I saw this below tweet a few days ago, and it invoked an experience I went through as a data scientist in the consulting industry.

It’s an experience that made me realize that I can’t stay in this industry. I’m working as a data scientist at a Thai consulting company. I’m leaving at the end of June to go study in a PhD program in cognitive and information sciences at UC Merced.

Now the story

My company worked with a think tank last year. The think tank was very ambitious. We worked on 4 public policy projects (4 different topics) to produce recommendations to shift the developmental paradigm of the country, Thailand. For every project, we had to deliver a visualization-focused content that they called “data story” and a formal report of analysis results that include recommendations for policy changes. We had up to 6 people in the team. “Up to” because some people had to work on multiple projects at the same time. And there was not a moment where everyone was available simultaneously to focus on a single project. It was a lot of divide and conquer. And do you know how much time we had to complete these 4 projects? It was 6 months! The progress they required of us was sequential. We had to finish the first project (let’s say on air pollution) within the first 2 months of the engagement period despite all the challenges like poorly formulated research questions or low quality data. We then started working on the next project on a different topic right after that, which we had to finish within 2-3 weeks Lol. Every time we delivered either a data story or a report, they provided us with endless (not just one-off) comments about the content, the writings, and even the analyses after a particular project was supposed to be over already. And we had to address all their comments. That’s why our team was stretched so thin. So my heuristic was to do as minimally as possible to address them considering the time limit. It kept going on like this for the entire period of the engagement. And for the entire period, I felt completely burnt out not only because of the tight timeline and overwhelming amount of workload but also because of the thought that I didn’t believe in any of the analysis results or recommendations.

To rub salt into the wound, my project manager and my boss didn’t even think the deadline was that unreasonable. They thought maybe it was a bit too tight but maybe 1 more month for each project would be enough extension. It was such a slap in the face for me.

I’m aware that my goal or value is not well aligned with the consulting industry. But this “collaboration” made it very distinct (“collaboration” because the think tank barely helped us with data and everything). It’s from this engagement that made me realize that time is everything too. People might not realize how time constraints actually limit or shape your thinking. If you give me a month to finish a project, I’ll automatically (or even unconsciously) pick an approach that I can do in a month and deliver something. That approach will probably be sloppy af, but it’ll give me something to deliver. In the current consulting culture, 1–3 months of engagement is very common. But I can never imagine a scenario in which that time length is enough for any meaningful work to blossom, especially for public policy. This could partly be due to level of expertise of our team as well, but I can imagine an alternate universe where our team is much more technically competent, but the outcome would still be the same due to the lack of understanding of how good quality research should be conducted to inform decision making on the side of project owners and managers.

The technical aspect compared against time constraints

It’s nice that we have an engineer who can take care of setting up a cloud environment and importing huge amount of data from clients for us. But that’s still not enough. I’m a DAG user for causal inferences. I try to encourage my team to adopt DAGs all the time not for the sake of using it but it will allow us to theoretically criticize our own models and deliver a more transparent result. It doesn’t really take off probably because cultural and behavioral change is hard. But even with DAGs in our inferential toolkit, it will still be difficult for us because we don’t necessarily have expertise on the project we work on. Sometimes, we have to read more papers. Sometimes, we need to ask a lot about the domain subject from our clients. But even the clients may not provide enough theoretical foundation for us to build any reasonable DAGs probably because theoreticians are not selected for in industry. Bummer right? So even just trying to grasp the nature of the problem domain itself will already take quite some time. Then data processing and cleaning. Then modeling. Then writing report. Then potentially revising the whole thing depending on how much time we have left. I really find it difficult to do all these things thoroughly in a period of 1–3 months. And to top that, for a single engagement, there are actually multiple projects we have to carry out simultaneously. But to not entirely vilify my company, not every engagement is as bad as the one with the think tank. I just think that the whole ecosystem of the consulting industry isn’t facilitative of rigorous research or analytics work. (This is based on my experience at a consulting company in Thailand, so takes it with a grain of salt!)

Why can’t we change the culture?

Mostly because of money I guess. You may not be able to compete and survive if other consulting companies deliver something that makes clients satisfied within a shorter period of time. Longer engagement and contract also mean higher price. And with a capitalist mindset, you will not want to spend “too much” money on things that might not give you enough financial return. As a result, consulting companies will try to optimize the process more and more. As far as I can tell, the trajectory is that our team is expected to use less and less time to finish a project, which means it’ll probably get sloppier and sloppier. And another thing is it’s really hard to vouch for scientific rigor when that’s not a value of the organization to begin with. This is why I am CERTAIN that I can’t stay in this industry. I’m not even sure if I can work satisfactorily in any other private sector. As long as you value profits over rigor and people’s well-beings, I can’t in a million years see our relationship thrive. Every time I deliver something that I don’t believe in, a piece of my soul dies. The money is actually decent, and I have a healthy work-life balance. But I still feel burnt out and really can’t bear it anymore. So, I’ll go out to do my roughly 5-year PhD and enjoy the hell out of it.

Academia vs. Industry

I’m aware more people are leaving academia because of the corrupted incentive structure to do rigorous research. And I can empathize with that because, to be honest, the incentive structure of academia or the research circle in Thailand is very perverse too. But my experience also shows me that the private sector doesn’t necessarily have the best interest of science in mind either. If you’re wondering whether you should keep doing research in academia or go out and do it in industry, I really don’t have an answer for you. I can only tell a story from my perspective. It might not even be a matter of academia vs. non-academia anymore. It might just be about which organization can create an environment that facilitates good research. I have my bias against letting the private sector takes control of the direction of science. But that’s a different conversation altogether. At the end of it all, we need to weigh pros and cons that best align with our personal values.

In the past few years, I’ve been studying statistical and causal inference while working as a data scientist. I realize that most people are taught frequentist statistics to perform statistical inferences. But over time, I find the interpretation of p-values and confidence intervals unintuitive and, most of the time, renders them useless or impractical. In this article, I select one scenario that can result in meaningless confidence intervals and p-values to demonstrate my point.

