2022 Reading List

Another year has passed in the blink of an eye. Shuttling between work and life, I have subjectively felt that my understanding of the world has shifted in various ways this year. That shift owes a great deal to the books I read over the course of the year. Compared to last year, I gradually read a great many books in psychology and philosophy. The initial motivation was the same as in previous years — to find enough inspiration for my own research — but as time went on I found myself increasingly captivated by philosophical argumentation. I share these here, hoping to meet people who have had similar experiences and insights to exchange thoughts with.

This year I grouped psychology and sociology books together under social sciences, ordered below by what I consider their importance. The one that had the greatest impact on my thinking was social choice theory. In earlier years I used to dismiss sociology outright, naively believing that social science was nothing more than the study of things that cannot be falsified by logic. It was not until recently, through the lens of my doctoral research, that I began thinking about the relationship between machine learning algorithms and human-labeled data, which led me to the philosophical mind-body dualism and then to theories about how societies emerge. Under orthodox Western rationalism, the fundamental problems of machine learning are framed as optimization problems. Yet any optimization problem can be conceived of as a single-objective or multi-objective optimization, which ultimately maps onto the currently fashionable formulation: the minimization of a loss function. When there is only one objective, the problem is quite simple — finding the extremum of a non-convex function, which in most cases can be addressed with gradient descent. But when the problem becomes multi-objective, we encounter the so-called Pareto Front, and the optimization problem reduces to one with no single answer. Any optimization process is then translated into a trade-off between different objectives, because on the Pareto Front no objective can be improved without sacrificing another. This idea is demonstrated with remarkable clarity in social choice theory, which proves with particular force — from a logical standpoint — the irreconcilability of social choice and individual choice, or more broadly, that society’s collective understanding does not match individual understanding, causing individual differences to disappear. All of this connects intimately with the foundational assumptions of statistics and philosophical epistemology, making it a genuinely fascinating domain.

Philosophy

Naturally, as I dug deeper into these sociological theories, I increasingly felt that every question was ultimately pulling me into the realm of philosophy. So this year I gradually began thinking about various grand philosophical questions: the limits of behaviorism, the relationship between free will and determinism, what nothingness means and what gives existence its meaning, and so on. In my earlier years I was a fervent advocate of behaviorism and held an instinctive contempt for certain social-science research methods — questionnaires, qualitative analysis, and the like — primarily because questionnaires seemed far too uncertain and riddled with limitations arising from participants' behavior, such as filling them out randomly. It was only when I began to engage seriously with philosophical inquiry that I started thinking carefully about how to better quantify human inner experience, and whether statistics can truly reveal the nature of things.

I have run a thought experiment with many people: suppose we have a new colleague, A, and we want to know what food they like best. From a behaviorist perspective, we would continuously observe colleague A’s food choices, accumulate statistics, and eventually arrive at a possible conclusion. Now imagine an extreme scenario: a year later, we find that colleague A has chosen from 10 different types of food, and by coincidence the probability of choosing each type is exactly equal. What conclusion can we draw? That colleague A has no food preference? That each of their food choices is random? We cannot conclude anything at all. Ironically, the questionnaire — which we regard as the least reliable instrument — turns out to be the most effective approach for this type of inquiry: why not just ask them what they like to eat? There is no need for the “foolish” behaviorist approach of observing something for a full year, only to end up with nothing. Of course, this thought experiment has several further versions, revolving around the limits of observational dimensions and the many confounding variables involved, which I will not go into here. It was later, when I began engaging with philosophical research on the problem of other minds, that my understanding of the nature of human behavior broadened considerably.

