Digital Transformation
cancel
Showing results for 
Search instead for 
Did you mean: 

One of HPE’s Chief Technologists on “models versus machine learning”

mikeshaw747

Over the last few weeks, I’ve recommended to a number of people that they take a look at an excellent video by the Chief Technologist for HPE’s high performance computing division, Dr Goh, about the difference between modelling and machine learning, specifically deep learning. The video is here

dr goh picture.jpg

The meat of the video is only 24 minutes long, but some people have told me they don’t have 24 minutes to spare. So, if you are in a similar situation, I’ve done my best to summarize the video below.


The past, we have created models of things, and then used these models to make predictions. For example, we might create a model for heart disease and then use this model to predict how likely someone was to have a heart attack.


Machine learning guesses at models

Machine learning is the exact opposite of this. With machine learning, we do pretty much the exact opposite. We literally throw masses of real-world examples at a the training part of machine learning (see my previous blog part on how a machine learning system is arranged). 

For example, we might throw 1.3m patient records at the machine learning system. In some, the patient ended up with hearth, in others, they didn’t.  Or, if we were trying to train a machine learning system how to recognize pandas, we would “throw” a million pictures of animals at the training system. Some would be of pandas, some not - we would tell the training system the “outcome” for each picture (this picture IS a panda, this picture is NOT a panda, etc). 

A tricky one - red pandasA tricky one - red pandasAnd the machine learning system would guess at what the “model” is. This is,  of course, a brute force approach. It requires a huge amount of data and a huge amount of guessing. Which is why Dr Goh, who is chief technologist for HPE’s high performance computing business, is talking about machine learning - the training part of machine learning often requires the sort of compute power that his team creates. To quote Dr Goh, “you take a huge haystack of data and you get your needle out”. 


Why is machine learning so popular right now?

Dr Goh believes there are three reasons why machine learning, which is certainly not new, is so popular right now..

  1. The massive compute power required for all this guessing is now relatively cheap
  2. The amount of data that we are using to build models is simply too great for humans to process. For example, I was reading the other day how machine learning is being used to cut down on hospital re-admissions. The researchers said that they had to use machine learning because there were too many variables for humans to build a model from
  3. Machine learning is bias free (provided, of course, they you feed it training data in an unbiased way). Humans have biases. Scientists and researchers try very hard not to exhibit biases, but research tells us that this is very tough to do

We have no idea how machine learning’s models work

Dr Goh points out that with the “build a model” approach, you have an underlying model that you can look at to understand how prediction based upon that model works.  With machine learning’s guessing, there is no model for us to look at. 

dr goh go.png

When DeepMind, Google’s AI unit in London, built the world’s Go champions, the DeepMind team were asked, “what is the model that DeepMind was using”. The DeepMind team said, “we have no idea”. When HPE’s high performance computing team beat the world’s best poker players, they were also asked what the model used was, how did the system learn to bluff so well. The answer was, again, “absolutely no idea”. 

This is a little unnerving. If we attach our machine learning to something that can take action, we have no way of knowing if the resultant “system of action” always behave in a way that we consider safe. This is why Dr Goh believes that most of the time, machine learning will be used to advise rather than to take autonomous action. 

Needless to say, Dr Goh’s presentation has a lot more detail, and I encourage you to watch it. You only need to watch up to 24 minutes (up to when the Q&A starts), although the short Q&A is a good discussion on some of the ethical issues associated with machine learning. 


Mike Shaw
Director Strategic Marketing
Hewlett Packard Enterprise

twitter.gif @mike_j_shaw
linkedin.gif Mike Shaw

Mike Shaw
Director Strategic Marketing

twitter.gif@mike_j_shaw
linkedin.gifMike Shaw

About the Author

mikeshaw747

Mike has been with HPE for 30 years. Half of that time was in research and development, mainly as an architect. The other 15 years has been spent in product management, product marketing, and now, strategic marketing. .

Labels
Events
28-30 November
Madrid, Spain
Discover 2017 Madrid
Join us for Hewlett Packard Enterprise Discover 2017 Madrid, taking place 28-30 November at the Feria de Madrid Convention Center
Read more
See posts for dates
Online
HPE Webinars - 2017
Find out about this year's live broadcasts and on-demand webinars.
Read more
View all