In this article, I will work with the Neyman-Pearson definition of confidence intervals. That is if you repeatedly sample from the same population and calculate 95% confidence intervals, approximately 95% of all confidence intervals will contain the true parameter value. Let’s say that we tracked exam scores of 100 students over time. Each time before an exam, students either took a tutoring class to boost their exam scores or they didn’t. We want to know whether the tutoring class was effective or not. Assuming that exams were conditionally independent of each other given the student, we will test the effect of the tutoring class by conducting a Welch’s t-test between exam scores when students took a tutoring class and exam scores when students didn’t take a tutoring class (Note that fitting a linear regression model with a variable indicating whether the student took a tutoring class or not as a predictor variable and exam scores as the dependent variable will return the same result).

Suppose that we know the true data generating process, and it turns out that the tutoring class affected the students’ exam scores differently for each of the students. Some students’ scores increased with the tutoring class. Some students’ scores decreased. And some students’ scores remained approximately the same. But overall, the average effect was 0. The effect of the tutoring class on students was drawn from a gaussian distribution: t ~ Gaussian(0, 5).

I simulated 100 studies and calculated their respective 95% confidence intervals for different number of data points per student. The results are shown in the figures above.

According to the definition of confidence intervals, roughly 95 intervals should contain the true parameter, which is 0. However, most of the calculated confidence intervals don’t contain 0 at all (false statistical significance). Moreover, with more number of data points per student (20, 50, and 100 from left to right figure), there is less number of confidence intervals that actually contain 0. This happens because we ignore the variation in the effect of the tutoring class on students resulting in more overconfidence intervals with larger sample sizes. Let’s try removing the variation between student effect and redo the simulation again.

Now, most of the confidence intervals contain the true parameter value even with the t-test. But what if there’s a real variation between student effect, and we want to account for it? A method to handle additional variance is to use multilevel/hierarchical/mixed effect model. This family of models will allow you to specify an additional level of data that causes a variation in the effect. In this example, that is the student-level variation. This model will also allow us to estimate effects for each level unit separately. Meaning, in this case, we can obtain the effect of the tutoring class for each individual student as well, but we will ignore them for now.

The above figure shows 95% confidence intervals of the tutoring effect when we use a mixed effect model instead of a t-test (or OLS). The confidence intervals become much wider (not overconfident) and contain the true parameter value (0) roughly 95% of the time now.

This is just one of many other scenarios that can throw off your confidence intervals and p-values. In reality, things are even more complicated than the example I used here. And anything can mess with your confidence intervals. I may write about them later when I feel motivated :)

One thing you should be aware of is that statistical inference with mixed effect model can be tricky and requires a specific algorithm to calculate confidence intervals and p-values. In the frequentist paradigm, once you move beyond traditional statistical tests, inferences (or the quantification of uncertainty) become complicated. Another thing is when you use the mixed effect model, the average effect can become a little less relevant. Do you still care about the aggregate effect if you can estimate the more granular level of effects now? The answer is dependent on context and theory that drives your research question.

P.S.1 This article is also inspired by a preprint “generalizibility crisis” by Tal Yarkoni, in which he discusses issues that arise with mismatches between statistical models and verbal hypotheses more extensively (https://doi.org/10.31234/osf.io/jqw35).

P.S.2 Here’s the link to an R notebook that I used to produce analyses and figures for this article: https://htmlpreview.github.io/?https://raw.githubusercontent.com/chasnew/data_simulation_repo/master/unacc_var_mixedlm.html

We should drop the use of the word “passion” and stop telling people to follow their “passion.”

And here’s why.

Passion is vague

First, it has a very vague concept in speaking language. But people use it in a way that it seems like a key to happiness in life. I feel like whenever people refer to the word “passion” they put their own context into it (think of research findings that get misinterpreted because people generalize scientific terms that have different nuances from speaking language) and make it sound like a secret ingredient to succeed in your profession. The consequence is that some people might be trying to find their own “passion” but the best advice people who use these words can give them is a very vague description of what it should feel like or, even worse, the exaggeration of what it actually feels like. Just telling people to get a job that they like or that will satisfy them is clearer and better than telling them to follow their “passion” in my opinion.

Passion has costs

My second reason is that not everyone could afford following or finding their “passion.” My own hypothesis is that people who suggest others follow or find their “passion” is just having a survivorship bias. Just because it works out for you doesn’t mean it will work out for others. Some people who go out of their way to follow their passion without having a safety net to back them up could fail horribly and will never get back up again, especially people with low socioeconomic status. And not every passion has the same financial return. If you’re passionate to be a farmer in Thailand, only God could help you from financial disadvantages that come with it. So make sure you can afford the cost of “passion.”

Abusive passion

Come to my last point, if you’re a business owner and tell people that you’ll only hire people with “passion” for the job. That will just sound like you’re trying to get the highest productivity with the lowest cost possible. You want people who’re willing enough to work more than they get paid, which, to me, sounds exactly like an overwork culture. I’m not telling people to only work as much as they get paid. If you enjoy doing it more, that’s good for you. But you have to realize when it causes you more distress and starts interfering with your personal life. Work is just work after all. No matter how much you like and value it, it still can cause you stress because it’s not a hobby, and it comes with responsibility.

I’ve seen some people argue that you can’t make significant impacts by subscribing to the work-life balance culture. And they like asking the question “how are you supposed to change the world by working 40 hours a week?” My answer is “I don’t.” I’m not too full of myself to think that I *alone* can change the world or make any significant impacts. If it’s a group of people that include me, that sounds more plausible. End of my little rant.

Around 3 weeks ago, I attended the Machine Learning Research School (MLRS) 2019 in Bangkok, Thailand. The event basically includes lectures of the topics relevant to Artificial Intelligence (AI) and Machine Learning (ML). But one particular topic that I was looking for the most was Causal Inference taught by Elias Bareinboim. It’s one aspect of data science that is crucial but not taught widely enough in my opinion. I only became aware of this area only recently as well. So, this blog is my attempt to raise awareness on this topic and, perhaps, help better the practice of data science in general.

What is Causal Inference (CI)?

