Artificial Intelligence and Machine Learning.

konfab said:
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/
Click to expand...

Actually, an academic who doesn't even try make their research available online (or even list that it exists on their academic page) is almost certainly not an active researcher. The Google Scholar results you list support this (very low citation counts, especially for first author titles). 5 citations since 2013? Basically, retirement.

The second person you mention is also barely doing research according to Google Scholar.

I'm not really sure what your point is, but you appear to support mine.
 
Last edited:
Nefertiti said:
And today this was added to Humble bundle :D

https://www.humblebundle.com/books/artificial-intelligence-books?hmb_source=humble_home&hmb_medium=product_tile&hmb_campaign=mosaic_section_1_layout_index_1_layout_type_threes_tile_index_1
Click to expand...

Thanks - bought them. The books look like a good introduction - something I've found with this topic is that reading multiple books is less redundant than most topics, since the approach of a given book is so often completely different.
 
cguy said:
Thanks - bought them. The books look like a good introduction - something I've found with this topic is that reading multiple books is less redundant than most topics, since the approach of a given book is so often completely different.
Click to expand...

I've found these bundle deals usually contain crap (bought some in the past on other topics) - from a quick glance do you think it was worth the money?
 
xrapidx said:
I've found these bundle deals usually contain crap (bought some in the past on other topics) - from a quick glance do you think it was worth the money?
Click to expand...

It's definitely far far from the quality of a book like Bishop's Pattern Recognition and Machine Learning, or Elements of Statistical Learning, but it's relatively accessible, and it's interesting to see how something that takes 10 pages in some of these books is summarized into two paragraphs. For a dollar, I would say it's worth it to get Statistics For Machine Learning.
 
cguy said:
It's definitely far far from the quality of a book like Bishop's Pattern Recognition and Machine Learning, or Elements of Statistical Learning, but it's relatively accessible, and it's interesting to see how something that takes 10 pages in some of these books is summarized into two paragraphs. For a dollar, I would say it's worth it to get Statistics For Machine Learning.
Click to expand...

Thanks - wondering if the more expensive option is worth it - or if I'm peeing money away. I guess its only $15.
 
xrapidx said:
Thanks - wondering if the more expensive option is worth it - or if I'm peeing money away. I guess its only $15.
Click to expand...

If you're new to the topic and interested in a language specific introduction (C++, R, Python, C++), or technology (OpenCV, Tensorflow), then it probably is. I looked through the R one, and it's not bad.
 
Did a course on Machine Learning for BSc Hons in Computing. I struggled a bit as I've never done stats before. cguy is correct when he says mathematics/stats majors would make the best hires, as ML is based almost exclusively on maths and stats with the end result of creating the most ideal output to fit most situations.

It does ultimately boil down to preparing data and running training sets. I like the online university courses as they focus on preparing scenarios to create solutions. UNISAs course was less than practical and even people that were in the ML field didn't know how to help with the proposed questions in assignments.

Azure has a nice ML interface for starting out and can be quite powerful for both managing the input as well as creating a linked training and processing system. Also done some really good pluralsight courses which made everything very easy to understand.
 
Last edited:
To dabble a bit with this, whats the minimum hardware requirements I need? I'll have to upgrade my hardware and thinking maybe I can get away with only buying a new nvidia gpu (1060, 1070 or 1080). I have an older system with an intel 3770k processor, 16gb of ram and a small ssd in. I am hoping I dont have to upgrade that as well :)
 
Later on in the year when I start playing with this stuff, I might use it as an excuse to upgrade my hardware!

However, I'm thinking that real horsepower comes from the cloud, and that upgrading my home desktop might be a waste of time as far as machine learning goes.

I have an AMD 1600X, 16GB of RAM, 1 x 1070 8GB and 1 x RX480 4GB.
 
Ancalagon said:
Later on in the year when I start playing with this stuff, I might use it as an excuse to upgrade my hardware!

However, I'm thinking that real horsepower comes from the cloud, and that upgrading my home desktop might be a waste of time as far as machine learning goes.

I have an AMD 1600X, 16GB of RAM, 1 x 1070 8GB and 1 x RX480 4GB.
Click to expand...

Cloud costs money :p so I want to avoid it for a little bit for now. Azure has a bunch of tools for machine learning and a course on it, so will dabble with it on that side and write the exam on it. Im already down one azure exam so might just as well try the others :)
 
I get $50 worth of monthly Azure credit through my employer's MSDN subscription, so I plan to use that when I get to taking that Azure ML course.

But there is a personal ML project I want to work on, an idea I've had, and I'd rather do that locally if possible. Just not sure if my machine has the horsepower to do so. I suppose the only way to find out is to build it and see what happens.
 
