Hey Folks,
The week was a bit hectic with work and intermittent guitar practice sessions. I guess these have been helping to curb my anxieties during the daytime.
I had almost no personal time to read essays from my to-read list. I’ve been going to bed as early as 8 pm and waking up around 4 am; feeling too stressed out, dozing off mostly in my chair.
I am probably bored with almost everything these days. Mindlessly browsing the internet. Putting my phone on airplane most of the time, and not touching it. No mood in talking with anyone. You know how it goes. I let it be.
First, a little bit of introspection
[I had many second thoughts before posting this section. But there you go. There wasn’t any other way to get everything off my chest. I might cringe in the future though.]
A lot of family issues here and there. I don't intend to get too personal in my letters. So, this is the last time I am (over)sharing everything off my head to cope with my anxieties right now. Know that, for the past few months, my parents had some big fights. As a responsible adult child, I have tried to fix that. They had verbal fights which I didn't interrupt [0] except asking if they really want to stay together. A few years back too it was so much mess that, due to certain circumstances, they had to divorce each other. So, yeah! That's that.
The past 15 years have been really tough for us. When we came to Kathmandu for the first time almost 15 years back, we were nearly homeless. I still remember times when our rented rooms were mostly empty, and a few initial winters were a bit rough.
I'd say, right now as I write this, I kinda feel secure to have a roof of my own. But then, it's another abstraction.
A home is an abstraction of struggles, miseries, dreams, wishes, love, pain, suffering, happiness, and all.
Being financially independent feels good. I can contribute to the housing loans we have. There is still a lot of loans left and I am sure I'd be able to help reduce that in the coming five years. [1]
For more than a decade we had so little money to even survive, that our survival was drowned in fear and a lot of improvisations; we had to think many times to even spend money on simple things. As a growing child fascinated by a lot of things, I didn't have money to buy books I wanted to read, technology I wanted to experiment with, guitar gadgets I wanted to try. So, definitely I was very excited when I bought my first electric guitar a year back. [2]
What saddens me really is when people around me become judgemental of me regarding this. They think that a guy living in his parent's house should be having a fulfilled life. They think that I don't have to pay rents, that I don't have to think about breakfast and dinner. But that's an abstraction of miseries. I am telling you this because there were a few "friends" who had told me exactly the same thing. “What does a guy living in his own house in Kathmandu have to fear?” But I digress strongly. Everything is an abstraction. And mostly they are leaky if given enough care.
Right now, half of my earning goes into the family. And remaining half is an improvisation with personal savings and expenses.
That's that. Living in the past sometimes provides me an introspection about my life itself. Now add a lot of personal shit (about love, life, and career/goals), the whole narrative is absurdly miserable. As someone said to me twice, "Nish! You like suffering", I like to think that these miseries provide a way to see life from different kinds of lenses. So, definitely, the Nish you are getting used to in all these newsletters is an abstraction of a weird (and paradoxical) existence who is trying to live for the sake of living, trying to connect the dots. Nevertheless, I am here alive. And that is all matters for now.
It feels lonely to not have anyone to talk to about all these things. So, writing has been a way of meditation. :)
For now, I like to pretend I am alright. And that I have to focus on some other things than being miserably lonely every other day. In all, my work-life has been a way to break out of this “lifey” rut and flow into another Sisyphean one which is less prone to miseries.
Anyway enough chit-chat! I don't have any good reads to share this time. So, I am going to talk about a few events on social media that made me think hard.
Weird sides of algorithms
This might be the most obtuse thing I have ever written till now. But bare with me. Probably, I am overthinking this topic too.
You might be wondering why the title isn't “Biases in algorithms” (Wikipedia link). Probably, that is a much better title. However, I "kinda" believe that biases are societal constructs rooted in the interpretation of events you encounter. So, definitely biases arise from underlying “prior” assumptions from human beings, be it algorithms or data modeling. [4]
[Figure: Bubble with data points]
Let's take a realistic example.
Suppose you do a small survey of company X. One statistic points out that it is mostly comprised of female workers. How do you interpret the statistic? What happens if you share the result with more people? Will they have the same interpretation as yours?