You’ve probably heard the catchphrase “correlation is not causation” before. But what does it mean to infer causation? The process of inferring causal relationships is called causal inference. This process is needed when we want to know whether a certain event/action causes another and doesn’t just correlate with or predict it. Still, we usually use statistical techniques to determine the causal effect. But statistics alone cannot inform us about causality (or anything really).

Causal Diagrams with the third variable

Let’s take a look at an example used in Elias’s lecture at MLRS. Let’s say we obtain a dataset containing data of people with a serious illness. The data indicates whether patients taking a drug (treatment) and whether they survive after taking the drug (outcome). We go ahead and explore the data and create a table showing statistics of recovery rate as in figure 1.

From figure 1, it seems that patients who take the drug have better chance of surviving. But when we later take into account the variable sex of the patients, it results in the tables with different story to tell like in figure 2.

From what we can notice from the tables in figure 2, it seems that the patients are more likely to survive by not taking the drug for both male and female patients now. This is a phenomenon called Simpson’s paradox. This phenomenon happens when some subgroup variables are neglected in the analysis resulting in a disappearance or reversal of patterns in the data. This phenomenon can be generally attributed to “confounder” variables. So, how do we prevent this from happening in the first place? We have to remember that statistics is just a tool. Analysis results can’t tell us anything without our assumptions or contexts. In a causal inference framework, Structural Causal Models (SCM), we express our assumptions using a causal graph, which is basically a Directed Acyclic Graph (DAG). Considering our example, there are 3 main potential causal graphs that would give us results as in figure 2. These three causal graphs are illustrated below in figure 3. Each of them reflects different assumptions of how the world works, or how taking the drug, sex, and survival rate relate to one another in this scenario.

According to figure 3, our first potential causal graph (graph A) indicates that the sex variable is a confounder variable between the treatment and outcome variables. Meaning the sex variable partially determines who do (or not) take the drug while it also influences the survival rate (Males and females have different base survival rate). Whereas, the second causal graph (graph B) reflects that the sex variable partially mediates the effect of the drug on survival rate. But we can use our intuition to filter out this possibility because the drug shouldn’t be able to change the sex of the patients. Lastly, graph C indicates that the relationships between drug, sex, and survival rate are confounded by unobserved variables. Learning from this example, you might start thinking that we should throw in as many variables into the analysis as possible. But there’s another type of variables in the causal framework that’s called a collider. A collider variable in a causal graph will look like in figure 4 below.

A collider variable is a variable that you must not incorporate into your analysis because it will produce a relationship between your variables of interest when there is none. From figure 4, variable C is a collider variable and it will establish a link between variable A and variable B if you include it in your analysis even though there’s no direct link between A and B. Another article here will give you a clearer example of colliders (Rohrer, 2018).

Causal hierarchy and identification

According to Judea’s causal hierarchy, there are 3 levels of causal questions we can ask, including association, intervention, and counterfactual. This model is described as a hierarchy because if you have access to information at the higher levels, it’s implied that you can answer the queries at the lower levels as well. The lowest level of this hierarchy is association that includes correlational analyses and most predictive models (Pearl, 2019).

When we try to determine a causal effect of 2 or more variables, we have to be able to determine at least the interventional probability of those variables. Meaning, we will need to focus on P(Y|do(X)). But what does the notation do(X) mean? It means that we have control over variable X and determine which subject gets which value of X like in an experiment. Does that mean we can only answer this type of queries with experimental data? No! This is where SCM comes in. In the SCM framework, with your causal graph, you can perform “do-calculus” to convert the interventional query into the form of associational query that adjust for relevant variables as in figure 6. If the conversion is possible, it is said that your query is identifiable. From the example in figure 6, it appears that our query P(y|do(x)) is identifiable by adjusting for the confounder z. This is how you would attempt to infer causality in the SCM framework in general.

But as you may have noticed already, SCM mainly concerns with identification of the effect. It doesn’t tell us what type of relationships our variables of interest have or what estimation method we need to perform to determine the precise effect. But again it doesn’t mean whatever estimation method you end up using is assumption-free. Each of them has different assumptions regarding the type of relationship between the variables too. After all this is what I really want to emphasize. Know your assumptions and be explicit about them! It would add much more value to your data science work than just misleading correlational analyses.

Now, are you wondering how we could come up with a sensible causal graph? In few circumstances, you can come up with them yourself. But practically, you need to consult the existing body of research. In other words, you need to read lots of papers about the topic you’re working on! That’s right. You have to read other people’s work who have done similar things before you and don’t pretend like you’re the first person who comes up with the idea. From there, you’ll get a sense of what causal structure people use these days and modify it according to what you discover from your data. You may be thinking it’s a lot to ask data scientists to know everything in every domain they work on. Even scientists struggle to do so. That is why we need collaboration with domain experts and people who have theoretical backgrounds of the topic to conduct causal studies.

Extra notes

SCM is not the only framework existing for causal inference. To my knowledge, this approach is widely used among epidemiologists but not economists. Many economists rely on a different causal inference framework called Potential Outcomes (PO). These 2 frameworks are mostly interchangeable, but there are some fundamental differences that lead to different practices and techniques used in estimation too (such as Regression Discontinuity Design, Difference in differences, and so on). And you’re more likely to encounter PO approaches when you try to find online courses that teach causal inference. But there are still many more materials about SCM out there

References

1. Pearl, J. (2019). The seven tools of causal inference, with reflections on machine learning. Communications of the ACM, 62(3), 54–60. doi: 10.1145/3241036
2. Rohrer, J. (2018, March 19). That one weird third variable problem nobody ever mentions: Conditioning on a collider. Retrieved from http://www.the100.ci/2017/03/14/that-one-weird-third-variable-problem-nobody-ever-mentions-conditioning-on-a-collider/

Causal inference & Data analysis with assumptions was originally published in NewStories on Medium, where people are continuing the conversation by highlighting and responding to this story.

2018 has been an eventful year for me.
- I graduated from my master’s program.
- I decided to get a job in industry instead of going for a PhD program right after graduating.
- I published a paper at one of the ACM conferences, which results in me going to Boulder, Colorado.
- I revamped my Twitter feed, and now it’s full of researchers and professors Lol
- I had to move twice including moving back to Thailand.

