Artificial Intelligence and Machine Learning.

Been upskilling myself on ML the past 6 months.

What I've learned:

Ignore the headlines. It seems media reports ML like they report Science. Badly
80% of the work is getting the data in the right format. And getting a lot of data.
20% is running the training model and tweaking it.
Installing the tools feel like installing Linux in the early days. The dependencies in Python is a nightmare. You need to upgrade, downgrade, incompatibilities.
Setup the tools with GPU enabled.
Some people have done the grunt work for you. Tensorflow, Theano etc. for CNN. But you will still need to understand the basics.
Dust off your Linear Algebra and Calculus textbooks.
 
SlinkyMike said:
uh dude... /look of disapproval
Click to expand...

Yeh - I know... didn't really like the idea of spending money on a home PC - but I'm probably going to have to, for this and other projects

Currently I only have a ****ty laptop that I use for work.
 
creeper said:
Dust off your Linear Algebra and Calculus
Click to expand...

Yea, this is my sticking point at the moment. So Im going to learn at coursera/edx/khan academy first to get the hang of it, might take me a while :p
 
xrapidx said:
Yeh - I know... didn't really like the idea of spending money on a home PC - but I'm probably going to have to, for this and other projects

Currently I only have a ****ty laptop that I use for work.
Click to expand...

reason then to get a nvidia gpu :D a 1080 would be just fine :p
 
Nefertiti said:
Im used to working with getting unformatted data into a mssql databases so i know pain :o I have no ML or AI experience however so kaggle looks like a good way to get that type of experience without having to worry about the data too much.
Click to expand...

It's great for that. A working knowledge of Python will also help immensely in the field.

xrapidx said:
Yeh - I know... didn't really like the idea of spending money on a home PC - but I'm probably going to have to, for this and other projects...
Click to expand...

I'm thinking of building the rig in the article I posted. It's over a decade since I built a PC so it will be fun to build one again, and a rig like this could actually pay for itself if you consider what you'd pay AWS for something like this over time. YMMV based on how much you'll use it and how many projects it is suitable for.
 
Nefertiti said:
reason then to get a nvidia gpu :D a 1080 would be just fine :p
Click to expand...

Had my eye on carbonite for a "cheap" midlevel PC for about a week now - apparently cheap isn't what I thought it was. :p

SlinkyMike said:
I'm thinking of building the rig in the article I posted. It's over a decade since I built a PC so it will be fun to build one again, and a rig like this could actually pay for itself if you consider what you'd pay AWS for something like this over time. YMMV based on how much you'll use it and how many projects it is suitable for.
Click to expand...

Probably the same time frame for me - lost interest in desktop PCs as my job didn't require it - and also don't play games anymore. (tried GTA5 recently - but can't get the hang of a PS controller).

I enjoy home automation, because I hate repetition - so have tons of odd PC connected devices all over - it would be interesting to see what could be done with the data off coming off that. Identify patterns for automation, etc.
 
Last edited:
startups dont have a lot of money though so unless they have deep pockets I dont see how they can offer more... they will probably go with graduates or junior guys with that salary
 
Nefertiti said:
startups dont have a lot of money though so unless they have deep pockets I dont see how they can offer more... they will probably go with graduates or junior guys with that salary
Click to expand...

Posted it because I'd been monitoring the company for awhile - I like the way that Michael Jordaan operates - I was actually going to apply a couple of months back when they just had the one ML challenge, and no mention of salary.
 
xrapidx said:
Posted it because I'd been monitoring the company for awhile - I like the way that Michael Jordaan operates - I was actually going to apply a couple of months back when they just had the one ML challenge, and no mention of salary.
Click to expand...
See the website is active from October 2016. They do have a nice group of people associated/working for the company so they probably know what they are doing :) Why not apply and see where it takes you? :D
 
cguy said:
I've heard of a few companies that use it in SA - mostly market research trying to get insight from client data. Very little on improving the state of the art or even doing large scale ML. The universities don't appear to be big on it either - the focus there is often on very abstract aspects of AI as opposed to ML.
Click to expand...

Wat?


http://www.ee.sun.ac.za/research/signal-processing/
 
konfab said:
Wat?

http://www.ee.sun.ac.za/research/signal-processing/
Click to expand...

There's almost nothing on ML there. Here's their publications page(???): http://www.dsp.sun.ac.za/research/list-of-publications/

The "Signal Processing & Machine Learning" group appears to do very little machine learning at all. Looking at their research staff, only two guys actually do ML, one only got his PhD in 2016 from Edinburgh, and hasn't published a single paper while in SA, or with his SA colleagues. The other is the lab head who appears to be into HMM's, but doesn't have his publications listed anywhere but a CV from 2009. Also, I'm not saying that SA doesn't do anything, just that it's somewhat half-assed pretty much everywhere I look.
 
