rasmi

Earlier discussion here: https://news.ycombinator.com/item?id=16619917

Previous discussion: https://news.ycombinator.com/item?id=13414776

That's a fair point. There are certainly technical challenges involved in bringing this to other domains.

I didn't downvote you, but I think the comparison to Sagemaker misses the point. This is literally just uploading labeled data and getting a finely tuned classifier out. Hyperparameter tuning is neat, and both Cloud ML Engine and Sagemaker have that, but (correct me if I'm wrong), only AutoML actually handles all of the model architecture decisions itself using transfer learning and learning2learn. See here for details: https://research.googleblog.com/2017/11/automl-for-large-sca...

This significantly reduces the level of expertise required to train models, and the AutoML models outperform "expert" human-created architectures.

Compared to ML Engine, where you provide the labeled data and model code, with AutoML, you just provide the labeled data (with your custom labels for whichever domain you may be working in) and AutoML builds the model for you.

"Soon, Cloud AutoML will release other services for all other major fields of AI."

It's a start!

https://cloud.google.com/automl/

Disclaimer: I do ML-based work in Google Cloud, but I am not on the AutoML team.

The post says there is transfer learning involved, which means in practice you need much less data than you would if creating a classifier from scratch. Of course, more (good) data may yield better results, but it seems one of the goals behind this release is specifically to give custom (your own labels, not just generic object detection) high performance image classification to those who don't have access to Google-scale training sets.

Take a look at TFGAN -- it might make these implementations much easier!

https://research.googleblog.com/2017/12/tfgan-lightweight-li...

https://github.com/tensorflow/tensorflow/tree/master/tensorf...

I see what you mean. To some companies, ML Engine's cost as a managed service may be worth it. To others, spinning up a VM with TensorFlow Serving on it is worth the cost savings. If you've taken other approaches to serving TensorFlow models to get around ML Engine's per-prediction cost, I'm curious to hear about them.

You can deploy TensorFlow model binaries as serverless APIs on Google Cloud ML Engine [1]. But I would also be interested in seeing a TensorFlow Lite implementation.

[1] https://cloud.google.com/ml-engine/docs/deploying-models

Disclaimer: I work for Google Cloud.

TensorFlow is introducing eager evaluation: https://github.com/tensorflow/tensorflow/tree/master/tensorf...

https://twitter.com/oshtim/status/879453382326353920

Disclosure: I work for Google. Google Cloud does this with Dataflow and ML Engine in a fairly simple way, and the internals are open sourced as Apache Beam and TensorFlow so you can migrate to other platforms if you want.

I purchased this book on sale at the Kindle store to get a detailed sense of what his methodology was, and it seems he just uses Google Trends to compare search volumes across locations combined with Google AdWords to give the data a numerical scale and break down the data demographically. It seems statistically dubious at best, especially the part where he makes assumptions based on search autocomplete results (which vary based on location, IP address, browser, OS, and other factors).

You can read more about his methodology in his papers here [1]. He wrote his PhD thesis on "Essays Using Google Data". I can't think of a better way to do these estimates with the limited data provided, but I'm just not convinced it's a strong enough proxy for the conclusions he's making.

[1] http://sethsd.com/research/

It's also a bit risky, as some University administrators may not understand that he isn't actually being malicious. Automated scraping can be considered a violation of some IT policies though, so OP should be careful about something like this.

Something similar has happened with Transmission's download DMGs being replaced on their servers [1] (twice! [2]) in recent memory.

[1] https://news.ycombinator.com/item?id=11234589

[2] https://news.ycombinator.com/item?id=12403768

If anyone is interested in learning this content through an in-depth video tutorial, I highly recommend Dan Abramov's two-hour "Getting Started with Redux" [1] and the excellent follow-up "Building React Applications with Idiomatic Redux" [2]. This is a great article, but learning Redux more thoroughly directly from the creator himself may be of interest to some!

[1] https://egghead.io/courses/getting-started-with-redux

[2] https://egghead.io/courses/building-react-applications-with-...

The job description above is for internal user-facing tools. Mission-critical systems are done in C++.[1]

[1] https://www.reddit.com/r/IAmA/comments/1853ap/slug/c8bpr00

This toolbox allows you to visualize what neural networks "see" at each layer, and what image patterns activate what sorts of neurons (and how). For example, in this video [1] you can see neurons that detect edges, faces, and text.

[1] https://www.youtube.com/watch?v=AgkfIQ4IGaM

This is one of the most concise and accessible explanations of fine-tuning CNNs that I've come across. I hope someone finds it as helpful as I did.

Probably rainbow tables: https://en.wikipedia.org/wiki/Rainbow_table

It is possible to correct for all sorts of different lighting conditions. See here: http://grail.cs.washington.edu/projects/timelapse/

You would need a significant number of images for a particular scene (more than Google Street View would provide -- maybe in the tens or hundreds depending on the complexity -- but perhaps mining Google Maps photo uploads for a particular location would suffice). A Google Image search for a particular landmark is likely to yield better results ([1], [2], [3]). Also, stitching together multiple scenes after reconstructing each individual scene may prove difficult. Apple Maps does a sort of satellite view projection on to 3D building models ([4]), but this is far from a dense reconstruction.

[1]http://grail.cs.washington.edu/projects/timelapse/

[2]http://grail.cs.washington.edu/projects/timelapse3d/

[3]https://grail.cs.washington.edu/rome/

[4]http://www.theverge.com/2015/9/29/9413583/apple-maps-3d-flyo...

Agreed. The name made me expect an official "GitHub for Audio" type service, something like: https://splice.com/

FYI, the latest Jupyter + matplotlib have interactive plotting inline with "%matplotlib notebook". The server must be running for this to work (so you can't interact with static ipynb files in nbviewer, for example), but a static version of the plot is generated for static notebooks. There are also some changes coming down the line to expose the traits of matplotlib plots so that libraries like mpld3 can be integrated more seamlessly.

(Also, Hi Christian!)

Also check out 18F [1], which focuses on pushing forward how tech and the government work together. They run a 12-month Presidential Innovation Fellows program [2] which works more intensely on innovating with specific groups within the federal government. USDS is more focused on modernizing and ensuring some basic technical functionality of a lot of the core aspects of the government (which is incredibly important and impactful, but is less about innovation -- see a comparison here [3]).

[1] https://18f.gsa.gov/

[2] https://www.whitehouse.gov/innovationfellows

[3] http://ben.balter.com/2015/04/22/the-difference-between-18f-...