|
ML is the new Big Data, and you're an ignorant rube if you don't see how it will immediately solve every problem
|
#
?
May 24, 2019 04:29
|
|
|
|
| # ? May 12, 2024 02:39 |
|
|
Maybe we have to combine ML with Blockchain to get there. It's my idea I get 50% of all profits arising from it cuz ideas are the hard part.
|
|
#
?
May 24, 2019 06:59
|
|
|
Che Delilas posted:Maybe we have to combine ML with Blockchain to get there. You have to patent it first. Otherwise i do it and I'll even sue your rear end if I catch you using it.
|
|
#
?
May 24, 2019 12:17
|
|
|
Che Delilas posted:Maybe we have to combine ML with Blockchain to get there.
|
|
#
?
May 24, 2019 12:38
|
|
|
shrike82 posted:I do consulting on machine learning with a focus on NLP. It's bizarre how the role "data scientist" or "ML engineer" has been diluted to mean nothing. I just did a project with an org where my counterpart was nominally an ML engineer but his background was in front end development and his contributions were largely parroting the info I fed him to management. (chances are what they actually need is pretty simple and does not need machine learning, and maybe they should let me just make those judgments instead of thinking "Can we build a model that...." because that's the wrong way to go about any of this!) I exist solely so they can say they have a data science team. I do want out into a general engineering role though, unless I magically can find a place with good data product management or where they pay me to do that job. Ghost of Reagan Past fucked around with this message at 12:53 on May 24, 2019 |
|
#
?
May 24, 2019 12:50
|
|
|
Vulture Culture posted:I have a great idea for an app that will make doctors a thing of yesterday using AI. I'm not a coder, but I'm prepared to pay you generously in a small portion of the stock *eye twitching intensifies*
|
|
#
?
May 24, 2019 13:59
|
|
|
Vulture Culture posted:I have a great idea for an app that will make doctors a thing of yesterday using AI. I'm not a coder, but I'm prepared to pay you generously in a small portion of the stock hmm i wonder why you don't hear about ibm watson for radiology anymore. could it have been a retarded idea that failed horribly?
|
|
#
?
May 24, 2019 14:18
|
|
|
I met someone from Google whose job largely involved meeting with different teams and explain why they shouldn't be using machine learning.
|
|
#
?
May 24, 2019 15:00
|
|
|
ultrafilter posted:I met someone from Google whose job largely involved meeting with different teams and explain why they shouldn't be using machine learning. The job itself is a reward.
|
|
#
?
May 24, 2019 15:11
|
|
|
I think they should train a neural network to search internal code bases for ML usage and send an email to the team responsible if found. 
|
|
#
?
May 24, 2019 15:14
|
|
|
Protocol7 posted:I think they should train a neural network to search internal code bases for ML usage and send an email to the team responsible if found. You joke but someone probably did this for a hackathon recently.
|
|
#
?
May 24, 2019 15:32
|
|
|
https://twitter.com/deech/status/1131701210979930112
|
|
#
?
May 24, 2019 15:35
|
|
|
I see the data science and machine learning thing so much too that I see it's a buzzword, but I do have a crazy--albeit really sad--story about a data scientist from around 2011-2012. This is at a semiconductor company with lots of electrical engineers around, and I'm in a supporting side for software. So I see a lot of situations where electrical engineers try to put on a different hat and it ended horribly. A friend of mine had just gotten out of helping with some analysis that took six months and cost millions of dollars dealing with a relationship between manufacturing quality, clock speed, voltages deltas, and temperature extremes. It wasn't my discipline so I don't know the details. I had been at the local university to try to find some people from different disciplines from electrical engineering and manages to snag the resume of somebody graduating with a masters in data science that had their bachelors in electrical engineering. So we take her in and my friend decides to scrub this whole scenario they did a bit and play it through with her as part of the interview. He doesn't expect her to understand the discipline much at all but just wants to see if she'd be able to grok it enough to have helped solve this problem better. He lays out the initial data and starts explaining the problem. Midway through, she puts her finger on a cell in a table full of very close numbers. "The problem starts here," she says. Which it did. She went on to say why and what automated controls could have been put in to the data collection to alarm on it in subsequent products. Forgot six months and millions of dollars: she just punched it in the face in the first minute of this interview! They didn't hire her because HR policy didn't treat data science as a valid degree for which to hire into the engineering groups and they would only look at her last-awarded degree--disregarding her bachelor's in electrical engineering. I'd get check-in emails from her on-and-off for a few years after. She was still looking for work. I later switched to a different organization where an electrical engineer PhD was put in charge of data science. He was responsible for leading development of a tool that showed project completion as a bunch of progress bars. It never was finished.
|
|
#
?