Graduation

For the first one, I don’t really feel like graduation is my big life moment. I know that I might have underappreciated my own graduation both this time and also my Bachelor’s graduation. Most people probably can see from my timeline that there’s not a lot of my graduation photos. But it’s probably because I didn’t think of graduation as my goals of going to college. My goal was to learn and gain as much as possible from my college life and whatever the universities have to offer. And in that regard, I don’t feel like I took the full advantage of it. But it was an amazing experience, and I learned and gained a lot.

Career decision

Towards the end of my master’s program, something happened (I’d like to reserve this information for privacy), and it turned out that the professor with whom I was working will no longer accept anymore PhD students. Don’t feel sorry for me yet. I was surprised at first, but when I actually thought about it, I realized that now I had so many options. Before that, I was fixated on doing a PhD with my professor, but now I started considering other opportunities and possibilities in my life. I did apply to another PhD position in Netherlands. I passed through the first and second rounds of the interview process. However, after the last interview, they decided to go with a different candidate. So, I started looking for a job in industry for real. Don’t get me wrong. I still like doing research and stuff. But I also want a different kind of experience. This is also besides the part where I’m contemplating switching my field of research to quantitative social sciences. And if my passion for research is real, I will be drawn back to a grad school eventually.

Paper publishing

The third one is one of my proudest accomplishment this year. It was a lot of struggles, and It was not even the best paper I could have produced. But the fact that I went through that experience and actually got a paper published at the conference it was supposed to (ACM ICMI 2018) was almost as priceless. And the result of me going to the conference in Boulder was even more invaluable. I had a chance to connect with many interesting and intelligent individuals. Most people who do research probably understand that it’s quite rare to find a circle of people who understand or are remotely interested in the same thing as you work-wise. Thus, whenever you have an opportunity to network with people who share some common interests with you, you don’t let it go. Moreover, the conference had several interesting works, I enjoyed many talks there, and Boulder was gorgeous.

Twitter reform

I started revamping my Twitter feed when I was looking for a job. Originally, I had my Twitter account to vent off some of my personal issues and set it to a private account. However, I decreased the number of my complaint tweets quite significantly when I started the master’s program. And when I was looking for a job, I found out that many people in tech industry use Twitter to announce some job opportunities. And some professors on Twitter also use it to share PhD opportunities and interesting papers. That’s when I officially decided to transform my Twitter feed into something more beneficial to me. And Twitter is now an app that I enjoy scrolling through even more than Facebook and Instagram since it has several interesting and intellectual conversations on it. (Here’s my Twitter account: https://twitter.com/NewRem in case any of you wants to follow Lol)

Migratory life

And here is the major struggle of my 2018 to say the least. because I graduated in May 2018, and the lease of my housing at that time would last until the end of July. Since I had no idea where I would end up getting a job (or would even get one in the US), I needed to find a place that would allow me to stay for only a couple of months. Simultaneously, I was looking for a job, more specifically a data scientist/analyst position. It turns out those positions require very high qualification. Moreover, the job function of data scientists vary from company to company. Some want you to work on the entire data pipeline from data cleaning to modeling. Some want you to build machine models and develop software products. It was almost like I had to know it all and be able to do them all from data structures and algorithms to statistics and machine learning. For a data analyst position, I never heard back from any companies, so I can’t tell what companies actually look for in candidates Lol. Hence, it was a lot of stress I was under, and the anxiety of moving didn’t help. But I found a place in the Westwood area, which I’m really happy about. The neighborhood was great, and the housemates were amazing. I was very sad that eventually I had to move back to Thailand due to issues with the delivery of my employment authorization documents. It took me a while to admit that I had to leave the US. But once I accepted it, I tried to enjoy every single second of my stay in the US. I was still quite emotional, but I managed to enjoy my life more. I went to the ACM ICMI 2018 conference around 2 weeks before leaving the US, and it was one of the best things I did before leaving.

Some of you might also know that I joined an archery club when I started my master’s program. I was not completely involved in the club during the first academic year. But in my second academic year, I joined their competitive team and grew very close to people in the club. They became my favorite group of friends that I liked hanging out with, and I continued shooting with them even after I graduated. I was aware that I had to leave them eventually, but the actual experience of leaving was still very devastating to me.

Toward 2019

Coming back to Thailand has been quite an adjustment for me. I’d like to reserve detailed information for private conversations or another blog post if I decide to do one in the future. But I would broadly use the term “reverse culture shock” to describe my coming back experience. However, one thing I learned from this year is I should be able to make whatever life throws at you into your own story and get familiar with spontaneity.

I decided to look for a job in Thailand, and I think of the coming back to get a job here is another challenge of my life that could help me become a better self. And I got a job offer as a data scientist in Bangkok one month after coming back to Thailand. My job will start on Jan 7th, and it’s the start of a new chapter of my life. I’m really excited about it and can’t wait to see what it’ll be like.

I don’t really do New Year resolutions, but I practice self-reflection every now and then. Sometimes, we go through a lot of struggles and try very hard to make progresses toward our goals, and our effort doesn’t come to fruition often enough. Self-reflection, to me, helps us see all small progresses that we have made, which can be quite rewarding and motivate us to keep going. Another new trick I learned recently is to revaluate your values at every turn in your life. The circumstances we live under keep changing constantly, and that could change our general worldview and life goals. And the revaluation could help us figure out what’s the next step of our lives.

That’s enough about me. I wish you all an amazing 2019 :)

From 2018 toward 2019 was originally published in NewStories on Medium, where people are continuing the conversation by highlighting and responding to this story.

It’s apparent that many people consume information from the internet or social media nowadays. And with every thing that’s going on around the world, there’s an overwhelming amount of information floating around the cyber space. And the social media makes it much easier for people to generate information and disseminate it by themselves. Consequently, the problem of untrustworthy information rises [1, 2]. Information literacy has become a crucial ability because the information you consume shape your perception whether of politics, healthiness, or the world in general [3, 4].

