Comment by kmax12 | Hacker News Reader

Comment by kmax12 | Hacker News Reader

I think this highlights the disconnect between AI researchers and industry. This research focuses on image data and then makes the claim that they want to use it to enable millions of businesses to solve problems with machine learning.

However, image data isn't what most businesses are using for machine learning. They use relational datasets. If you look at the recent "State of Data Science" by Kaggle, relational data is ~3x more common than image data in every industry besides academia and the military [0]. While Google wants to 1000x the number of organizations using AI, they aren't focusing on the problems companies actually have.

Basically, academics love building AI for images, but what companies really need are better ways to build AI systems on tabular and relational data. Images will be a piece, but shouldn't be the focus.

Disclaimer: my company develops an open source library called Featuretools [1] to automate feature engineering for data with relational structure.

[0] https://www.kaggle.com/surveys/2017

[1] https://github.com/featuretools/featuretools

Because that's not a place where deep learning shines. Most of the simple tabular data competitions are won by simple methods like random forests and manual feature engineering. Even simple linear regression or at least SVMs will usually get reasonably close to the best possible model.

But images are more important in the long term. Robotics has been held back for decades because of lack of good AI. You could build robots that can do incredible things, but they could only perform rote actions. They were blind and couldn't see the objects they were interacting with.

Now that we have decent machine vision and reinforcement learning, there will be many more interesting applications of robots. Automation will be a lot cheaper and more convenient.

This is really great. I actually struggled to find ML solutions for a large relational data set at a previous employer. Ended up figuring out a way to flatten the data, but was not ideal. Glad that people are working on this, keep it up!

Presumably that's because relational data is much easier to work with when you aren't using ML, so most companies have a lot of pre-existing data to work with and can jump right in and start using ML without collecting a lot of new data.

Image ML will take much longer to catch on because companies haven't historically had much of a reason to collect image data in quantities that are too large for humans to deal with manually, so they are pretty much starting from scratch when it comes to data collection.

They also did experiments on the Penn Treebank dataset - i.e. text data. This was 53% to RDBs 65% in the Kaggle survey, so certainly not an obscure problem...

Yep, text data appears to be widely used too. Hard to tell from the Kaggle survey, but a large percentage of text data I see in industry is part of relational data or connected to tables in a relational database. Things like forum comments, product descriptions, etc

Visual input is the primary means we humans consume information. We use while driving and reading and for everything else. Clenching to the data structures and information storage/encoding probably not the way forward.

in the terms of intelligence, it's just one of 6 senses. MEMS sensors, signal filters, and coding schemes (compression) are going strong, too. Touch is still problematic I understand.

In that regard, flexible LED screens could be interesting, using the photoeffect in the diodes like a camera, lighting the area at the same time and maybe integrate a resistive touch grid matrix, too.

They used image data because they had to do something first, but I don't think the techniques are at all limited to image recognition.

Reply on news.ycombinator.com