The Man Machine

Written by Julian Shaw

“Man Machine, pseudo human being

Man Machine, super human being

The Man

Machine

Machine

Machine

Machine

Machine

Machine

Machine

Machine

The Man”

And so on…

Never ones to overdo the lyrical content, Kraftwerk surpassed themselves here. But somehow the starkness of the words mixes with the cold-warm, technological-yet-beautiful music to convey the meaning. (Or what I think is the meaning, the existential nature of being a human, how akin are we to simple machines, with our repetitive lives seemingly predetermined to stick on repeat until we fade away? Oh yes, I am feeling cheery…)

You can take that one step further; we live out our lives in repetition, machines operate by repetition when we teach them. Will machines ever be able to learn the routine independently of us? When they do, what will happen? It’s a massive topic, one that invariably leads to dystopian predictions or dreadful films, and that would need a far more learned discussion than I’d ever hope to attempt. But there are some parts of the comparison that I find interesting.

Take Machine Learning –how far will that go? Can we ever teach a machine enough to surpass our own understanding?

Baby Steps, or Droolian Explains It All

As a baby learns by exploration, so we can teach software to explore data and reach hitherto unforeseen conclusions. We lay out educational toys, point the baby at them, shake them about a bit (the toys, not the babies, before you start) to get interest and sit back. Then we look on amazed as the little one picks up a stuffed monkey with a drum velcroed to its chest, works out that the drum comes off, fits it partially in a pre-tooth mouth causing torrents of dribble and makes a huge noise banging on the floor, slobbering drool over your expensive wooden floors.

Same for Machine Learning really, except without the dribble.

Basically, Machine Learning is the process by which we tell machines how to examine data sources, generally massive (“BIG”) data sources, and thence how those results are used to further generate more understanding by the machine – in other words “Learning”.

CRISP-DM https://en.wikipedia.org/wiki/Cross-industry_standard_process_for_data_mining

The CRISP-DM framework is very much a process methodology, based on axiom based decision making.

(In fact, if you go deeper into researching it you’ll see that Process Modelling and Decision Making are crucial to its format. Which makes sense as Machine Learning in this context is aimed at understanding business data, and process modelling is about understanding how a business works. They sort of go hand in hand.)

Although it is more complex, CRISP-DM has six main phases. To my mind these broadly match a software deployment (but that’s another story…) and allow a structured definition to be imparted to the software. Here’s my precis.

Business Understanding

– Get to know the business aims and goals before you even look at data

Data Understanding

– Analyse what the data is; what’s it saying, where is it from? Can you create more information, statistics and data?

Data Preparation

– make sure your data is cleansed, conformed, up-to-date, veracious, viable and visible.

Modeling

– build a model for the software to use; base it on algorithms, statistical techniques and pattern recognition; “train” your software how to interpret and report back on the data. How does it do that? Where does the resultant data go? How should we examine the data about the data?

Evaluation

– how do the results stack up against eventualities; were they accurate; did outliers and anomalies mean anything; what about false / true positives and negatives; how do we reprogram, if at all? At this point we go back to stage one.

Deployment

– Only now can we send the model out into the world. Without all of the prior steps, we could not ‘trust’ our results. 

A Non-Linear Process, or, A Loop-Hadoop-Garoo (…)

This penultimate point is crucial. We have to go back on what we’ve found otherwise we can’t learn. 

The thing with learning is that it’s not just a straight line. You don’t read a book and take it in and then that’s it. You carry that knowledge with you, and in the future you’ll dredge it up and it will inform whatever you are doing, thereby influencing the secondary learning.

That’s how Machine Learning works; that’s how babies learn.

Many moons ago, when I was teaching myself how to code and writing software, I locked myself away in a room, listening to all manner of music. One of the albums that got worn out was Kraftwek's Minimum-Maximum. It has a ‘live’ version of Man Machine. The music and the words sank in. Now, years later, that sub-conscious learning has altered the way I began to examine Machine Learning. I equated machine learning with human learning. I am a LEARNING MACHINE!!

It’s an ongoing circle. The graphic below explains it really well.

As you can see, the results feed back for further analysis and modelling.

Man Machine, or a woman, as the case may be.

The pun was intended. In “Pattern Recognition”, William Gibson used his heroine, Cayce Pollard, to examine the way machines and humans analyse data. Whereas she had a semi-mystical way of seeing patterns, we have to use logic and algorithms. (Or do we? Or does she? Are Cayce Pollard’s insights merely the product of some super-advanced algorithm in her head? And where do our insights come from? Are they any different, albeit far less complex?)

Whatever the process, the neural connections we create from memory, is there nothing to stop us writing down our algorithms and logic, and applying them to life? Other than the vast complexity of the calculations needed – how can a rug-rat quantify that the monkey’s drum is much more fun if it’s left on its chest – and the time it would take to process this, and the Deep Thought like storage space we’d need to contain it all.

Or maybe we could get a computer to do the information gathering and assimilation for us, get machine learning to do the grunt work for us. Like Tom Stoppard explored in Arcadia (https://en.wikipedia.org/wiki/Arcadia_(play) ), we can predict anything, as long as we can write the algorithm for it. So we can tell a machine to learn for us if we can tell it how to learn and what to do with the information it learns. 

And that’s what Machine Learning is – us telling the software what to go looking for and what to do with it when it gets the data.

Something like IBM SPSS is exactly that. Underneath all the mysticism of SPSS and the like is a very logical, ordered, structurally complex system that feeds off itself, generating its own data to be further analysed based on human generated algorithms. But that doesn’t make any less powerful.

When you point SPSS at a selection of data sources it will go off and do its thing, then come back with insights you probably hadn’t thought of. Then you examine the results and go again, and then again, and so on until you are happy (Like the diagram.)

Then, once this process of Machine Learning has been carried out, you put IBM SPSS in place and it will start to predict future outcomes with a startling degree of accuracy. In today’s instantaneous-information driven world, the organisations that react fastest, react most accurately will thrive. And you can almost guarantee the ones that do this best will be pumping massive amounts of data through predictive engines like SPSS that have gone through the process of Machine Learning to provide the results.

Like everyone, I’ll carry on learning in my own leftfield way, baby-steps and strange connections embedding things into my consciousness. And all the while machines will be primed and set free to become more learned, bringing insights we mere humans could never hope to uncover in the mass of data flying around.

Postscript:

Incidentally, I wasn’t aware that SPSS was one of the initial signatories of the CRISP-DM, which sort of tells you that they know what they are talking about. Of course, it’s now much more powerful by being connected to the likes of Cognos TM1, Watson and all manner of Social Media data.