The problem of unreliable information can come in many different forms. Some are easy to notice, and some are more well-hidden. I’ll be discussing them in the following sections.

Arguments with no premises (references)

I believe you have seen those blogs or articles that try to state facts or share knowledge but plainly fail because they don’t really have evidence. It’s sad because it doesn’t mean that they are wrong. They are just untrustworthy or unreliable because their arguments are invalid.

Some of you probably were taught how to craft a legitimate and proper argument. Basically, an argument consists of premises and a conclusion. A conclusion is the proposition that the author of the argument offers, and the premises are reasons to validate the conclusion. An argument will be invalid if it doesn’t have premises or if its premises are false [5]. Does that sound too formal? What I’m saying is, in order to validate the information you present, you need references or premises. Otherwise, whatever you tell other people is just your own beliefs or opinions. And just because you believe it doesn’t make it true. And this is why you were taught to make a citation when you write essays or papers. It makes your papers sound and reliable.

Sufficiency of references

Some people know that they need references to support their arguments. So, they add a list of references at the end of their articles. That should suffice, right? Not if your references are invalid. As mentioned in the previous section, your argument will also be invalid if your premises are invalid as well. For example, if I argue that the original version of the saying “blood is thicker than water,” which simply means family is more important than other people, is actually “the blood of the covenant is thicker than the water in the womb,” which insinuates the opposite idea, and cite two other articles [6, 7], does it make my argument true? Obviously no, if you check out the articles I cited, you’ll see that their articles don’t have any references either, which demolishes the legitimacy of the argument.

To make my argument valid, I need a reference that will help people trace back to the actual evidence, which is the first writing that uses that saying. As you can see it’s not that you can just cite any articles to make your arguments valid. Some trustworthy sources of information that you can cite usually include government websites, educational institution websites, credible media organizations, and peer-reviewed research papers.

So, what if I refer to a peer-reviewed research paper? It seems like a very valid premise, right? Here’s where it gets tricky. There are at least 2 two things about research you need to know. First, a single research paper usually doesn’t contain a lot of information. If you have read a research paper before, you might notice that even a 4-page paper could contain more than 20 references. That’s a general number of references that we need to provide a legitimate and comprehensive background for our readers. When I see some articles that make up an entire theory out of a single research paper, I usually get skeptical. However, it’s worth keeping in mind that some long papers can contain a lot of information (many contributions) as well. A research paper is basically arguments for researchers.

Research arguments

In the world of science or research, scientific communities might not quite a united front. Like in everyday life, we all have different opinions on many topics. Researchers and scientists also adhere to different theories in their research areas as well. Even though some studies can be quite undeniable such as climate change [8], some still can be as controversial as politics.

I’ll use the above paper as an example. This paper was written by Paul Ekman, and a part of it basically argues that basic emotions are discrete and each emotion has its own neural circuit resides in our brains. Ekman is a renowned psychologist who was also a consultant of the movie Inside Out [9].

Although the movie was very well rendered, I’m not a big fan of his theories, especially his work on microexpressions. But my personal opinions aside, there are many other researchers disagree with his work. James Russell, for instance, argues that a dimensional model would better characterize emotions[10]. And because Ekman uses the work on universal facial expressions as his premises for the basic emotion theory, Russell has crafted an argument pointing out flaws in the studies of facial expressions and casting a doubt on the legitimacy of Ekman’s theory [11].

Another thing about citing research papers is that they can be dated. Note that the Ekman’s paper I use as an example was published in 1992, which is rather dated. There have been many further studies conducted to support or oppose the theory. It’s crucial to stay up-to-date with the most recent study. To realize the more recent situation of the field of emotion, read this [12].

For any nerds who might be interested, another example of research controversy is the following series of papers (arguments) on the evolutionary aspect of language:
The Faculty of Language: What Is It, Who Has It, and How Did It Evolve?
The faculty of language: what’s special about it?
The Evolution of the Language Faculty: Clarifications and Implications
The nature of the language faculty and its implications for evolution of language (Reply to Fitch, Hauser, and Chomsky)

The conversation includes a few well-known linguists consisting of Marc D. Hauser, Noam Chomsky, W. Tecumseh Fitch, Steven Pinker, and Ray Jackendoff.

Last thought

I know it’s a lot to take in and will require a lot of you exercising your logical thought. But in this digital era, there’s a lot of false information floating around. It’s important that we can distinguish useful and trustworthy information from the invalid ones. I hope this article will help you realize how to navigate social media in a more astute manner and also hope that you enjoy reading it :)

References

[1] False Information on Web and Social Media: A Survey
[2] Twitter under crisis: can we trust what we RT?
[3] Prevalence-induced concept change in human judgment
[4] How Memories Shape Our Perception of The Present
[5] Argument
[6] 4 quotes that you have been terribly misquoting
[7] 31 Famous Quotations You’ve Been Getting Wrong
[8] Climate change: How do we know?
[9] An Argument for Basic Emotions
[10] The circumplex model of affect: An integrative approach to affective neuroscience, cognitive development, and psychopathology
[11] Is there universal recognition of emotion from facial expression? A review of the cross-cultural studies
[12] What Scientists Who Study Emotion Agree About

Information validity and argumentative thinking was originally published in NewStories on Medium, where people are continuing the conversation by highlighting and responding to this story.