May 24, 2019 15:45
|
|
|
Just sent this to my scrum master and she confirmed it was not agile lingo. TMYK
|
|
#
?
May 24, 2019 15:54
|
|
|
Hmm it says here that you were a scrum master at XYZ corporation for four years. Can you answer what the following terms mean then? (monotone voice) Dumpster fire Tire fire Pull out and nuke it from orbit Oh my god. The blood. The blood.
|
|
#
?
May 24, 2019 16:00
|
|
|
Rocko Bonaparte posted:Hmm it says here that you were a scrum master at XYZ corporation for four years. Can you answer what the following terms mean then? A) The previous company I worked for.
|
|
#
?
May 24, 2019 16:04
|
|
|
Those words definitely apply to my previous employer's implementation of SAFe and we had a lot of turnover in the scrum master roles. Coincidence? Nah.
|
|
#
?
May 24, 2019 16:09
|
|
|
Bruegels Fuckbooks posted:hmm i wonder why you don't hear about ibm watson for radiology anymore. could it have been a retarded idea that failed horribly? https://med.stanford.edu/news/all-news/2018/11/ai-outperformed-radiologists-in-screening-x-rays-for-certain-diseases.html What's dumb is using AI in lieu of radiologists for clinical environments that can afford radiologists (i.e. not developing nations and underserved rural hospitals), but using it to spot pathologies that have been missed by visual inspection is extremely great
|
|
#
?
May 24, 2019 16:28
|
|
|
Vulture Culture posted:Nah. AI can perform equal to or better than radiologists at detecting certain pathologies: No, ML is effective in radiology. I'm not making GBS threads on ML for that purpose. My jab is at watson specifically and their retarded marketing campaign and expectations they set.
|
|
#
?
May 24, 2019 17:03
|
|
|
Bruegels Fuckbooks posted:No, ML is effective in radiology. I'm not making GBS threads on ML for that purpose. My jab is at watson specifically and their retarded marketing campaign and expectations they set.
|
|
#
?
May 24, 2019 17:10
|
|
|
It'd be cool to work on something that actually merits using ML and I think that line of thought is a contributing factor to why it's so over-applied and mis-applied these days.
|
|
#
?
May 24, 2019 17:12
|
|
|
Bongo Bill posted:It'd be cool to work on something that actually merits using ML and I think that line of thought is a contributing factor to why it's so over-applied and mis-applied these days. Not to mention that it is now extremely easy to get started and train some simple models with some simple data and see results almost instantly.
|
|
#
?
May 24, 2019 17:13
|
|
|
TensorFlow is 2019's Hadoop for sure I have all this unemployed time to play with it but I'm learning Rust instead because I'm a dumbass and an unmitigated disaster of a human being
|
|
#
?
May 24, 2019 17:30
|
|
|
Don't you just instantiate some ml package, put your data in and wait while magic happens until the result comes out anyway?
|
|
#
?
May 24, 2019 17:33
|
|
|
Vulture Culture posted:TensorFlow is 2019's Hadoop for sure the actual use cases for practical application of rust are surely more than tensorflow
|
|
#
?
May 24, 2019 18:54
|
|
|
Bongo Bill posted:It'd be cool to work on something that actually merits using ML and I think that line of thought is a contributing factor to why it's so over-applied and mis-applied these days. "It works now but I have no idea why" has 100% of the time come back to bite me in the rear end later and isn't that how ML is all the time. Except when it's not working, or is working but revealing things about you that you didn't want to know. edit: oh why is this relevant. I replied because the rear end-biting experience is why I have zero desire to ever work with anything ML or adjacent like data engineering.
|
|
#
?
May 24, 2019 19:02
|
|
|
BabyFur Denny posted:Don't you just instantiate some ml package, put your data in and wait while magic happens until the result comes out anyway? Only if your company's AI can't do it for you? 
|
|
#
?
May 24, 2019 19:19
|
|
|
Yeah a combination of hiring a bunch of new junior engineers and running out of obvious high-priority feature work has led to an ML-fest on my team the last few months and I'm absolutely not here for it. It's just the type of domain where even if the models are 90% accurate, the 10% is going to annoy the users enough that they will learn to ignore or resent.
|
|
#
?