Mathematics and Statistics

The core thread running through my mathematical statistics reading this year was a single idea: the comparability of objects, or more generally, the rankability of multiple things. This line of thinking can actually be traced back to ten years ago, when I first entered university and encountered some foundational ideas in set theory — the notion of “order” on things, such as what kinds of sets can be sorted, partial orders, total orders, and so forth. That may sound abstract, so let me be concrete: the original impulse that first drew me deeper into these concepts was the study of human preference. Growing up, we all encounter the experience of being scored in one way or another — exam marks, rating a film out of five stars, and so on. This behavior can be classified as absolute scoring. Yet psychological research has long established that human cognition is not fully rational when assigning absolute scores to things (as famously demonstrated in works like Daniel Kahneman’s Thinking, Fast and Slow), which suggests we may need to assess things by preference ranking instead. The core idea then reduces to choosing between two alternatives. Following this line of reasoning further, from a rational standpoint we might consider a global preference relation among several objects, from which we can derive the single best choice for a given individual among multiple options. The subsequent complications are considerable, however: logically, we might hope that preferences over multiple objects should be transitive, but in practice human preferences violate transitivity quite easily, making it impossible to solve the optimization problem algorithmically, among other difficulties. In all of this one can glimpse certain philosophical puzzles — for instance, the mathematical study of comparison relations is typically developed in terms of binary relations, which carries more than a hint of dualism.

Engineering

Last come the engineering books, a category that, if I am honest, I find rather hard to get excited about these days. On one hand, having worked through so many engineering problems, I have accumulated a certain breadth of technical knowledge that lets me see straight through to what is actually driving a technology — whether it is pure rationality or a developer’s stubbornness. On the other hand, my understanding of computing has shifted from wanting to know “how to do it” to wanting to know “why do it at all.” A big part of the reason is that “how to do it” is now, for me, mostly a matter of either being able to build up an understanding of the approach immediately, or being able to judge right away whether the problem is even solvable. So if the “how” is no longer the obstacle, the more pressing questions become: what is the motivation behind this thing, how significant is its impact, and is it really worth my time?

If you’re interested in previous years' reading lists: Reading List Archive.

转眼又是一年，这一年里奔波在工作和生活之间，主观上感受到人的认知得到了不同程度上的转变。这一转变得益于自己在这一年里度过的各种读本，对比去年，这一年里我逐渐读了非常多的心理学和哲学相关的书籍，最早的原因和往年一样是为了给自己的研究有足够多的启发，但越到后来愈发的对哲学上的论辩着迷。我在这里分享出来，希望能够遇到有着类似经历共同感悟的人一起交流。

社科类

这一年里我将心理学和社会学相关书籍统一归类的到社会科学中，并按照我认为的重要程度进行下面的排序。其中对我认知影响最大的是涉及到的社会选择理论。早年我对社会学一类书籍嗤之以鼻，无知的认为社会科学无非是研究一些不可被逻辑证伪的东西。直到近年来因为博士研究的关系，我开始思考当机器学习算法和人类标注数据上的关系时，开始研究到哲学上的心物二元论，并进而接触到社会的产生机制等学说。在正统的西方理性主义下，机器学习的基本问题被定义到了优化问题的框架下，然而任何优化问题都可以被考虑为一个单目标或者多目标的优化问题，最终转化到如今比较热门的说法：损失函数的最小化问题。当我们只有一个目标时，问题非常简单，即求非凸函数的极值，这样的值在大部分情况下是可以梯度下降的。但当问题来到了多目标时，我们就会遇到所谓的 Pareto Front，这时优化问题会被归结到一个没有单一答案的问题，任何优化过程都被转化为不同优化目标之间的权衡，因为在 Pareto Front 上，任何目标都无法在不牺牲其他目标的情况下获得提升。这一思想在社会选择理论中被展现的淋漓尽致，并且相对突出的从逻辑的角度证明了社会选择和个人选择之间的不可调和性，或者更进一步说，社会的公共认知存在与个人认知的不匹配，从而导致个体差异的消失。而这些内容又紧密的与统计学的基本假设以及哲学认知论紧密相连，非常有趣。