I am sure the interpretation isn't the same. Some might say the company is doing great in [[Women Empowerment]]. Some might interpret it as being biased against males. Whatever it is, it's not possible to find common grounds.
Maybe, people are being blatantly aggressive to the statistics? Maybe the survey didn’t have any malicious intent? It was simply done to get insights on age groups to optimize working hours. The “gender” narrative is simply the wrong measure to look at, and it probably arises from people's non-neutral thought processes? Or perhaps that's also the consequence of people having a specific [[Belief Template]] they can't get out from?
Wrong interpretations have unforeseen consequences, especially in times where information (and misinformation) spreads rapidly than ever. (See: Issue-11).
The situation is uncannily similar for algorithms that power the tech world. One wrong interpretation with heated discussions and it creates a catastrophe. (I have talked about face recognition in issue 07 regarding AI Ethics. However, ethics and bias are two different things. This time the context a bit different.)
Twitter’s Smart Cropping
Recently there was an experiment related to Twitter’s cropping algorithm.
The algorithm inevitably cropped out people with darkish skin tone. People blatantly chimed in that Twitter was biased and that their face recognition was racist. However, that's not even the case. Twitter doesn't use, verbatim, any face recognition in its smart-cropping algorithm. Rather it uses a saliency-based neural network to get the region of interest.; that is it uses the part of the image a person is most likely to look at. And it just happens that the cropping out people of darker skin tone is a side effect. There's no inherent racism in the algorithm itself. Noone coded it. Using saliency kinda makes sense for non-human pictures where, instead of showing the whole picture, you show only a relevant region that is highly saturated and has some good contrast.
This doesn't mean there's no bias in the algorithms. Surely, the dataset on which the neural network was trained on might have some kind of biases, and perhaps Twitter should have done rigorous analysis on the model (although they had done it and didn’t find any statistically significant bias). Hopefully, they figure out a way to remove the bias to some extend. Probably, they need a better metric to evaluate? (That might also be…well…complex…because of Goodhart’s law.) That's why there's a whole another field in fairness testing and explainability of such algorithms.
(Related: Responsible AI Practices)
However, that's not what made me a bit salty. It's the why part. I wonder what the original intention was. Why only use the black/white people narrative? [3] I think there could have been other variants of the experiment. Like putting in people with similar skin-tone and changing the contrast/color saturation. Or probably non-human scenes and such. Nobody actually conducted such an experiment (as I write this) to generalize the actual cause. People (probably) flagged the algorithm as “being racist”?
These algorithms aren't "right/wrong/racist" in themselves. Nor the people who write these algorithms. They are very smart people, and probably they know what they are doing (I hope). However, it's the underlying assumptions that backfires while writing such algorithms.
Of course, “proper” fairness checks should be done. Datasets on which machine learning models are trained on should be checked for inherent biases. The model is as good as the data itself. So, it's almost inevitable that sometimes the model backfires in certain sensitive cases.
Another example might be Zoom's face recognition system. It seems Zoom's virtual background doesn't do good for darker faces. In this case, it's riskier. Just imagine having your face cropped out of an important video conference because of a weird algorithm.
Let's take another example.
Suppose you are writing an algorithm to detect “Hate Speech” in social media.
You decide to use a list of words and phrases for the task. You assume that such words will make good candidates. If there are certain words/phrases in the comment, it’s tagged as “Hate Comment”. You test the accuracy and find it to be good. So you deploy it.
What could go wrong?
It might tag more non-hateful comments as hateful because words/phrases are not a good representation of the problem. Sentiments and context should be taken into account.
Accuracy might not be a good metric to evaluate. Since the domain is highly imbalance (good comments overshadow hateful ones in general), a better mechanism should be used to measure the important minorities (hate comments).
Direct release is probably riskier without prior real-world testing. Pr-release testing should be done. Just allow a group of people to use the new feature. It will give good insights into where the algorithm might fail. This might include doing statistical analysis as well as interpretability tests. Or evaluating different types of algorithms.
