The most important part of machine learning is not the machine
At least since the year…“2000”….
Web development used to be cool or possibly due to “the positivity effect” that’s how I prefer to remember it. Every day I would strut into a stuffy suited office space, with a smug smile, in my casual shorts and shirt. I would sit astride of two large monitors and plug myself into the matrix.
People were mystified by what I did and how it worked, I can admit to myself 15 years later that I took much pleasure in that mystification and wonder from my colleagues. Just the tiniest self reflection reveals data science/AI to be the equivalent job for the same reasons today, it seems I like the attention. The only difference now is my experience, I’ve learnt some transferable skills in that time, I’d like to share them.
Know your limitations well or ignore them completely…“2003”…
Back in the days of corrugated carpets and plywood partitions, my junior web development job begun with a ping from Microsoft Office 2000. After a desktop rumbling 10 seconds, an e-mail would present itself pixel by pixel, leaving the download of a huge 10Mb attachment in its wake. Once downloaded, I would scroll through the, 4 page long, Times New Roman, heavily bullet-ed list of solution requirements, vaguely adding up days of dedicated work in my head, multiplying the number of days by 4 and sending the “quote” back to my manager.
Subsequently, I would be asked to join a bravado led, chest beating, dry air conditioned, 2 hour PowerPoint meeting where I would invariably be asked in the last 15 minutes what the limitations of “the server” and “the code” were, “to ensure technical limitations were considered”.
Although I appreciated — some may say lived for — the acknowledgement that my job in some way mattered to others is not the point. By artificially considering my “technical needs” without understanding what my “technical needs” were, just meant limiting the solution they wanted without even helping to make my job any easier.
I was young, inexperienced and quiet, so I didn’t say anything. I kept being fed bad requirements and politely giving crappy feedback in return.
Today, people design solutions utilising machine learning
Aren’t designed screen prototypes basically just the new way of handing over requirements?
They don’t understand machine learning, but they pay lip service by talking about “considering its requirements”. By “considering requirements” that you don’t understand, you’re actually limiting a solution in ways you never had to, you’re designing something that could be better.
If you don’t understand the considerations, it’s better not to consider them at all, design the ideal solution and let the technical experts pull in the limitations. Or do some machine learning code yourself, it’s not that hard honestly.
Designed screens don’t actually — in any way- describe any of the AI/machine learning underneath. The best screen designs cover up the AI and technical complexity. Maybe we need a visual way to design algorithms? I tend to think we do.
More data doesn’t mean a more impressive algorithm…“2006”…
Fast forward to an important and educational number of years in my life, I scraped through a bachelor’s degree in Information Systems, to think back then only 1 of my modules was web based, 1 module — that’s insane.
Along the way I made some awful fashion choices, before half drunkenly stumbling into a job in Germany where PHP was the cool language of the moment. At this point the setting had changed, I was now surrounded by bare foot, casually dressed Germans drinking a beer at lunchtime.
I had what I thought was a lot of experience. I had started to query and question who was giving me these long lists of rambling requirements and why I had to listen to other people tell me what to do whilst beating their chest for 2 hours — what knowledge did they have that I didn’t?
A quick “ask Jeeves” (just kidding I’m not that old) Google search, revealed the early beginnings of a User Experience discipline. Suddenly I had some best practices and objective theories to base web solution decisions on, ones that weren’t based on inputs from a single navel gazing set of people with ulterior motives. I could take information from multiple sources and start to form a balanced opinion on what the outcome of the solution should be. But my solution didn’t make any money, even though users said they liked it.
I’ve witnessed the same things happening with machine learning and AI. Just because your algorithm has historical data available doesn’t mean it makes the right decisions. Machine learning needs to know what data to give what weight to. Not only that, it doesn’t know what it doesn’t know. I didn’t know that I hadn’t considered that what people say and do are different things. I had given too much weight to UX and not enough to business.
Unless you tell it otherwise machine algorithms will do the same thing, algorithms do not have life experience, they have not been conditioned by societal norms. They are as naive and as inexperienced as you let them be.
You don’t have to look far to find examples of this. The most famous being the racist image recognition algorithm.
http://www.bbc.com/news/technology-33347866
Not to mention many times you start to type a search query into google…
Some may suggest that this Google search actually shows Google is experienced — I’ll leave that to the comments section.
The oldest ways are the best…“2011”…
After an enlightening stint in Germany, I begrudgingly returned to the UK and began working as a developer for an agency in London. I still had to wear a suit, but now at least I didn’t have to wear a tie, although I always did for my salary reviews.
I no longer had the stress of an email client, we had web-mail, online docs were not far away and we had started to use post-it notes in the more “creative” meetings. I really thought that I was living in the future, on the cusp of big technological things.
Everything seemed new, we were developing new models, tools, methods left right and center. I spent a mass of time in PowerPoint moving around arrows and shapes with various buzzwords in them, very few of them stuck or got any notoriety whatsoever. I learned a lot, but I don’t think it added much value to company I was working for. I was just repeating old news in new shapes.
There were a few that did stick, the purchase funnel did and also AAARR metrics, which also took off. I didn’t realise at the time that this model was actually from 1898. I started to become acutely aware that many of the best models had existed for years.
If you hadn’t realised the format of this post yet… this finding can also be related very closely to machine learning. Many of the roots of many of the most common machine learning algorithms are at least 20 year old. Although many of the origins are debatable, the table below still illustrates the point that they were conceived before we even knew big data and highly scalable distributed computing was a thing.
A highly debatable origin of algorithms table
This insight could be taken a number of ways. You could suggest that it means we haven’t found the best ways to work with big data yet, but you could say that this means we have really solid foundations in which to work. This further illustrates my point that there is nothing new at the core with machine learning, our technological capabilities around the same core principles may have developed but that’s it.
The hype is unavoidable around new fun and cool technologies, I don’t want that to stop, I like the enthusiasm and imagination it can bring to people. But the people in the industry who are relied on to consult on such things should not get carried away and should keep grounded in the reality.
The future can be at odds with the past when really it should be informing it from the present.
Fast forward…“2017”…
I can’t think of any good analogies that fit with my narrative for the next 5 years so I am conveniently fast forwarding to now. I work in a space where I am helping to combine the scientific with the creative, the machine with the human, the quantitative with the qualitative, ontology with phenomenology, maths with art.
Although my expertise leans towards the machine, data and algorithm side, I always see that people are the challenge, if we get the people part right, everything else falls into place. My past experiences seem to point to this too.
Peoples perception of what data is and what it can be used for has a large impact on the solutions we can design and build. As I’ve also touched on in a previous post about calling “data the new oil” if we don’t understand the nature and power of data or admit what we don’t know, we cannot make best use of what’s in front of us. We as humans must give the AI and machine learning what it needs and tell it what we expect of it with as much clarity as possible. That clarity only comes with understanding and being honest about our own abilities.
The most important part of machine learning is not the machine, it’s the people.