Nefertiti said:
To dabble a bit with this, whats the minimum hardware requirements I need? I'll have to upgrade my hardware and thinking maybe I can get away with only buying a new nvidia gpu (1060, 1070 or 1080). I have an older system with an intel 3770k processor, 16gb of ram and a small ssd in. I am hoping I dont have to upgrade that as well :)
Click to expand...

Ancalagon said:
I get $50 worth of monthly Azure credit through my employer's MSDN subscription, so I plan to use that when I get to taking that Azure ML course.

But there is a personal ML project I want to work on, an idea I've had, and I'd rather do that locally if possible. Just not sure if my machine has the horsepower to do so. I suppose the only way to find out is to build it and see what happens.
Click to expand...

It does depend on the data set and how well your technique is suited for the data, etc., but one can generally look at the curve of the loss/error function to see whether or not more time will help, and the rate that you're getting diminishing returns. For many data sets, you will be able to see this in seconds or minutes or hours, using just a single home GPU (I recommend something high end like a 1080 assuming you can't get a Volta ;) ). This allows you to do most of your development and testing at home, and then kick it to the cloud when you're ready to scale.
 
The price of the volta though :eek:

but by the time I have monies maybe there's a cheaper version :p
 
cguy said:
It does depend on the data set and how well your technique is suited for the data, etc., but one can generally look at the curve of the loss/error function to see whether or not more time will help, and the rate that you're getting diminishing returns. For many data sets, you will be able to see this in seconds or minutes or hours, using just a single home GPU (I recommend something high end like a 1080 assuming you can't get a Volta ;) ). This allows you to do most of your development and testing at home, and then kick it to the cloud when you're ready to scale.
Click to expand...

Thanks.

Yeah I'm nowhere near ready to start working on it. I want to cover a few more courses on basic machine learning stuff before even attempting it.

What performance metric is important for machine learning performance? Talking about model training here, not inference. Is it double precision performance? If so, neither of my GPUs really has a lot of double performance grunt.

In any case, it would be an interesting exercise to see if I can even build a model and get it to run on any GPU, even if the performance is terrible. Since obviously I am a newbie, I am not setting my expectations too high.

Nefertiti said:
The price of the volta though :eek:

but by the time I have monies maybe there's a cheaper version :p
Click to expand...

The problem is that Nvidia segments their market very well. They will never have a consumer GPU with tensor cores in them, nor will they ever make another consumer GPU with decent double precision performance. AMD's Vega is not bad in double precision performance because it was made with two markets in mind. AMD does plan to release an updated Vega next year some time - time will tell whether this will retain the good double precision performance of the current Vega chip.
 
This Stanford University course, and this teacher in particular, were recommended to me by a friend who works in data science (for Google):

https://www.coursera.org/learn/machine-learning
This course provides a broad introduction to machine learning, datamining, and statistical pattern recognition. Topics include: (i) Supervised learning (parametric/non-parametric algorithms, support vector machines, kernels, neural networks). (ii) Unsupervised learning (clustering, dimensionality reduction, recommender systems, deep learning). (iii) Best practices in machine learning (bias/variance theory; innovation process in machine learning and AI). The course will also draw from numerous case studies and applications, so that you'll also learn how to apply learning algorithms to building smart robots (perception, control), text understanding (web search, anti-spam), computer vision, medical informatics, audio, database mining, and other areas
Click to expand...
 
Ancalagon said:
Thanks.

Yeah I'm nowhere near ready to start working on it. I want to cover a few more courses on basic machine learning stuff before even attempting it.

What performance metric is important for machine learning performance? Talking about model training here, not inference. Is it double precision performance? If so, neither of my GPUs really has a lot of double performance grunt.

In any case, it would be an interesting exercise to see if I can even build a model and get it to run on any GPU, even if the performance is terrible. Since obviously I am a newbie, I am not setting my expectations too high.



The problem is that Nvidia segments their market very well. They will never have a consumer GPU with tensor cores in them, nor will they ever make another consumer GPU with decent double precision performance. AMD's Vega is not bad in double precision performance because it was made with two markets in mind. AMD does plan to release an updated Vega next year some time - time will tell whether this will retain the good double precision performance of the current Vega chip.
Click to expand...

One can get away with single precision if you understand your data, chunk its processing correctly and have well calibrated feature set (I.e., features aren't fat tailed). Double precision is a safer bet when you don't know. Once again, it depends on your dataset and tools used - many tools may hierarchically accumulate sample covariances and such so that you don't end up adding tiny numbers to big ones unless your features are badly behaved.
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter
X