Even if you are going to use the list of words for the sake of performance, you should cross-reference whether those words overlap with good comments or not. That way, you will get an idea of how much bias is in your assumptions. It might help you to provide a score for the prediction.
Maybe the problem is very complex to solve even by humans?
As someone who is both the producer and the consumer of similar algorithms, I can surely empathize with a lot of these things. My assumption and [[Belief Template]] might be different than yours but almost no programmer will have malicious intent while solving impactful problems. [5] That's why it feels weird to have heated arguments and contentious discussions on these topics without proper prior knowledge.
[[Recommender System]]
Another weird algorithm is the recommender system (pdf link). You open a video of a cute penguin chick on YouTube. You like it. You find another video of dolphins playing catch with pufferfish. Another video of a humpback whale. A few hours later, you are suddenly watching a penguin having an existential crisis. Damn. That's a long rabbit hole.
This happens with almost every platform that is algorithmically putting out content to the users. Amazon’s product recommendations, Facebook's newsfeed, Tiktok's videos, recommended items from YouTube and Netflix. You name it. These systems are optimized to make users engaged; users are mostly guided by recommendation algorithms (recsys). No matter how “good” (or worse) the recsys feel, the ultimate goal is to not only provide quality service but also to get each and everyone hooked in.
For instance, Facebook's newsfeed is algorithmic by default instead of chronological because it focuses on the relevancy of the posts and contents. As Mark Zuckerberg says — "At the end of the day, if someone close to you has a baby or a certain milestone, you'd want to see that post immediately instead of seeing chronologically where you'd it encounter while scrolling after 30 minutes" — the newsfeed kinda makes sense most of the time. However, since it's optimized for engagement it might also backfire, especially with a lot of (inherent?) political chaos (I am not going into details here.)
Similarly, YouTube wants you to stay in the platform longer because it's probably one of the major platforms where ads can reach a large user base.
Tiktok video recommendation is probably the most addictive one. It uses contents plus user preferences to tune the next video. (No wonder my sister and mom are constantly hooked in.)
Netflix is doing fine to get the user hooked in and binge-watch as many series/movies as possible. (I haven’t used Netflix so far. Haha. )
[[Online Interaction]]
“Today” is definitely the reign of big recsys. If you think these platforms are for free, you have to reconsider what free means. YouTube, Facebook, TikTok, and the likes thrive using/selling your data for targeted advertisements. Your behavior is the actual product they are streamlining. While most of the users can see positive impacts, a more subtle, yet unhealthy, impacts are psychological.
Let’s take instagram. You spend some time on Instagram, see a lot of good posts about people living their life. This might make you feel shit about your own life. This is the very reason, I am not that much into social media. It makes me feel more anxious than I should be, comparing my life with others.
Also, let’s not even get into the domain of addiction and the attention economy.
The scariest part might be misinformation spreading among elderly people. They can rarely differentiate whether the news is authentic or not since they are using only a single platform (say, Facebook) and don’t really know how to manually verify the information.
Probably, that's the reason why the movie The Social Dilemma seems good enough to simplify this knowledge for the general audience? [6]
You might find this Twitter thread interesting. It summarizes the movie nicely:
Another topic I like to introspect is about Siraj Raval incident. Being big on the internet has its own perks. A few years back, a very naive Nish used to dislike Siraj for insensitively copying things without credit. Now (for more than a year), as I think about the event surrounding him, I can empathize with him. Probably, he had started to “learn in public” by making videos on Data Science; simplify various topics for beginners. In doing so, he might have grown overly ambitious to get big, and hence improvised on a lot of things. Eventually, everything backfired and now he might be struggling to regain trust.
This is just another thought in my head. I went through some of his recent videos, and I can see a lot of hate in the comment section; people throwing out disparaging sentiments. That's why I have so much skeptcism towards online interaction.
In all these, one thing makes sense:
We are strongly trapped in a filter bubble. Letting an algorithm fully automate a sensitive task is probably not a good idea. Human-in-the-loop is an important component. Algorithms are just tools that we shape which in turn shape our lives.
Leaving on a more positive note: I guess it all depends on how you use these platforms. Too much of anything is probably not healthy.
