Startups Democratize Deep Learning So Google and Facebook Don't Own It All | Hacker News Reader

Startups Democratize Deep Learning So Google and Facebook Don't Own It All

forbes.com

47 pointsjonbaer17 comments

"Once you can recognize key objects in images, you can target ads not just on keywords but on objects in an image."

It is sad that so much amazing tech is being used primarily to deliver ads more effectively. Deep learning is not the first, nor will it be the last, technology for which that is the case...but, it doesn't make it any less depressing. I like free markets, but somehow our collective consciousness needs to shift from valuing making money over all other things.

And, it's alarming that many of the most valuable companies primary products are ads. Many companies that make real things (like electric cars and solar panels) are worth vastly less than many of the companies that spit ads about real things at people (like Google and Facebook). Something seems unbalanced about that. It seems so wrong, in fact, that it feels vaguely dangerous, to me, though I can't quite put my finger on why.

Yes, I agree -- seeing our best minds work on our lamest problems is distressing. Though arguably, it's an improvement from the previous condition, in in which our best technology was primarily used to produce weapons of war.

At least some of the products of the social media & web advertising era give people access to information and one another, two fundamental human drives. That's a bit better than "better technology to persuade people to buy stuff."

>seeing our best minds work on our lamest problems is distressing

I've seen this a lot, and it's probably true. But because of cuts to research and capitalism's "market failure" as Gates calls it, when it comes to research, does it really matter? The highly motivated and brilliant people will go off into research anyway.

The fault isn't the people of course, it's the markets.

If it helps, this is primarily what I'm setting out to solve building a sort of cloudera/red hat for deep learning at skymind. An open source google brain for big companies to solve interesting problems is appealing provided the right interface.

"The best minds of my generation are thinking about how to make people click ads. That sucks." – Jeff Hammerbacher, fmr. Manager of Facebook Data Team, founder of Cloudera

What's critical to understand is that he's not just talking about the minority that is directly working on ad technology. He's talking about everyone who works on an ad-supported website that depends on page views, which is nearly all of them. Since revenue is proportional to page views, the driving concern is to increase page views rather than maximizing end user satisfaction, which only has secondary importance as one factor in page views, and often not the largest factor.

As to the free market, advertising distorts it very badly. The free market is based on consumers voting with their wallets. What happens when the wallets are held by advertisers?

And for what do we accept the above state of affairs? So that we can get content and services for free? Then we are all suckers! It's an absolute lie that ads give us anything for free. It actually makes the web much more expensive: https://news.ycombinator.com/item?id=7485773]

Another thing that feels so wrong - people who save lives (doctors, cops, EMTs etc), people who shape the next generation (teachers) etc make a pittance compared to sports persons, people who make shitty music etc Something is so wrong about our priorities as a species (not saying sportspersons are worth less, simply pointing out teachers/doctors etc are worth much more than they get credit for)

Deep learning is an awesome topic. The problem I find isn't the concept or the maths. It is the enormous amount of damn data janitorial work that I have to process everyday.

Oh data didn't come back in a clean format? 10 hours gone.

All these text need to be tagged? 2 days

What categories do these images fall under? Another 3 days spent tagging

Wanna parallelize your algorithmn? Ok, good luck learning how to deploy to multiple machines on AWS.

I will pay easily for a service that allows me to run my analysis in parallel without any sysadmin knowledge.

What would you like to do? I made deeplearning4j for this express purpose? Don't need hadoop, command line aws (just specify the number of machines), with a consistent data pipeline interface usable in either single threaded or distributed mode. Happy to take feature requests and answer questions.

It looks like Azure ML will scale out and run parallel code paths simultaneously automatically - any feedback?

http://blogs.technet.com/b/machinelearning/archive/2014/06/2...

I love deep learning and I am sympathetic to the concern that too much of it will end up patented. However, these garbage popsci tech journalist articles need to slow down. I can't imagine anyone who doesn't already love it not starting to hate deep learning after all the garbage articles about it.

I think we should revisit the meaning of the word "democratic."

I think its a relevant description. Technology like deep learning shouldn't be the domain of large companies and graduate researchers.

A friend is working on DeepLearning4J (http://deeplearning4j.org/) which is an open source distributed deep learning package.

Thanks for mentioning me! Admittedly a lot of work to be done. Looking forward to a stable release here soon that will hopefully allow easier access to models while not compromising on speed and ease of use. It can be hard problem to solve and I myself have learned a lot along the way.

> AI For Everyone: Startups Democratize Deep Learning So Google And Facebook Don't Own It All

> In particular, several companies aim to democratize deep learning by offering it as a service or coming up with cheaper hardware to make it more accessible to businesses.

Couldnt spot the "democratization" of AI here (unless democracy only to those that can pay?); also, given the current trends, small interesting startups get acquired by the big ones early.. and we end in big monopolies again..

The only way to runaway from this cycle is open-source.. otherwise we would still using windows XP and IE6 by now

Until good open-source initiatives can happen, we will be lock in the hands of a few forever.. I hope we still have some open-source heroes left out there..

I discussed this very concept at[1]. The industry as a whole could really benefit from representation learning. There's a lot of unstructured media to analyze. What's missing I think is the proper interface to do so. I think the tools will emerge in the next few years that will allow us to do this.

[1]: https://gigaom.com/2014/08/11/gigaom-meetup-talk-learn-and-h...

Everytime I think about deep learning, I am in awe of how important mechanical turk is.

A "turker" essentially a "probabilistic", "general purpose" classifier.

Don't forget Microsoft. We are way ahead on using DNNs in speech and training those with GPUs.

Reply on news.ycombinator.com