tasks such as using mathematical modeling to predict trends for global warming, which in turn can be used as supporting evidence for change in politics, law, and corporate policy, with the ultimate goal of protecting Planet Earth. Artificial agents lack this deep level of cognition!
In addition to human learning, there also exists the ideas of the human purpose. Prof. Zhu mentions that intelligent agents should all be driven by a causal chain and a value chain. The value chain is our own utility function that dictates our preferences over the decisions and choices we make. We can think of our utility function’s local maximums as the states in life we desire; we can think of these states in life as goals that we set for ourselves to achieve.
We set small goals every day (and maybe unrealistic goals every year on New Year’s Eve), but what about our lifelong goals? Are there some universal goals that humanity can relate to?
At 16:12 in the video, 马未都 makes the following observations about the 3 goals we have in life:
“We have 3 goals in a lifetime (achieved sequentially): Goal #1: 趋利 — Take care of yourself. Goal #2: 趋名 — Build and value your reputation. Goal #3: 趋静 — Seek peace.”
趋利 (Take care of yourself) - This is the first and foremost goal that is everyone’s priority. In a way, taking care of yourself means surviving autonomously in this society. We do not want to be an infinite burden to anyone, so we work hard to not only live, but live better every day. Once you do that, you can help others, such as your family, loved ones, or those in need, so that they may reach the same first goal.
趋名 (Build and value your reputation) - Once we can take care of ourselves, we start to become more aware about our actions and their consequences. Valuing reputation is not the same as striving to be a famous celebrity; instead, we simply ask ourselves the question “What do you want your legacy to be?” We begin to build values of accountability, responsibility, and trust. With the prior knowledge that actions have consequences, we program ourselves to distinguish between right and wrong, to ask ourselves if it’s worth ruining our reputation to do something shameful, and to generate a moral compass.
趋静 (Seek peace) - There’s a cliché quote out there that says “money can’t buy happiness”. We can derive from this quote the importance of having intrinsic values. The Mission published a powerful article about the rise of children mental disorders and speculates that there is a shift towards extrinsic goals instead of intrinsic goals that contributes to this pattern. Some background on these ideas (from the article, quoted below):
“Intrinsic goals are those that have to do with one’s own development as a person — such as becoming competent in endeavors of one’s choosing and developing a meaningful philosophy of life. Extrinsic goals, on the other hand, are those that have to do with material rewards and other people’s judgments. They include goals of high income, status, and good looks.”
The final stage of the human experience is essentially you asking yourself the question, “Have I found all the intrinsic goals that I seek, and do I value them over my extrinsic goals?”. After achieving survival and maintaining reputation, the next step is reaching inner peace. We can always stress our physical bodies, but we should never stress our minds.
These ideas that 马未都 introduced— these philosophies — have far-reaching roots in our society, culture, values, and more generally, our common sense. The end game of modeling human intelligence in AI is essentially to build an agent that exhibits common sense. Taken straight from Wikipedia:
“Common sense is sound practical judgment concerning everyday matters, or a basic ability to perceive, understand, and judge that is shared by (“common to”) nearly all people.”
Human learning and intelligence are complex phenomenons, and we are only beginning to scratch the surface of modeling these phenomenons for an artificial agent. There is, without a doubt, an enormous potential for revolutionary development in the next decade (or less). Exciting times lie ahead in the study of intelligence!