On Intelligence, it is a masterpiece anyway. Despite the fact that this book on the theories of intelligence is filled with Jeff Hawkins’ excessive self-esteem and seems to be a little bluffing, the scope it provides us with as well as its clear arguments and narration still make it an extraordinary piece of writing. As the Nobel laureate James D. Watson commented, “On Intelligence lays out the framework for understanding the brain”. At least for me, many things I used to take for granted started to make sense after reading this book.
In the following essay, I am going to identify the most interesting thought of the author from my perspective and a proposition I am skeptical about, and to give the reasons behind.
II. The Most Interesting Thought: the Way Cortex Learns
Keyword behind the whole book is “prediction”, which is the proof of intelligence instead of behaviors. Neocortex is the part of our brain that actually takes this responsibility. In Chapter 6, there is a section for cortex learning methods which I found both original and meaningful. Its basic idea is that representations of objects move down the cortical hierarchy with repetition.
A. Brief Introduction of Author’s Opinion
We are born with almost blank cortexes. All we know about the world is acquired, and the formations of pattern classification and sequences turn out to be the most fundamental components of this learning process. Mr. Hawkins held the belief that during repetitive learning, representations of objects move down the cortical hierarchy. When we get to understand a higher-structure of a subject, we actually do not need to pay much attention to the simple basics concentrated on before.
B. My Perceptions of the Argument
This argument makes a lot of sense to me. As far as I am concerned, it reveals the principle of human’s knowledge acquisition and accords very well with the prediction system, or cortical hierarchy theory, described in preceding sections.
Throughout our lifetime, learning is not what happens just in schools. After acquisition of sophisticated skills, I usually find it hard to recall the difficulty at the beginning. The truth is the individual details I used to spend long time to react to now require little concentration. With Hawkins’ moving-down-hierarchy assumption many things fall in place. Our memories are stored in cortex in the form of sequences. When we learn something, we actually build patterns and sequences in our 6-layer cortex. Higher layers store sequences of sequences of lower layers. When we learn something completely new, we store the tasks higher up the hierarchy, which requires longer time to react to because of invoking of lower hierarchies. With repetition the abstract representations move down to lower hierarchies, then frees the higher ones to learn more and saves the time.
Therefore, the argument corresponds to the hierarchy theory closely. In other words, it is also a strong proof of the existence of cortical hierarchy.
C. An Example
The first supporting evidence of this argument stroke my mind is the so-called “10,000-Hour Rule”.
A study done by the psychologist K in the 1990s found three different groups of violinists at Berlin’s elite Academy of Music: the stars, merely “good” and only likely to be teachers, with three different amounts of practicing time: 10,000 hours, 8,000 hours and 4,000 hours. Of these violinists, the psychologists failed to find any so-called “naturals”. It turns out that successful people actually work harder, and 10,000 is a magic number equaling greatness.
In this case, it is practice that makes an expert, or “repetitive learning”. They can play excellent music just because of repetitions. If Hawkins’ hypothesis is true, over time the violinists move the established knowledge down the hierarchy and free the higher layers to learn more higher-order structure. The brains of those experts are able to see structure of structure and patterns of patterns beyond what others do. Thus, as far as I am concerned, Hawkins’ theory can be a possible argument to explain the 10,000 Hours-Rule.
III. Proposition I am Skeptical about: Separated Intelligence and Emotional Drives
Even though I find most of the ideas of Hawkins original and insightful, there are still many propositions in the book that I do not completely agree. For example, in a section called “Should we build intelligent machines” in the 8th chapter, Hawkins attributed the fear for intelligent machines may work against us to a conflation of intelligence with the emotional drives of the old brain. He implied that these two parts are different and can be completely divided so that intelligent machines can have only intelligence and no emotions, which I consider as being too idealistic and optimistic.
A. Hawkins’ Proposition
As Hawkins put it in Chapter 5, old brain regulates blood pressure, hunger, sex, emotions and many aspects of movement. But mammals are more intelligent because of their neocortices, which allow them to use old brain behaviors more intelligently. Hence, he asserted that understanding of the neocortex is the key to unlocking the mystery of intelligence. And after expanding his hierarchy hypothesis in several chapters, he drew the conclusion that machines can be intelligent without emotion drives.
From my point of view, Mr. Hawkins seems to be a little bit self-contradicting on this question.
Given the intelligence of neocortex exists in the assumption that it uses old brain behaviors intelligently, how can a machine be intelligent without the existence of a part corresponding to old brain? The author said it is prediction, not behavior, that proves intelligence. I agree. However, that does not mean an intelligent machine does not need a behavior part to interact with the world. To be specific, old brain behaviors are the fundamentals of neocortical algorithm. And a machine can hardly possibly have well-functioning higher control units without fundamentals.
Defenders of Hawkins’ idea may argue that we only need to preserve some behaviors of the old brain on machines, such as those movement behaviors. Fair enough. But how? If the whole neocortical algorithm is built on the basement of an old brain part, just like the real cortex does, would it still function perfectly well with an incomplete old brain? Till now there is still a question mark on it.
C. A Counterexample
Consider the creation of art and literature. Do they the result from pure emotions or pure intelligence, i.e. uncommon or higher-level predictions like Hawkins indicated in Chapter 7? Of course they are the combination of both. At least neocortex invokes emotional drives during this creation, if we follow Hawkins’ assumptions. Therefore, if we want to build a machine of high intelligence level, which means the machine also has some degree of creation, we need to preserve the emotional behaviors of the old brain.
Great works are always controversial. And limited by personal experience and knowledge, my opinions may not necessarily be flawless. Nevertheless, all the agreements and disagreements, author’s being too categorical or too much self-esteem, do not bother On Intelligence being an insightful must read for all artificial intelligence enthusiasts, and even everyone living in the 21st century.
With the scope offered by this book, I got a brand new understanding of intelligence and many things related. Such foresight is essential. As electronic engineers we are obliged to know where we are going.