Hacker News Reader

Comment by swah | Hacker News Reader

I'm a layman but I find it quite interesting that a big release such as TensorFlow doesn't affect more people outside Google - or at least thats my impression. One would think, at least, that online store recommendations would become better or something like that.

View on HN

TensorFlow doesn't make the algorithms more effective, it just makes them easier to describe, and recently, more quick to train / test. Also, with the kind of predictions Google is making, it's very unlikely that you'd notice improvements, since they would be gradual.

...but if you want to make your algorithms more effective, you'd probably benefit if they were easier to describe, quicker to train and test, and you'd want to take advantage of gradual improvements. Right?

Not so, for the same reason that low level languages are more effective computationally but less easy to describe and more difficult for code development.

Lua is more low level and has an extremely isolated and fractured community relative to the current Python ecosystem. It is also non-intuitive and has negligible benefits compared to the current scientific Python ecosystem.

I find the abstractions offered by Python and its standard library to be very easy to comprehend, write, and maintain relative to Lua.

..."easier to describe" makes it sound like it's a HIGHER level language, not a lower level language.

Yes, probably. The improvements would still be gradual, though.

This is probably because many people are using pre-existing tools that do very similar things as TensorFlow. To the best of my knowledge, it's not implementing some radical new algorithm with better classification performance, so "online store recommendations," etc. won't become better but may become more widely used.

Why yes of course recommendation engines will get better with these improvements in machine learning. I'd even say that significant, noticeable improvements are not unlikely. Consider image classification, playing Go etc., where RNNs have advanced the state of the art from "research interest" to "reliable product" within a few years.

There may be some reasons why recommendation engines may not advance as fast:

- It's always been more valuable than playing Go, allowing vastly larger ressources to be dedicated to optimizing the current models - Image-processing and NLP both each profited from specific inventions (i. e. RNNs allowing position-independent feature detection in images).

I'm actually looking forward to a time where amazon can recommend me books based on the shoes I bought last year. I don't think I've ever seen a recommendation engine that impressed me.

OP is referring to the existence of mature tools like Theano and Torch. If Deep Learning can help, Amazon is already using it internally, possibly with either their own framework, or with one of the other open-source ones. If something will increase revenue or decrease cost, you can be assured there's already an initiative for it.

Tensorflow is not adding some magical improvements to Machine Learning, it's just one more framework (from a reputable company). The hard(er) part is getting data, cleaning it, testing, making sure it works in production, and updating as data changes.

The business cycle of implementing an open-source takes a while. For instance many e-shops are relying on 3rd party services. In those somebody needs to proactively start playing with TF, then convince the leadership to adopt it and then it will be deployed. As with every other solution, it takes couple of years until the circle closes up (look how long it took for the industry to adopt hadoop, spark, etc.).

It takes a while for the eco-system around TF to get settled. I feel Google wants companies to spring up around TF, and use it in every industry.

I'd say now is the time to start a niche company on top of TF. You'll be acquired faster than you can say "minimum viable product".

Makes sense. If you can make a viable product atop TensorFlow, then acqui-hiring you is probably a good bargain.

Tensorflow is just a development framework. It's not a plug and play solution.

Reply on news.ycombinator.com