เนื่องจากไม่นานมานี้ได้เห็นข่าวการออกความเห็นของ พล.อ.ประยุทธ์ จันทร์โอชา เกี่ยวกับการอบรมสั่งสอนนักเรียนจาก https://www.facebook.com/thematterco/photos/a.1735876059961122.1073741831.1721313428084052/2005802166301842/?type=3&theater

เราก็มีความรู้สึกว่า โอ้โหความอนุรักษ์นิยมพุ่งเกินระดับ 10 เลย
ณ จุดๆนี้เข้าใจว่าท่านนายกแกคงจะมองโลกกันคนละแบบกับเราอย่างจริงจัง
แต่ๆไหนๆก็ไหนๆ ข่าวนี้มีเนื้อความที่เรารู้สึกว่าน่าสนใจอยู่ คือท่านนายกบอกว่า เด็กก็เหมือน “ผ้าขาวที่พับไว้” แล้วถ้าไปสอนเรื่องความขัดแย้งเนี่ย “ผ้าขาวมันจะไม่บริสุทธิ์” เรารู้สึกว่าเป็น keywords ที่น่าสนใจ เพราะมันดันไปเข้ากับทฤษฎีด้านศีลธรรมอันหนึ่งคือ Moral foundations theory [1]. ทฤษฎีนี้เขาวัดความรู้สึกด้านศีลธรรมของคนเราจาก 5 ด้าน (เพิ่งเพิ่มเป็น 6 ไม่นาน) ประกอบด้วย

1. Care/harm
2. Fairness/cheating
3. Loyalty/betrayal
4. Authority/subversion
5. Purity/degradation
6. (Liberty/oppression)

แล้วในเนื้อความคร่าวๆที่ท่านนายกพูดมามันก็บ่งบอกถึงการให้ความสำคัญด้านความบริสุทธื์ (purity) สูงมาก ซึ่งช่วงแรกๆความบริสุทธิ์จากทฤษฎีนี้ถูกผูกเข้ากับเรื่องศาสนา แต่เขาเริ่มแก้แล้วเปลี่ยนคำนิยามให้เกี่ยวกับแนวคิดของความรู้สึกขยะแขยง (disgust) กับความแปดเปื้อน (contamination) ซึ่ง 2 อย่างนี้เขาก็ให้แนวคิดมาคร่าวๆว่ามันเกี่ยวกับการละเมิดบรรทัดฐานทางสังคม (social norm violation) แล้วเขาก็มีการทำวิจัยโดยใช้ทฤษฎีนี้เพื่อดูว่า คนที่มีความเห็นทางการเมืองต่างกันเนี่ย มีความรู้สึกด้านศีลธรรมต่างกันยังไง [2] ผลคือพวกเสรีนิยม (liberal) จะให้ความสำคัญกับ 2 ด้านแรกมากกว่า (care, fairness) ในขณะที่พวกอนุรักษ์นิยมจะให้ความสำคัญกับ 3 อันหลังมากกว่าพวกเสรีนิยม (loyalty, authority, purity)

ก็นอกจากเรื่อง Moral foundations theory แล้ว เราก็มีความเห็นว่าประเทศเรากำลังถูกนำไปในทางอนุรักษ์นิยมอย่างขีดสุด จริงๆเราก็เห็นข่าววิจารณ์รัฐบาลเยอะนะ ทำไมอะไรๆก็ยังเหมือนเดิม รัฐบาลไม่อ่านข่าวประชาชนเหรอ
จริงๆก็คงอ่านน่ะแหละ แต่แนวคิดทางการเมืองจริงๆเป็นเรื่องที่เปลี่ยนยาก เพราะมันมีความเกี่ยวข้องกับแนวคิดของตัวตนเราเองสูง (self-concept) มันบ่งบอกได้ว่าเราให้ความสำคัญกับเรื่องแบบไหน แล้วเขาก็มีงานวิจัยออกมาเหมือนกันว่า ถ้าเรานำเสนอข้อมูลที่มันขัดแย้งกับความเชื่อแกนหลักของเขาเนี่ย แทนที่จะเปลี่ยนใจเขา มันจะส่งผลให้ความเชื่อที่เขามีอยู่เนี่ยมันแข็งแกร่งกว่าเดิมแทน [3] ซึ่งเหตุการณ์แบบนี้เขาเรียกว่า Backfire effect ในทางจิตวิทยา

เอ่า อย่างงี้เราก็ไม่มีทางเปลี่ยนใจเขาได้เลยเหรอ
มันก็ไม่จริงซะทีเดียว มีงานวิจัยออกมาเหมือนกันว่าความอนุรักษ์นิยมเนี่ยมันเกิดขึ้นเพราะความกลัว มีการทดลองที่เขานำเสนอผลร้ายของนโยบายต่างๆให้พวกเสรีนิยมดูแล้วปรากฎว่าพวกเสรีนิยมเริ่มมีความเห็นตรงกับอนุรักษ์นิยมมากขี้น ในขณะที่อีกงานวิจัย สร้างสถานการณ์สมมุติขึ้นมาให้พวกอนุรักษ์นิยมรู้สึกปลอดภัย แล้วปรากฎว่าพวกอนุรักษ์นิยมมีความเห็นด้วยกับแนวคิดแบบเสรีนิยมมากขึ้น[4]
แล้วสุดท้ายคือจริงๆความเห็นทางการเมืองมันก็ไม่ได้แบ่งคนเราออกอย่างชัดเจนขนาดนั้น ถ้าเราลองคุยกับเขาด้วยเหตุผลด้วยความจำกัดหัวข้อที่ไม่กว้างเกินไปเพื่อลดความรู้สึกว่าความเชื่อแกนหลักเขาถูกคัดค้าน เราก็อาจจะเปลี่ยนความเห็นของอีกฝ่ายได้เหมือนกัน

ส่วนใครที่อยากลองทดสอบความรู้สึกด้านศีลธรรมตัวเองก็ไปที่เว็บ http://www.yourmorals.org/explore.php แล้วเลือกทำ Moral foundations questionaire ได้นะครัช

Source:
[1] Moral foundations theory: http://moralfoundations.org/
[2] Liberals and Conservatives Rely on Different Sets of Moral Foundations: http://www-bcf.usc.edu/…/GrahamHaidtNosek.2009.Moral%20foun…
[3] Neural correlates of maintaining one’s political beliefs in the face of counterevidence: https://www.nature.com/articles/srep39589
[4] A Yale psychologist’s simple thought experiment temporarily turned conservatives into liberals: http://www.businessinsider.com/how-to-turn-conservatives-li…

ความเห็นทางการเมือง, ทฤษฎีศีลธรรม, และวิธีการเปลี่ยนความเห็นทางการเมือง (Liberals and… was originally published in NewStories on Medium, where people are continuing the conversation by highlighting and responding to this story.