May 24, 2019 19:22
|
|
|
Teams that dabble in ML tend not to have a proper benchmarking protocol (clean, held out test data set of reasonable side) or misuse metrics (e.g. using simple accuracy with imbalanced datasets), so I tend to be skeptical of self reported claims about performance.
|
|
#
?
May 24, 2019 19:44
|
|
|
Vulture Culture posted:TensorFlow is 2019's Hadoop for sure I have the unmitigated disaster thing going on, too, but if I wanna learn a new p-lang it'll be in lieu of Factorio and welp gotta get off that rock!
|
|
#
?
May 24, 2019 19:44
|
|
|
Does anybody know of some static site generators that work on Markdown while supporting embedded Plant UML? I just thumbed through a bunch of them in the Python ecosystem (just because that's this project's language of choice). Mkdocs made it sound like it could through a python-markdown module, but the directions for that are very Linux-centric and I'm not entirely convinced I can run that on a Windows machine. I then tried Pelican and was turned off by the required metadata in the Markdown files. I decided in all of this that the actual language implementation didn't really matter if I just worked with it through config files or whatever, so now I'm opening it up. I'm just kicking various SSG tools around randomly and hoped for a pointer.
|
|
#
?
May 24, 2019 20:43
|
|
|
Rocko Bonaparte posted:Does anybody know of some static site generators that work on Markdown while supporting embedded Plant UML? I just thumbed through a bunch of them in the Python ecosystem (just because that's this project's language of choice). Mkdocs made it sound like it could through a python-markdown module, but the directions for that are very Linux-centric and I'm not entirely convinced I can run that on a Windows machine. I then tried Pelican and was turned off by the required metadata in the Markdown files. I decided in all of this that the actual language implementation didn't really matter if I just worked with it through config files or whatever, so now I'm opening it up. I'm just kicking various SSG tools around randomly and hoped for a pointer. Why don't you try scrum
|
|
#
?
May 24, 2019 23:21
|
|
|
BabyFur Denny posted:Why don't you try scrum Scrum is too Agile for static site generation. Also, I meant to ask that in the general questions thread. I just checked, didn't see it there, freaked out, and thought I posted to military history again. Some day I won't get probated by this, but instead get visited by social services to take me to assisted living or something. Also, I figured out mkdocs.
|
|
#
?
May 25, 2019 00:24
|
|
|
Subjunctive posted:We’re trying to figure out levels for our ML roles and looking at peer companies is a disaster because “ML position” can mean anything from “has an excel formula set that nobody understands” through “used SGD once” to “Jeff Dean blows me every Thursday”. No one's really figured this out yet, but the industry seems to be standardizing on two different roles. One is the ML engineer, who is basically a software engineer that knows the basics of machine learning. The other doesn't have a standard title, but is often referred to as an ML researcher, and they're the ones who know a lot more than just the basics. For the ML engineer, you can pretty much just reuse whatever levels you already have. These people aren't expected to be anything more than basically knowledgeable. Some of them will learn more than that and that's a good thing, but the bar is low here. What's more important in most companies that have these roles is knowledge of how to do things at scale. The ML researcher is a different story, and where exactly you set expectations is going to depend on your company's structure and needs. At the entry level I would expect someone to be comfortable with the various algorithms and the basic theory around them. If they have experience in TensorFlow/Keras/whatever the flavor of the week is, that's great, but it's not something you should really be screening for (and for God's sake don't hire someone just because they know one or two of those packages!). As they get more senior, I would expect them to be able to discuss more of the following topics in depth:
The hard problem that no one has solved is how to build out an ML organization if you don't already have one. You really have to get the first few hires right, but it's not clear how to do that when you don't have the in-house expertise to make those assessments. I don't have any good advice there.
|
|
#
?
May 25, 2019 16:55
|
|
|
I agree with those two categories basically, but want to emphasize that only one of them will be commonly capable of software architecture and development. In my experience all machine learning scientists will be unable to devise methods to perform any of the support necessary to build a machine learning system, namely: How to get data into the system, how to schedule retraining of the system, how to actively update data used for retraining, how to roll back when a retraining fails, how to provide revision control and configuration of models, how to monitor the behavior of the model, how to scale the model for production use, how to release the model beyond their desktop, how to deploy their model through a pipeline, or how to provide integration testing or regression testing for the model before releasing it into a production environment. I have also seen a fair number of machine learning scientists that can provide no business interpretation for the results, so not only will you need a team of five developers to build the infrastructure, you will also need several analysts familiar with your data set to define the inputs and outputs that the machine learning scientist can use. In all honesty, I have tried over the years to keep my eyes open looking for benefits to ML, but it's very difficult to get a good picture of when it is needed verses when it just happens to work for a particular problem that could be solved more simply and directly. There are certainly a few systems that benefit from it, but in many cases a simple statistical model will be sufficient to move your business forward. In my opinion anyway. 
|
|
#
?