Recommended Read: The Hardware Lottery
“This essay introduces the term hardware lottery to describe when a research idea wins because it is suited to the available software and hardware and not because the idea is superior to alternative research directions.”
#Watching
Rodney Mullen: Pop an ollie and innovate!
Rodney Mullen | TED Talk | 18 min
At the start of the week, I watched this skateboard video which triggered an endless rabbit hole.
Rodney Mullen is a genius in many ways and I feel I should have known about him earlier. He is, as people say, the godfather of freestyle skateboarding, with the majority of tricks attributed to him.
Here he talks about being creative and how each individual stunt is made of fine-grained sub-movements. (Damn, his skills are beyond comprehension.)
I loved his enthusiasm and energy throughout the video. He even talks about Richard Feynman, Open Source, and hacking. Cool. :D
I also liked this video by Physics Girl where Rodney explains some of the physics behind his freestyle.
Additionally, I found this particular trick very artistic.
Finally, this spiral loop trick by Tony Hawk is insane. Damn!
Being Ugly - My experience
Never Give Up | 24 min
I binge-watched every video from this channel and I'd say I am highly inspired.
I can totally relate to this. A very naive Nish would have felt similar emotion, especially during childhood where he had a weird outlook and was highly insecure about himself - the long nose, pants touching belly button, short hair, and such. Now, I guess he has grown so much, especially with his long hair and beard. Haha. He has gained more confidence. I guess…
Also, you might want to do a follow up: Being Ugly 2020 update (what have changed)
Wander - Natascha McElhone at Royal Botanic Gardens, Kew
Natascha McElhone reads a passage about 'Trees' by Hermann Hesse, at Royal Botanic Gardens, Kew
This is so beautiful. Profoundly moving.
#Poetry
Mary Oliver - Night and the River
…
the fish had vanished, the bear
lumped away
to the green shoreand into the trees. And then there was only
this story.
It followed me home
…
Hauntingly beautiful.
#Music
I have been listening to a lot of Megadeth during the daytime, especially their Rust In Peace album. I have shared some of Marty Friedman’s music before. Tornado of Souls is probably my most favorite Megadeth music with Marty’s insane guitar solo. Songwise, I love Hangar 18 and Psychotron.
During the evening hours, Bartika is filling up the void with Najeek.
Add the Monkey Temple band in the mix and the week was kind of “sad-happy”. :)
Timeless.
Deepak Moktan's guitar solo evokes “weird-good” emotions. :)
#Ending-Thoughts
That's all for now. I hope this write-up wasn't boring.
Love,
Nish
PS: I recorded, yet another, music (although recording is very raw): Poison Rain
#Footnotes
[0] - Let's face it, we are all adults who should know what they are doing! PERIOD
[1] - Five years is a long time. But, I have this less-hopeful wish that I have to do something with my life before 30.
[2] - I wanted an electric guitar from my early engineering days when I was playing guitar for almost 3 years at that time. Now that I have it, I still feel hesitant to spend money on gadgets because of fear of squandering my savings.
[3] - If the world was idealistic and non-discriminating, the cropping wouldn't even matter. Having said that, I am not implying it wouldn't actually matter. Rather, it depends on what the platform optimizes for. In the case of Twitter, the higher the engagement in a Tweet, the higher the retention rate. Probably, that's why they are using **Smart Cropping** instead of a manual one. For the same reason, they might not have gone for scaling down the image and showing it as a whole.
[4] - I don’t mean it’s not serious. Of course, biases in everything should be taken seriously. What I am trying to convey is, everything comes from us humans and hence the bias is the side effect of our own thought processes and beliefs.
[5] - It doesn’t mean there are no malicious actors. Of course, there are. The world is not a fantasy. In such cases, human biases (ignorance + negligence) actually leaks into algorithms. Other than that, experiential biases and statistical biases are two different things.
[6] - I haven’t watched the movie…yet. But I am sure there’s nothing new it can provide to me. Perhaps, I should stop being too much egocentric and just consume things for the sake of entertainment? Sometimes, there shouldn’t be any purpose, right? :)
As always, great post. Thank you nish. keep it up.