As machines have become more and more intelligent, people heavily focus on the field of artificial intelligence (A.I.), including machine learning and deep learning. My purpose here is to give a general idea of what this area of study encompasses and relate it to machine learning.

Some computer science students who are interested in research might say their research interest is A.I. But A.I. is an extremely broad area of study, practically universal. The field is built based on the idea of making machines intelligent, which brings up the another question that is “what is intelligence?” The concepts of artificial intelligence have been roughly categorized into 4 schools of thought as on table 1 (Russell & Norvig, 2010).

The row on the top shows approaches that try to model thought processes. And the row on the bottom shows the approaches that try to address issues concerned with behavior and actions. Each column corresponds to a different performance standard for AIs, where the ideal performance captures the concept called rationality. Some fields that adopt the notion of AI agents thinking humanly are neuroscience and cognitive science while agents that think rationally would exploit the field of logic. For the agents that act humanly, you might relate them to the Turing test, a test of machines’ ability to trick humans into believing that they’re humans. There have been several revised versions of this test due to its flaws.

Regardless of approaches, in general, a computational problem is widely recognized as an A.I. related problem when it’s NP-complete (even though that’s not necessarily the case). Informally, NP-complete defines problems in computer science that cannot be solved in deterministic polynomial time. In other words, with existing algorithms, we don’t know how long the program will take to find a solution if there’s any. By how long, we don’t just mean minutes or hours, but we means years, decades, centuries, or more. Thus, we need some tools to deal with this problem.

A.I. as a search problem

As CS students progress in their major, especially if they take an algorithm course, they will learn how to look at a problem from different perspectives resulting in different representations or formulations of the problem. Think of origami. You have tools, usually just paper in this case, and you can transform it into many distinct shapes. The same idea can be applied to problem-solving in the field of computer science. Hence, in this article, I will try to formulate A.I. problems as a search problem. What do I mean by searching anyway? In this context, it means searching for a solution(s) (automatically). Let’s use the figure below as an example.

From the figure, we’re trying to get from Lyon center of University of Southern California (USC) to one of the Panda Express branches. As the map shows us, there are many ways to get there. We can take a shuttle, drive, walk or bike. It’s not only the mode of transportation that we can choose, but also the path that we want to take. We can even walk around the campus twice before walking to Panda Express. But Google map searches through the solution space and selects only these perceivably efficient ways to get there for us, which are our solutions here.

Formalism of search problems for A.I.

How does it search exactly? Didn’t you just say A.I. related problems are NP-complete? In computer science, classical search algorithms, such as breadth first search (BFS) and depth first search (DFS), are very common. But they can’t deal with our NP-complete problems, which most real-world problems are NP-complete. Thus, there is another family of algorithms called local search that trade systematicity for computational efficiency. First off, let me introduce you to some components that are essential to search problems, which are states and actions. States represent the current condition of the agent(s) and the environment, and actions are how the agent responds to its environment usually resulting in the transition of the state.

Let’s revisit our Google map example. Now we decide to walk from Lyon center to Panda Express. Google map recommends three different paths for us. So, I visualize them using a graphical model. According to the figures above, the states represent where the agent is geographically. And the actions will correspond to which path the agent take. If our problem is to find a shortest path, BFS will examine every possible path from our origin to our destination and return the shortest one, which is why it’s called exhaustive search. Classical search algorithms search systematically and keep track of every action and path unlike local search.

Turning to local search

As you can see from our classical search example, its solution is actually a path. What if we don’t care about the path? Local search algorithms return their state(s) as a solution, and they store only small number of states in memory. As a result, it requires significantly lower capacity and much lower time to execute. Local search also deals with problems better when we take into account the nature of real-world problems, such as stochasticity and partial observability. One basic algorithm in this family is hill-climbing. Hill-climbing is an iterative algorithm that will always select an action that leads to the adjacent state that maximizes our objective function, a function that quantifies the objectives that we want to achieve. It will stop when there’s no adjacent states that will increase the value of the objective function. Let’s say that there’s a classroom that has only one row of seats. And somehow you come up with a mathematical function that takes the number of a seat and output your preference score of the seat as in the figure 5. And you want your program to find your favorite spot.

Here, the preference score is our objective value. Now when we run the hill-climbing algorithm, it’ll randomly initialize the starting point(seat’s number) and begin to find an optimal solution. Considering this set of preference scores, no matter where the program starts, it’ll always find the optimal solution, which is seat#6 as in figure 6. Now we try changing the function that calculates your preference scores, and it outputs a new assignment to the seats as in figure 7.

Figure 8 shows that if our initial seat is seat#0, seat#1, and seat#2, the algorithm will get stuck on a local optimum, which is seat#2. Local optima are very common for local search. The number of local optima depends on the complexity of our objective function, which is usually high. Problems’ complexity also expands as the dimension of our functions grow.

Figure 9 is an example of a potentially real search space that has two dimensions, meaning the objective function takes two arguments to calculate the value. Objective functions of real-world problems typically have a myriad of local optima. Hence, there are other algorithms of local search that try to address this problem including simulated annealing, beam search, genetic algorithms, and etc.

How is machine learning a part of A.I.?

While A.I. is just the way to formulate problems, machine learning is actually an approach to solve them. The idea of machine learning is to use available data to train a program that can learn to make a prediction or a decision. The core component of the program that allows it to learn and make a prediction is called model. And the learning process of it is called training. When we train a model, we need an objective function to guide our model. One well-known numerical method for supervised learning problems is gradient descent. Gradient descent shares close affinities with hill-climbing search. Instead of trying to find a maximum point, gradient descent tries to find a minimum point, which is only a flip of the sign. It uses a loss function as the objective function. Loss functions indicate the distance between the function learned by the model and the actual function that describes data’s characteristic. At each iteration, it will compute the gradient(s) of the objective function, which is comparable to finding an action that will lead to a better state in hill-climbing. Then it subtracts the computed gradient from the model’s parameters to update the model’s state.

