
Sponsored by: NVIDIA
Industrial AI - Part 2: “Dirty, Dangerous, Difficult & Dull” - The Case for Ethical AI Automation.
Why AI should tackle ‘The 4 D’s’ - Dirty, Dangerous, Difficult and Dull.
Like the ‘4 Horsemen of the Apocalypse’, I think the 4 Ds are a good ‘litmus test’ on whether we as a society should be ‘replacing humans with robots’. The concept has roots in the Japanese idea of “kitanai, kiken, kitsui”, (meaning respectively 汚い "dirty", 危険 "dangerous", きつい "demanding"). The 4th ‘D’, Dull, only being added relatively recently.
Data Scientist, Hans Rosling gives some fascinating provocations about just how dire these 3Ds are to our life expectancy; where most of the world in 1800s would not live beyond their 40th birthday, and it’s very clear that the cause was Dirty, Dangerous and/or Demanding work.
Above: Hans Rosling’s Interactive Data Visualisation Tool: “Gapminder”. (link)
This means that the work at Lion Vision most certainly satisfies alleviating the burden of those 3D, as handling partially damaged, or leaking Li-Ion batteries is certainly Dirty, Dangerous, and Demanding. Even when filming, the smell is repugnant, and we all wore gloves for safety, and had a fire extinguisher at the ready! I can see no reason why anyone would insist Lion Vision be disregarded, and humans reinstated. If you have any doubts, please read about the possible skin, eye, and respiratory damage caused by prolonged handling of such waste (to humans, not robots - they’re fine).
The tension of course arises from the fear of unemployment, and is likely to be contentious to any companies and countries who feel it necessary to use human labour for the foreseeable future, either because it is genuinely too difficult to get a robot to do the human’s job (perhaps because of a particular dexterity, multifaceted and complex intelligence, or nuanced empathy and compassion, etc.), or because it is simply cheaper than the cost of AI/ML and Robot alternatives.
“Gradually, Then Suddenly”. (Why Disruption is Hard To Predict).
However, this is where technology can ‘upend’ society when we least expect it. Like me you may have been tinkering around with Arduino and RaspberryPi computers for some years. They are fun, handy, and certainly do ‘automated tasks’ with aplomb. However, in the grand scheme of things, they are a ‘cost down’ on what Industry has had available at a cost prohibitive to the general public, and certainly school kids. Yet, we now have most 8-16 year olds learning a bit of Python coding language, and indeed, this too can control a Machine Learning computer like an NVIDIA Jetson Nano.
The ‘disruptor’ here is not the ‘computer’ but the ‘compute’.
If you’ll excuse the irresistible wordplay, ‘compute’ is what is often referred to as the act of a Machine Learning computer ‘thinking’. The ‘cost of compute’ has been compared to that of Moore's Law (the size and cost of silicone chips have been falling exponentially year on year), and similarly, we are at the dawn of a new ‘Machine Learning Arms Race’, with the cost of ‘compute’ now being slashed year on year.
The difference between ‘automatic doors’ and ‘machine learning detection of cancer’ (and the ‘machine learning drug synthesis’ to cure said cancer), may seem a quantum leap to laypeople, but to data scientists and machine learning professionals, this is the moment they have been waiting for. We just handed a laser-guided missile to people who were already a crack shot with a rifle. We now hit targets with more precision, more speed, and more intelligence than all but Sci-Fi fans could have imagined 50 years ago.
Many engineers are familiar with the apocryphal story of ‘putting a man on the moon with less processing power than is in a modern school-kid’s graphical calculator’, and I’m sure there would be something equally mind-blowing one can say about the scale of a NVIDIA Jetson Nano, below, which is a $200 computer with its 472 GigaFlops of compute.
To just define a GigaFlop, for one second: that’s “one billion floating-point operations per second”. And the key phrase is ‘floating points’ here, in that with a GPU, as opposed to a CPU, we’re able to run calculations in parallel, simultaneously cross-referencing vast arrays of data points. So this is where the ‘man-on-the-moon-compute’ thing fails us as a meaningful descriptor of just how massive this shift in opportunity is.
