Expect Deeper and Cheaper Machine Learning | Hacker News Reader

Expect Deeper and Cheaper Machine Learning

spectrum.ieee.org

92 pointsjonbaer16 comments

And the drone manufacturer DJI is already using something akin to a deep-learning ASIC in its Phantom 4 drone, which uses a special visual-processing chip made by California-based Movidius to recognize obstructions. (Movidius is yet another neural-network company recently acquired by Intel.)

Wow, this is just so wrong. The Movidius Myriad chip that's used by DJI is not an ASIC. It's a programmable vision co-processor. It was not designed for running neural nets. Movidius is based in Ireland, not California. They are not a "neural network company".

WTH is a "neural network company" anyway?

> WTH is a "neural network company" anyway?

Being able to design a network that appropriately solves a problem may become a valuable skill. I wouldn't be surprised if a startup popped up indicating that they could solve your ML problem.

The Myriad is an ASIC: http://www.movidius.com/technology

Myriad was Ireland-founded but is currently based in San Mateo.

They really meant "machine learning company" not "neural network". But the other two facts are correct.

"WTH is a "neural network company" anyway?"

A company specializing in anything having to do with neural networks. Before the deep learning craze, that was many companies making tools for designing/debugging them, developing ASIC's to accelerate them, further developing the tech if a R&D outlet, or applying them to industrial problems. The number of things such groups might be doing has only expanded with the recent leaps in the technology.

Why would I not expect a greater version of a spreading technology that developers and companies have a great interest in?

I mean telling me to expect it is akin to saying expect the economic realities that have affected the spread of every technology heretofore will continue to work as they have in the past. To which I reply - will do.

I look forward to learned models as a service. Which you can then script to do more intelligent things. This stuff is too low level for me. In fact that's a great idea for a startup. Learned models as a service via an API.

This is what we're working on over at Asteria. It's been fun and challenging. It's still early days, but we're running HITL feedback for model building as well.

You mean like https://cloud.google.com/ml/ ?

More like this https://cloud.google.com/vision/

There are API aggregation sites like Mashape that are already working toward this goal. My startup thing is about using learned models for text processing [1]. Quick tests are also available here [2].

[1] https://market.mashape.com/atrilla/nlptools [2] http://nlptools.atrilla.net/web/api.php

I've thought about that too. I could imagine a near future where people build models and sell access to them in a marketplace, either for direct use or transfer learning - sort of like breeding farm animals. Perhaps in five to ten years neural net farming, or learning agent husbandry, will be common ways for people to earn a living.

Currently, there is a lot of reinventing the wheel. Many models start from scratch instead of transferring already good results from past models. It should be easy to plug vision models, audio processing models, behavior models together, the way we use various libraries in programming.

if everyone is going to earn a living doing the same thing, it'll become easy accessible and understood by everyone and thus it'll defeat its own purpose.

> I look forward to learned models as a service.

My current disappointment with machine learning as a service is pricing. For example, the same job that could be solved by a cheap $5/month instance with open source software, would cost $45/day if it were implemented in Google's Cloud Prediction API. My task was to classify news and I had a volume of about 100K examples to classify per day.

Why would cloud API cost hundreds of times more?

That $45/day vs. $5/month comparison ignores the set up and maintenance costs of your personal instance. You would also have to train models, keep them up to date, etc. If you are happy trading your dev time that way, fine, but the value proposition of APIs are clear.

NVIDIA's P40 and GTX Titan XP already have 8-bit inference instructions.

Reply on news.ycombinator.com