哲学类

显然，随着对这些社会学理论的深入挖掘，我愈发的感受到所有的问题最终都把我带入了哲学的范畴，所以这一年里我逐渐的开始思考各类哲学上的大问题，比如行为主义的局限性、自由意志与命定论的关系、虚无和存在意义是什么等等。早年的我其实是非常推崇行为主义而从骨子里鄙视社科类的一些研究方法，例如调查问卷、定性分析等等，原因通常都是问卷这种东西的不确定性太大，存在相当多受到被试影响的局限性，例如乱填。指导开始深入接触哲学思辨时，才开始深入思考如何才能更好的量化人的心灵体验，以及统计学是否真的能揭示事物的本质。在这里我跟很多人做过一个思想实验，说我们有一个新来的同事A，我们希望知道他最喜欢吃的食物什么。从行为主义的角度来说，我们会不断的对同事A对食物的选择进行观察，统计并最终得出一个可能的结论。这时，我们假设一个极端的情况：即一年之后，我们发现同事A一共选择过10类不同的食物，但巧的是，每个食物选择的概率是相同的。这时候我们能够总结出什么结论呢？同事A对食物没有偏好吗？同事A每次的食物选择都是随机的吗？我们什么也总结不了。相反，我们认为最不靠谱的问卷，在调查这类问题的事情上反而是最有效的：直接问他喜欢吃什么不就完了吗？甚至不需要像行为主义那样"愚蠢"的对某件事情进行观察长达一年的时间并且到头来可能什么都得不到。当然了，这个思想实验还有后续的好几个版本，围绕着这个观察纬度的局限性，存在很多的混杂变量等等的讨论，这里我就不细说了。后来直到我开始接触哲学上对他心问题的研究，扩宽了我对人类行为本质的认知。

数学统计类

这一年我读数理统计类的书籍核心其实围绕的是一件事情：对象的可比较性，或者更一般的说多个事物之间的可排名性。这些阅读甚至可以追述到十年前我刚进大学的时候接触到集合论的一些基本观点。谈到事物的"序"，例如什么样的集合是可以被排序的，偏序、全序等等。当然，这样说可能相对抽象，最开始让我进一步接触这些概念的原动力是对人类偏好的研究。在成长的过程中，我们多多少少都会经历被评分的体验，比如考试多少分，这部电影我给几分等等，这类行为可以被归类为绝对评分。然而，心理学研究早就已经证实了人的认知在事物进行绝对评分上并不能完全的理性（比较有名的著作例如丹尼尔卡尼曼的《思考快与慢》），从而启发了我们可能需要对事物进行偏好评级，核心的思想便转化为两者之间选一个。如果按照这个思路进一步走下去，从理性的角度上，我们可以考虑若干个事物时间的一个整体的便好关系，从而得出某个独立个体在多个选择之间的一个最佳选择。当然，这个问题后续还很复杂，例如逻辑上我们可能会希望多个事物的偏好应当具有传递性，但事实上对于人类而言这种偏好的传递性很容易违反，从而导致无法从算法上对优化问题求解等等。这里其实或多或少又看到了一些哲学上的迷思，例如数学上对比较关系的研究通常以二元关系展开，或多或少有一些二元论的意味。

工程类

最后就是工程类了，工程类的书籍对于今天的我反而比较提不起兴趣，一方面是因为做过太多的工程类的问题，对技术了解积累了一定的广度，能够直接看穿一个技术背后到底是什么在驱动，到底是纯粹的理性还是开发人员的固执等等。另一方面，我对计算机技术的理解从原来的希望了解"如何做"变成了"为什么要做"，其中一个很重要的原因是"如何做"这一问题对于现在的我，更多的是要么能立刻建立起如何做的认知、要么就就是能够立刻判断这个问题是否可解。所以如果如何做已经不是问题，更多的问题则是做这件事情背后的动机是什么，影响有多大，我到底要不要在这件事情上花时间。

如果对往年的阅读清单感兴趣的话，可以查看这里：往年阅读清单合集。

Social Sciences

Philosophy

Mathematics and Statistics

Engineering

社科类

哲学类

数学统计类

工程类