(Just in case you were freaking out about a Terminator 2 / Skynet take over and enslaving the human race, a GigaFlop is 10^9 Flops, whereas the Human Brain is estimated at 10^18 Flops, so we’re safe for a couple more years - and “I for one welcome our new overlords!”).
Above: NVIDIA Jetson Nano, “C100” (252-0055) , selling for under £180. (RS Online - see full range). The last image shows a basic Raspberry Pi camera as an example of expansion / accessories.
Anyway, getting back to Machine Learning and your job, with my Terminator 2 / Simpson’s references aside, we are unlikely to see ‘robots’ effectively navigating creative, emotional, cultural, tasks. Although not a straight comparison, the human brain still has about a billion times more ‘compute’ than a Jetson Nano, but certainly, just as a ‘doorman’ has been replaced by ‘automatic doors’ in all shops that do not require elite symbols of grandeur, it is reasonable to assume that if something is reasonably well defined as a recurring problem, then Machine Learning probably has merit.
It may only take a few weeks to input say pictures of 1000 good apples, and 1000 bad apples - so the Machine Learning part of the process will probably operate at 95%+ accuracy, then it's hard to justify ‘saving’ a human job, that even if not Difficult, Dangerous or Dirty - is most certainly Dull, or at the very least not using the full ExaFlop capability of that human!
It’s not really my place to debate if someone is entitled to a job, even if dull, but otherwise safe (not 3-D), but as a technologist, looking at history, it’s hard to see a majority of evidence where we didn’t use ‘new tools’ to take us forward in some regard, at the short term detriment of some who are made redundant, but overall at to the benefit of many more. How else do we explain how life expectancy languished at a maximum of 40 years for millenia, and yet in the last 200 years, largely coincident with the Industrial Revolution, we are now almost double that. To quote Marty Neumeier, in his excellent book in education, Metakills (2012); “we are moving from an ‘Industrial Age to a Robot Age’”.
Coming to Our Senses: What to ‘Feed’ Your Machine Learning System?
The somewhat whimsical example of picking certain Sweets over others is to illustrate that you can train the Machine Learning system to detect pretty much *anything*, so long as you define basic parameters of ‘positive’ and ‘negative’ results. Whether it’s spotting LiPo Batteries, or spotting Purple Sweets (my favourite Quality Street sweet) - but it could just as easily be that you could use this to help modernise your work, be it detecting:
- ‘bad apples’ in farming [Visual / Cameras],
- ‘abnormal heat signatures’ in animals that may be unwell [InfraRed Monitoring],
- ‘eccentric loading’ of machinery, creating a ‘peak in vibrations’ [Audio Signatures],
- undesired ‘harmonic resonance’, in motors, [using Microphones] perhaps a precursor of failure,
- sudden ‘spikes in force’ if something is over/under-loaded [Force/Strain Gauges],
- dangerous emissions of prohibited toxic gases/fumes/smoke [VOC/Gas/Particulate sensors].
Above: A range of sensors, (optic, acoustic, force, chemical), all of which can in principle be inputted to a Machine Learning environment, so long as the ‘desired’ and ‘undesired’ bounds are easily defined.
Having worked in the Environmental Control department of Dyson from 2009-2013, as my first graduate job, it was not long before we realised the Sensor Array installed in Fans could certainly detect Volatile Organic Compounds (VOCs) such as smoke, hairspray, pollen, as said ‘pollutants’ passed through a fan’s internal sensors. And also when an engineer ‘cut the cheese’ in the lab!
Bodily-functions and toilet-humour aside, I say this to make the point that ‘Sensing Technology’ has been around for years, and yet it is rare that we apply a layer of ‘Machine Learning’ to make it more sophisticated. So long as one can say what is ‘acceptable’ or ‘not acceptable’ (good or bad apples) the system has a chance of being able to replicate to high confidence levels. Conversely, ‘subjective’ things like fashion styles, taste in music, or who we find attractive may have some ‘common data points’ or ‘trends’ perhaps, but intuitively we know these are considerably more complex, so it is hard to have a binary ‘yes/no’ outcome for a rudimentary Machine Learning system.
