Interesting piece! Thanks for sharing.
Nitpick if you are the OP: while I appreciate the distinctive style, the font really made it hard for me to read
I immediately reach for the "read clearly" button in firefox but it doesn't always appear and this time it didn't.
Funnily enough I clicked the chevron and there are some settings. It's a tiddly wiki. I tried changing the font but didn't succeed at first try so I gave up.
As a fellow belgian, I have been following you and segments closely. Congrats on being the first belgian YC company :).
From the beginning onwards I was wondering why you chose to put such an emphasis on segmentation labelling. Do you see this usecase as the Computer Vision application with the biggest (future) market or maybe the least saturated offering at the moment?
Thanks! The existing tools on the market for image segmentation are not very sophisticated, so it's a niche where we can immediately make a difference.
In a sense, image segmentation labels are strictly more informative than bounding box labels: you can trivially extract the containing bounding box from a segmentation mask. One big reason that segmentation labels are not used more often, is simply because they are too expensive. Labeling a bounding box requires only two clicks, while labeling a segmentation mask requires much more time with manual tools. We're trying to solve that problem.
In the future we want to dig even deeper into this problem, and expand our scope to video and 3D segmentation labeling. We believe there will be a huge need for such tools now that everyone is getting smartphones with Lidar and AR/VR capabilities in their pockets.
I genuinely wonder what this means for the future of DeepMind. Is it supposed to be profitable or is it just R&D spending for Google? Where are the paths to profitability for them? Are they counting on more expensive GPT-4,5,6 APIs?
My impression is that places like DeepMind are more of a social good than a profit seeker. Most universities simply cannot afford such big and expensive AI. So I suspect that DeepMind is Google’s way of expanding the frontier of AI knowledge through charitable donations. It’s ultimately good PR for them and maybe influences businesses to pick Google Cloud Services for brand appeal. Also I’m guessing since it’s research, they get a huge tax write-off.
Perhaps they will deliver once-in-decade unicorn breakthroughs that can ultimately be monetized? Maybe Google’s next phase will be AI generated content. With that they can completely cut out other content creators, and thereby collect affiliate cash through sales. For example, generating those “Top 10 Holiday Toys” guides automatically.
I do really like the idea of an MNIST alternative to very quickly verify ideas. However I have a few nitpicks:
1. 10 classes is way too small to make meaningful estimates as to how well a model will do on a "proper" image datatset such as COCO or ImageNet. Even a vastly more complicated dataset like CIFAR-10 does not hold up.
2. I feel like CIFAR-100 is widely used as the dataset that you envision MNIST-1D to be. Personally I found that some training methods will work very well on CIFAR-100 but not that well on ImageNet so TinyImageNet is now my go-to "verify new ideas dataset"
Genuine question: are there real-world image recognition tasks that requires training on more than, say, 10 or even 100 classes? I'm personally aware of only one that might come close, and that's because it's an image-based species detection module, and it's whole purpose is actively trying to be able to recognize a large number of very specific subgroups. But most of the other I can think of get maybe a couple dozen, and sometimes even as few as 4 or 5 classes where they're useful, and the accuracy within those cases is much more important than the sheer number of possibilities themselves.
I guess I'm just asking if COCO or ImageNet-trained networks are actually noticeably superior for most real-world tasks, or if it's just a metric that's used because the performance differences only show up in the long tail of the distribution.
> I guess I'm just asking if COCO or ImageNet-trained networks are actually noticeably superior for most real-world tasks, or if it's just a metric that's used because the performance differences only show up in the long tail of the distribution.
Given that for any real-world vision task you start from a pretrained model om those datasets they will in fact be noticably superior on the real world task after finetuning. Just because the quality of the features extracted through the backbone is better.
The League of Legends x Apple announcement is bigger than people might think. I 100 % believe that this is the start of a massive shift from pc/console to mobile which we have seen before in China. In fact, it is clear from this announcement that iPhone IS considered a console now for Riot. Huge stuff.
I think you might have it reversed. I think the LoL announcement is mostly catering to the LoL playerbase (which is mostly located in China) rather than a great ploy to steal PC gaming share.
Separately from that, its pretty interesting how obviously faked the "playing" of LoL in the video was, in an presentation that is otherwise so focused on attention to detail.
It's not the start of a massive shift. You'd be surprised what the gaming spending per platform breakdown looks like - phones are a MASSIVE chunk, and growing by far the fastest.
I don't have the numbers on me to back this up but maybe someone else can fetch them.
The rule of thumb when it comes to sizing gaming audiences is:
- Mobile: Billions
- Console: 100s of millions
- PC: 10s of millions
In other words, a decent mobile game (not Fortnite or LoL, but something like Among Us) can get 100s of millions of users, a good console game can sells 10s of millions of copies and a PC game can sell millions of copies.
According to one source[1], steam alone has 95 million monthly active users. Not every pc gamer has steam running, so actual PC gamers are definitely in excess of 100 million. By comparison, ps plus and xbox live have 103 and 90 million monthly active users respectively. I'd say that PC and console gaming are within the same order of magnitude.
10's of millions of PC gamers? I think you're off by a huge margin. I believe DFC reported that there's over a billion PC gamers in the world in 2020. It's a billion dollar market in itself.
I have a sneaking suspicion that this is partly why Epic made such a big move trying to put an Epic store on IOS (to the point of sueing Apple).
They managed to get Fortnite to run at 120FPS on Ipad Pro just this March and probably realized that high performance gaming on IOS is here and will only get better as CPU\GPUs improve on mobile, and sued Apple by August.
What better way to monetize your immersive, high performance, addictive game when it's in everyone's pocket all the time ie no need to sit at your console or PC.
I think we can both agree that there is a lot of value in knowing the "actual value" (as opposed to speculative value) of things in our society. Knowing the actual value of things is impossible for multiple reasons it (it changes very fast and is just intrinsically very hard). However, given a big incentive (making money) we can let the market estimate the value of goods for our society. This speculative valuation is far from perfect, but it is the closest we can get to the actual valiation of goods/institutions.
Fair point, and same probably goes for other billionaires, who have their money in a variety of investments. That probably does a lot to juice the stock market, pumping up valuations, but it doesn't do as much to generate actual economic returns. Happy to be educated otherwise.