Messugga said:
Depends what you want to achieve. cguy can confirm, but I suspect it's rather unlikely that you'll get very far on coursera courses, when interview at boutique investment firms that run algos and similar. So at some point, you need a more formal qualification.
Click to expand...

SlinkyMike said:
Nonsense.

Come work in my business, I hate PhDs and we do plenty of ML, Data science, etc.

DISCLAIMER: only scrappy punks who regularly overflow with passion allowed.
Click to expand...

It really depends on the level one is working at, and what one is trying to achieve. We've almost never actually hired someone with an "ML" PhD. Most of our ML experts have their PhDs in stats, maths or physics (who now average over a decade of data analysis/science, ML/prediction work at the highest level each), and the emphasis is looking at the data and predictions, and figuring out what factors are preventing optimal prediction, what optimal prediction actually is, how the data should be transformed, how the underlying algorithms should be changed in order to accommodate this, and how this can be done efficiently in a performance competitive live environment, and also how training for our system can be done at scale.

The optimal training is typically going to come from having a very solid mathematical and/or statistical background, and for those implementing the production code, an excellent CS background with low-level performance skills, in conjunction with the theoretical background.

For many, the job involves cleaning data, and using pre-bottled ML algorithms/APIs (TensorFlow, SciKit, Amazon/Microsoft's ML libaries, OpenML, etc.) (e.g., Creeper's post), and although there is debugging required to get it to work, the kind of statistical debugging required to make novel breakthroughs in prediction quality, and the tight run-time and training requirements that require custom algorithms are often not necessary, and don't need the same kind of people.

So basically, like many things in the world, there are those who build the tech, those who use it, and those who use it and can build their own tech as needed. Coursera courses are fine for most users, and someone who runs through all of them and fully understands their content, will probably be way ahead of the hoards of people using ML tech naively, but they're probably not going to figure out what change to a MLNN learning algorithm is going to improve their out of sample prediction for their particular data sets.

Edit: after replying to Creeper's post, one other thing to emphasize is that the type of work done for the first pass (using existing libraries and tweaking parameters), is very different from what one would do if their goal was to continually improve prediction quality and performance thereafter. For some applications, the first pass may be sufficient in which cases the more advanced skill sets aren't really required at all. For competitive ML (e.g., finance), where better prediction is always an edge, it is an absolute necessity.
 
Last edited:
creeper said:
Been upskilling myself on ML the past 6 months.

What I've learned:

Ignore the headlines. It seems media reports ML like they report Science. Badly
80% of the work is getting the data in the right format. And getting a lot of data.
20% is running the training model and tweaking it.
Installing the tools feel like installing Linux in the early days. The dependencies in Python is a nightmare. You need to upgrade, downgrade, incompatibilities.
Setup the tools with GPU enabled.
Some people have done the grunt work for you. Tensorflow, Theano etc. for CNN. But you will still need to understand the basics.
Dust off your Linear Algebra and Calculus textbooks.
Click to expand...

That is literally < 1% of what we do. The vast majority of work for us is figuring out how to improve our predictions beyond what a library and clean data would give us. What you have described, is the workload of what someone building an ML system would do in their first iteration.
 
cguy said:
That is literally < 1% of what we do. The vast majority of work for us is figuring out how to improve our predictions beyond what a library and clean data would give us. What you have described, is the workload of what someone building an ML system would do in their first iteration.
Click to expand...

Therefore the upskilling part ;)
 
cguy said:
I've heard of a few companies that use it in SA - mostly market research trying to get insight from client data. Very little on improving the state of the art or even doing large scale ML. The universities don't appear to be big on it either - the focus there is often on very abstract aspects of AI as opposed to ML.
Click to expand...

My uni we are doing practical, but more due to it being my research project and having a professor who is quite interested in it.

Did have Intel going to all the universities marketing their student kits for machine learning, wasn't much practical use besides learning to recognize numbers, no real use case that interests the majority. I have talked with some companies that are implementing it for analysis of banking transactions, can't mention names though :(.
 
cguy said:
The other is the lab head who appears to be into HMM's, but doesn't have his publications listed anywhere but a CV from 2009. look.
Click to expand...
That is because there is this thing called Google scholar, which is an important tool for the academic who doesn't bother to update their CV.

https://scholar.google.com/citations?hl=en&user=uAJ515sAAAAJ&view_op=list_works&sortby=pubdate

The ESL at Stellenbosch also dabbles in Machine Vision
http://staff.ee.sun.ac.za/cvdaalen/index.php/research/
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X