In essence, the skill is matching the tech to the problem, in such a way that is on the cutting edge, but no further, so it flounders and delivers useless results. Most of the Machine Learning applicable to most companies will be in the form of:
- Training Examples - Inputted, (data that is ‘good’ vs ‘bad’, so the system can ‘learn’ the difference)
- Machine Learning - Sorting said data into categories based on probability, of being ‘good’ or ‘bad’,
- Consequence: Something gets rejected, or selected for a special process, or ‘flagged’ for review, etc.
In summary, it's worth reflecting on how as humans, if we see a dog in the far-off distance, we may be 80% certain, using our eyes alone. However, if it barks, that confidence interval goes up to 99% let’s say.
AI is no different, and arguably the more ‘sensory inputs’ we add, the better the results. Unlike Humans we should not fixate on our limited ‘world view’ - as for example, to an AI Model, it is no more difficult to process IR or UV light data, as it is the visual spectrum we see. Likewise, why not examine Ultrasonics and Subsonic sound, why be limited to 20Hz to 20kHz, Elephants go below this, and Bats of course above. Gas sensors detect things our noses cannot, and so on. And we have not even got into the risks of the 4Ds that you wouldn’t want a human to do in the first place. Now is the perfect time to put those robots to work on projects no human should rightly be doing anyway. From this, your team has a grace period in the interim to hopefully upskill and retrain on other employment opportunities, before those are up for review. With that said, as I’ll explore in later sections, the outcome need not be ‘all or nothing’, but rather a Co-Pilot / Hybrid model or working.
In Machine Learning circles, the ‘Muffin or Chihuahua?’ The question relates to AI being ‘fooled’ by the apparent similarity between these two very different things. If you squint your eyes, one can be more sympathetic to this all-too-robotic error. And yet, I’d like to make the point of ‘why not add a microphone’ - if it barks, it’s 100% not a muffin. You could add an ‘electronic nose’ or gas sensor, but then shampoos these days are so fruity, who’s to say you’d be as confident.
OK, yes, in seriousness, you’d just use a couple of seconds of video and that would do it too. But the intellectual point is that a confidence interval can be radically improved not by just ‘more brute force’ but by adding another corresponding input. Consider likewise with your work.
What James Dyson’s Ex- Pool Cleaner Taught Me About Innovation.
If you’ve made it this far, perhaps the ‘take-away’ here is that innovation needs space to ‘play’, to ‘explore’, and ‘get familiar’. I’m not saying one ‘hack-a-thon’ will solve all AI woes, but if even straight-laced gurus of ‘business transformation’ such as McKinsey, through to rebel-scholars of ‘change management’ like the late, great, Clayton Christensen (Innovator’s Dilemma) advocate such things, it’s hard not to give it a try.
When I visited Lion Vision, I ‘trusted the process’, as we say in Design, that something good would ‘come out of it’. Sure enough, I proposed the ‘sorting sweets’ example, using Machine Learning. George and I worked on it for a couple of days, and it was in an evening when I got the idea to think of a new way to better detect batteries in eWaste processing - using high viz tape (or if not at least making them pink so they are uniquely incongruous in a sea of greys and silvers).
Above: General mess from eWaste processing (smashed up appliances and batteries) - next to a Hi-Viz Jacket - note this is in ambient light, all looks ‘greyish-silver’. Second image, with camera aperture closed down, and a small light shining on the subject - the Hi-Viz now ‘popping’ with high contrast!
The point is, the ideas come by doing, (not just thinking).
Asking a team to ‘think of a great idea for the company’ almost never works, as it’s inherently stifling and stressful - not creative! Yet if you say, ‘ok for fun, shall we get a Jetson Nano for a few hundred quid, and some booze and some peristaltic pumps, and make an AI cocktail robot, that makes you a drink based on you telling it 5 facts about yourself, and your facial expression - and we can exhibit it as the company xmas party to explain why AI is fun (and less scary).