May 26, 2019 01:21
|
|
|
PhantomOfTheCopier posted:I agree with those two categories basically, but want to emphasize that only one of them will be commonly capable of software architecture and development. In my experience all machine learning scientists will be unable to devise methods to perform any of the support necessary to build a machine learning system, namely: How to get data into the system, how to schedule retraining of the system, how to actively update data used for retraining, how to roll back when a retraining fails, how to provide revision control and configuration of models, how to monitor the behavior of the model, how to scale the model for production use, how to release the model beyond their desktop, how to deploy their model through a pipeline, or how to provide integration testing or regression testing for the model before releasing it into a production environment. I built a bunch of really flashy looking "machine learning" features on top of extremely simple heuristics and queries. These get called machine learning because engineering management thinks they need to do this to defend my salary, which is a pretty sad state when I don't make anywhere near what I could make elsewhere (for the record I'm a data scientist/machine learning engineer that doesn't actually want to keep doing that poo poo, because I'd rather do something that's less fake).
|
|
#
?
May 26, 2019 01:48
|
|
|
ultrafilter posted:The hard problem that no one has solved is how to build out an ML organization if you don't already have one. You really have to get the first few hires right, but it's not clear how to do that when you don't have the in-house expertise to make those assessments. I don't have any good advice there. It's an interesting problem. I've observed a number of (non-tech core business) orgs launch ML teams and the build-out strategy varies drastically depending on the business/IT stakeholders. I've seen the following :- 1. Business stakeholder-driven initiative: they tend to employ vendors/consultancies wrapped with a layer of internal business analysts + software engineers. 2. Software engineer-driven initiative: the engineering team attempts to upskill itself to cover ML topics. 3. Quant- ("PhD") driven initiative: common in finance with a quant team that hires a bunch of PhDs to cover this. The second approach is the most problematic because software engineers don't have the business knowledge to assess the business value, and don't have the ML theoretical knowledge to assess the limitations of any given approach.
|
|
#
?
May 26, 2019 03:38
|
|
|
PhantomOfTheCopier posted:I agree with those two categories basically, but want to emphasize that only one of them will be commonly capable of software architecture and development. In my experience all machine learning scientists will be unable to devise methods to perform any of the support necessary to build a machine learning system, namely: How to get data into the system, how to schedule retraining of the system, how to actively update data used for retraining, how to roll back when a retraining fails, how to provide revision control and configuration of models, how to monitor the behavior of the model, how to scale the model for production use, how to release the model beyond their desktop, how to deploy their model through a pipeline, or how to provide integration testing or regression testing for the model before releasing it into a production environment. We're currently "adding" a Data Science segment to our company and offerings (we're a small data consultancy that is roughly split into analysts/consultants and engineers), and this is what drives me nuts at the moment. The lady who is tasked with this has a math background and comes with a background in ML, but her technical expertise is nonexistent. However, since she currently falls in the "engineer"-category, it is kind of assumed she will be able to do the software architecture and development parts just the same, because "it's all programming, duh". It's a really tricky problem, because ideally, you'd get people who can do at least two of those roles at the same time, i.e. data science/ML + engineering, engineering + analysis, ML + analysis. But since it's hard enough to find people who are even half competent in any single one of these things, good luck. And if you then consider that we are doing consulting, i.e. you might actually have to engage with customers, and you are in for a lot of fun. I feel that this overlapping of skills is the really hard problem, because otherwise, ML-people build models that cannot be run, engineers build systems without purpose, and analysts build analyses without technical merits.   
|
|
#
?
May 26, 2019 10:42
|
|
|
Thoughts on working for an internal frameworks team at a big company rather than a product team?
|
|
#
?
May 26, 2019 17:26
|
|
|
|
| # ? May 12, 2024 02:39 |
|
|
smackfu posted:Thoughts on working for an internal frameworks team at a big company rather than a product team? Depends entirely upon how much political power they have. Could be chill or could be a nightmare. Basically, do they have the power to say "we'd like to do this for you, we'll put it into the backlog and prioritize it, but teams X and Y also need Urgent Feature Q and P so we're committed to that right now."
|
|
#
?
May 26, 2019 17:33
|
|