Figure 10 illustrates the update step of gradient descent, where s represents model’s state, F(s) is the objective function, and 𝜸 represents the learning rate of the algorithm that is a constant that indicates the size of the update. Gradient descent also suffers from local optimum problems. Therefore, people have come up with many variants of the algorithm. (More about gradient descent and its variants here)

Last thought

As I mentioned at the beginning of the article, there are several stances that we can take to approach A.I. related problems. However, I’m personally disinclined to stick to any particular school of thought. Different approaches will reveal different facets of a problem. Furthermore, solving an A.I. related problem hardly requires a single approach. For example, artificial neural networks or deep learning is a machine learning technique inspired by the human’s thought processes. Still, we could build an agent that thinks rationally using this technique. A.I. is a field that has a lot of potential, left alone a prevalent field like machine learning itself. I write this blog in a hope to deepen people’s comprehensions of the field and as the initial attempt of a blog series about my works during my graduate school.

Caveat: There are alterations to the definition of some terminologies to keep the content accessible to general readers. Nonetheless, you should get an idea of what A.I. field includes and how to approach its problems as a search problem and generalize your understanding.

Anyway, any comments on the content will be appreciated. And thanks for reading :)

References

• Russell, S., & Norvig, P. (2010). Artificial intelligence : A modern approach (3rd ed., Prentice Hall series in artificial intelligence). Upper Saddle River, N.J.: Prentice Hall.

Artificial Intelligence and search problems was originally published in NewStories on Medium, where people are continuing the conversation by highlighting and responding to this story.

Have you ever asked your Thai friends for their names, and the answer you got was some names that sound strange and also are not their real first names? It could even be an English word. For example, in my case, I like to introduce myself as “New” and tell people that it’s my nickname. And here are some possibilities that I imagine going on in the heads of the people I talk to.

1. “That is probably his English name that he made up when he came to the US.” I understand that many international students do that. But that’s not the case here.
2. “That is probably the name that his friends named him because of something he’s done.” It’s true that we call it a nickname. But you’re still far from correct.
3. “This guy is trying to hide his identity!” Ughhhh, you’re the worst here.

So, what is it?

The truth is every Thai person has a nickname! And it’s given by parents at birth! (Though, it could be changed later.) And we can be named literally anything, such as colors, English words, animals, fruits, and etc. And 90% of the times, we use our nicknames. One occasion that we’re called by real names is when teachers teach many students and don’t have time (or don’t want to spend time) to ask students their nicknames. Obviously, using the names that are already printed on the paper is easier. Some other occasions that we use real names are when we have to do paperworks because, apparently, real names are official names. Besides those scenarios, when someone asks for our names, the answer will usually be the nickname.

Why do you need a nickname then?

I’ve read an article that explains the reason behind Thai nicknames. It says that people in the past believed there were evil spirits looking for newborn children to take them away. And giving a nickname to a child somehow could confuse the spirits. The belief may no longer have an effect, but the practice still persists.

It is one of the commonest Thai culture and will probably last as long as Thai culture exists. And I’m somewhat surprised that it’s barely known by people in other countries. This doesn’t count the part that I feel slightly disappointed when people mishear my name, and it becomes “Neo.” Or if I don’t spell it out, many people would think that it’s “Nu.” I mean the latter one is more acceptable. At least, they got the right pronunciation. Still, I don’t want to have a new name just because I come to the US. Otherwise, does it mean that I need a new name every time I go to a foreign country? I think a name is an important part of a person. It could induce a cocktail party effect. And in our case, it creates a sense of friendliness in our interactions. Our names are integrated into our personhood in a very profound way. So, next time you see Thai people, try asking for their nicknames. They will be more than happy to tell you :)

References

• https://www.into-asia.com/thai_language/reference/nicknames.php
• https://www.sciencedaily.com/releases/2012/04/120418135045.htm

Facts about Thai nicknames was originally published in NewStories on Medium, where people are continuing the conversation by highlighting and responding to this story.

The first two things I had missed the most since I came to the US were my people, friends and family, and the food!! Hence, what my friends and I did when we met was having lunch together Lol.

Obviously, we have dim sum too. As including all pictures of the dishes we ordered might be cruel to some people, I posted only some of them haha. We then went to a dessert cafe.

This is a dessert called “(Pat)Bingsu” that is Korean-inspired to my understanding. It’s basically shaved ice topping with a lot of stuff, as you can see, depending on our choice. It’s pretty popular in Thailand nowadays.

It was almost New Year; thus, People started decorating the building and the shopping mall for the celebration. We went to the Bangkok Illumination event. The lighting was gorgeous.

After the trip with my friends in Bangkok, I went back to my hometown, Khon Kaen. It might not be a tourist city, but there’s so much memory there. Another thing that made me happy to go back home was my lovely and cuddly cats!

The left one is Jerry, and the right one is Poo (it means crab in Thai). They’re adorable right? I have to see them when I make a video call with my family haha. Anyway, I also met up with my college friends that were still in Khon Kaen and went to some restaurants that were reminiscent of our college life.

Despite all the fun and nostalgic moments, I went back with a different attitude towards everything. I’d immersed myself in a very different environment for quite long. Obviously, my perception of humidity had become more sensitive. When I stepped outside the airport I felt that the weather was uncomfortably humid, which I had never felt before I came to the US. It was probably because I got accustomed to the dry weather of LA. Also, when I accidentally walked into someone, the first word I said was “sorry.” Yep, it’s literally “sorry” in English not “ขอโทษ” (pronounced: koh-tod, meaning: sorry). I had to take taxis, which are notorious for their service, instead of Uber. It was my relief that I got a chance to meet with my foreign friends because they were people who I can share my experience of living oversea with. I figure it’s because I’ve lived in the US alone and only relied on myself for so long. I started becoming more independent, which contrasts with the way of living of Thai people, and started questioning if this was my place anymore. Now you might have heard about reverse culture shock. But I doubt that this is it. It was very fun visiting my country, but it’s just also something that I have to ponder whether or not it’s where I belong.

Visiting home (Thailand) was originally published in NewStories on Medium, where people are continuing the conversation by highlighting and responding to this story.