(Disclaimer: the idea and content collection of this article were begun months ago. But the process of publishing this article was rushed by the recent announcement of introducing DeepMind AI to YouTube. So some of the information or critique may not be relevant when you read this article.)
Introduction
I have met this issue of a contained media feed a few times, considering that I am using YouTube since 2012. In these years using YouTube, I have found this "barrier" of YouTube not recommending videos from channels I don't follow. Sometimes it lasts for a week, sometimes up to a month. But I am seeing the most recent "barrier" for 3 months. And I am still struggling to leave this circle.
How does YouTube decide what videos to recommend?
Summarising the official documentation of YouTube, the algorithm works by determining the category of videos (including the title and description), if the video is watched by similar users and the objects in the videos.
Basically, it works like a treemap. The algorithm finds most viewers of channel A also watch channel B. So the next time you watch a video from channel A, channel B will be on the recommended list. But the recommendations will get more personalised as you watch more videos. (And Google has a list of subjects and topics that your account is linked to. You can check them with this guide from Google.)
How to fall into the YouTube feed circle?
(Disclaimer 2: this part is totally personal. I only have one YouTube account and have been using the same one for all these years. And it is really difficult to replicate as the amount of data that YouTube has for my current and any new account are vastly different. But if you are experiencing the same issue, let's discuss.)
I have experienced this YouTube feed circle a few times. Although I cannot recall the exact year, the issue usually resolved within a month. A YouTube feed circle (as I named it) is simply an invisible barrier to reach content from other YouTube channels in the recommended video list.
From my guess, this YouTube feed circle is created because I mostly just consume videos on the recommended list but not searching for other videos. As views on some channels were exceptionally high over the years, this forms a niche of channels and topics that YouTube thinks I like. Subsequently, the platform only recommends videos from these channels.
To note, I have many fields of interest, including tech, cars, news, music, and lifestyle. My girlfriend sometimes may add cosmetics and fashion in the mix (as the house iPad is using my YouTube account.) And we also watch videos in English, Cantonese (and sometimes Mandarin). So I consider my watch history as "broad." But you may disagree.
The introduction of DeepMind
If I am just saying the name "DeepMind," you may not know what it is. But did you remember the highly intelligent AlphaGo which won every human in Go? DeepMind is the Google subsidiary behind it. And recently, Alphabet has announced that DeepMind is involved in improving the YouTube algorithm to provide even more individualised content to every user.
But why introducing DeepMind is so important? Although I believe this is part of a corporation act to save DeepMind from its 1.6 billion pound loss and demonstrate how the findings from developing the best Go player in the world can benefit us, the advanced AI from DeepMind is (kind of) promising.
Before going any further, we should know that most AI that we praise are statistical AI, which meant the AI mostly runs on churning numbers. While it is true that AI is not a living being, graphical AI is more than churning numbers but understanding objects visually.
Taking watermelons as an example, if we see black and green strips, it will be a watermelon. But AI "sees" watermelon differently, if we train the AI with watermelons that have 4 green stripes and 3 black stripes, it will not recognise a watermelon if it has 3 green stripes and 4 black stripes. This is an overly simple example. AI nowadays are able to understand this simple logic, but still far from our normal understanding.
ReCaptcha is a common way that Google uses to train its graphical AI. Notice sometimes that reCaptcha ask about identifying a crosswalk, it will provide three crosswalks. In most times, the Google engine only recognises two of them as crosswalks. The other one is for the human (you) to tell Google if it is right or not.
Saying so, crosswalk is designed to be a simple, distinctive feature that drivers can see from a far distance and slow down for pedestrians. The need to validate the outcomes from the graphical AI engine with reCaptcha demonstrates the ability of graphical AI currently. And not to mention it is just a still image in reCaptcha, recognising an object in a video could be an entire world for AI to discover.
Is DeepMind going to change how YouTube recommends videos?
In short, yes. But it could be a long time away. The automotive industry has been developing object identification for years to understand what looks like a cyclist, a car, or a plastic bag. Considering how many more different objects that the DeepMind-YouTube AI needs to learn, it could be years away. Recent findings of neural networks needing "naps" have also extended the time required for neural AI to learn.
And the development of this graphical AI is vastly different than the statistical AI that YouTube is using to recommend videos to you and me. Linking the findings from the graphical AI with the statistical AI can take time as well.
But the future of introducing DeepMind can open a different universe for YouTube, especially with the competition from Tiktok. By randomly showing videos of different topics, rumors are that Tiktok can understand topics that you're interested in less than 15 minutes. Graphical AI may give YouTube the advantage over Tiktok again. (But we also have to understand that content on YouTube is slowly moving into a more professional manner with proper production.)
Comments