Hacker News Reader

Comment by eksemplar | Hacker News Reader

The thing about machine learning is that, unlike with robotics RPA, new graduates in the field are so much better than anyone else.

I'm in the public sector in a Scandinavian country mind you, so it's not like AI experts are lining up. We tried retraining, didn't work. Then we tried retraining some more, this time with outside consultants, didn't really work out either. So we hired a new graduate who had specialized in tensorflow during his degree, who (obviously?) blew our small progress away.

Now this is anecdotal, but being the public sector we share experience openly with each other, and I'm hearing the exact same story from every other muniplicity in our network.

View on HN

Could it be that you simply didn't select the right candidates for retraining. Wanting to do it, is not the same at being able to or good at it.

It's hard to get the specialization that a degree provides you, working on the side with normal work.

Graduates are just better.

They go through a hardcore selection and intensive training for 3-5 years in the field. It's jackpot if he has an internship related to the job.

Picking a random people who wants to reconvert is unlikely to yield any result. Certainly missing the years of training on maths/statistics. Probably missing any experience in any related area.

My point, which I think you validated, is that if you want to retrain people, you need to put serious effort into it. A graduate will have at least 1 full year focusing on the topic and the large part of the remaining time was spend on fundamentals and related topics.

How is that possible? TensorFlow has been released 17 months ago.

It's not. everyone just wants to feel like they are the chosen ones.

Maybe it was a two year masters degree and he spent his last 17 months using it?

A master thesis is only one semester. He would have had plenty of time.

Hello from Oslo

Would be interesting to hear about what you do machine learning for in the public sector. Could you share a bit?

It variates, and it's in the earl stages. One example we've done is sort through 500.000 personal cases, containing multiple documents, some documents containing up to 50 scanned pages to find cases missing a specific law required form.

Once machine learning has OKed a case we used a software robot to archive it automatically.

Because it was an experiment and also a job that had to be done. We did it simultaneously with as a competition of man vs machine.

The man power was a team of expienced case workers working on it 37 hours a week. The machine was two new teams, one in RPA and one in machine learning.

Both proceses ended up taking roughly 3 months, but the "tech" team spent a lot of the time learning how to use the technology and quite a bit training the algorithm.

Once the tech team and algorithms were ready, the actual processing time took 5 hours in the Azure cloud.

Now we have a process for finding specific documents, however, and I'm told it'll take 1-2 weeks to retrain it, meaning it'll be 1-2 weeks + 5 hours vs 3 months next time.

    > One example we've done is sort through 500.000 personal cases, containing multiple documents, some documents containing up to 50 scanned pages to find cases missing a specific law required form.

Forgive my ignorance, but how is this not a case for OCR + indexing? I assume that the form would have some specific texts that can be found once it's processed with OCR, what is there to machine learn?

You're correct, it was mainly OCR, but we categorize this as machine learning. Partly because we use public machine learning algorithms to do it, but also because we're planning to do a lot with OCR coupled with other data sets and a range of IOT initiatives. Things like predicting where we'll need to water trees, prioritize areas for snow shoveling or schools for renovation, how to control traffic flows and a range of other things.

We are doing some more traditional stuff in BI, but a lot of it is relatively secret, in the very early stages or even getting shut down because legislation in the area is changing rapidly.

We tracked citizen movement in our inner city based on wifi from smartphones for instance. Then we compared the data with various attempts at directing crowd flows. With a decent success rate, much higher than before we started using machine learning to score the results. We've killed the project though, because the new EU privacy protection acts makes that sort of wifi tracking illegal.

As for Scandinavia and Denmark in particular, I can highly recommend looking in to people coming out of DTU Compute. There is also a yearly fair (DSE) where companies can meet with students looking for work/internship. I went as a student this year and had great time talking about practical ML with SKAT, the Danish Customs and Tax Administration.

Reply on news.ycombinator.com