In weeks to come, you may have a chance to softly explain how Large Language Models work - but in a relatable way, of ‘Personality vs Cocktails’ or whatever, it’s obvious, it’s silly, sure it may go viral on social media (like this similar example), but more importantly, your team will have to learn something new - and as every school teacher knows, kids (and adults), learn best when they are having fun.
Worry about the originality *after* you have had fun with the basics.
This, in case you're wondering, is why I wrote this article to not be “just another ‘how to use AI in your company’ blog”, I wrote it to explain why you need to re-frame it as being the start of upskilling your team somewhat through ‘stealth’ - making it fun, keeping it as simple as possible, and trusting that this also avoids the very real fears of ‘being replaced by robots’. Sure enough, not everyone will be employed by a given company in 3, 5 or even 10 years time, but I, like many of us, have been made redundant in the past - even when doing a terrific job with metrics to prove. There are just no guarantees in life.
Likewise, I’ve unfortunately had to make people redundant just because profits are down, not because they were ‘bad’ or ‘dead weight’ or not trying, just economics. It sucks. However, what has invariably been a good strategy for employee and employer alike is to ‘keep up’ with the new.
It’s why Dylan ‘went electric’, it’s why big brands got into Tik Tok, it’s why eWaste industrial companies in Manchester hire fresh-faced engineers to ‘try something different’. I genuinely worry about how AI will impact people’s lives and livelihoods, but I also believe that getting them involved by ‘lowering the bar’ is a good strategy. It may be that the person who works in IT might ‘do the code’, but the *insight* for the next big AI-idea for your company, might actually come from the person who dropped out of school, has worked on the production line, but understands the issues at a visceral and intuitive level. The point is that it’s about the team, not the lone individuals - and creating space for them to collaborate. I’ve seen this happen many times, and it’s one of the less discussed aspects of ‘diversity’ - that great ideas can come from anyone and anywhere, so long as we create space and foster it.
Whilst I was at Dyson, I was seated next to the guy who years ago used to clean James Dyson’s pool, but who worked his way up to be a Principal Engineer. To say he was unconventional was an understatement, but he was in my opinion an essential ‘x-factor’ amidst all the other more classically trained engineers (myself included!). With many patents to his name, and having been a leading force in creating two inventions that took Dyson beyond vacuum cleaners alone: The [coanda] Bladeless Fan and the Hairdryer. Over 3 decades, Kevin Simmonds has been a leading inventor, mentor and pioneer at the company, but he is also perhaps too iconoclastic to be a ‘typical’ engineer you read about in the usual fare of industry articles… (until now it seems - haha!).
Speaking personally, I can certainly say he was always a huge inspiration to me and he regularly solved problems which had stumped folks with more letters after their name. His ‘super power’ was that he simply built things and tried them. My point being, I’m not romanticising that having formal education is not an advantage, or that everyone on a production line is as ingenious or motivated as this Kevin was, but I genuinely have seen on numerous occasions in other companies - ideas come from unexpected people/places, and if we all can have the humility to have fun with it, it achieves great things.
I caught up with him, whilst writing this series, and I loved his take on this as inspirational as ever:
“I would say you have to make something, to know how it is made. You have to try something to see if it is right and what is most important, to listen to other people’s opinions - as there are many ways of doing the same thing. With AI you run the risk of ‘idea overload’ - as it can generate more than you handle. You still need a brain and some common sense to understand what is useful and what is fantasy.”
To my eyes, Machine Learning, GPT, Mid-Journey, and all the rest - it’s the same - they are tools, and it’s we as humans that give them meaning. We do it best when an unlikely group is thrown together, fuelled with fun and opinions of what needs to be better, and through being inspired to experiment and prototypes such ‘crazy’ ideas, until they’re not. It sounds like a ‘hack-a-thon’, yes, but the culture persists and lingers afterwards - and that's what matters - a sense of what could be, and that we can make it happen, again and again.
Creativity is a muscle, and it needs repeated working-out in opportunities like this.
We invent on our terms.
Industrial AI Blog Series Contents:
Part 1: Lion Vision, AI vs Automation, and Why a Game of ‘Go’ Changed Everything.
Part 2: “Dirty, Dangerous, Difficult & Dull” - The Case for Ethical AI Automation.
Part 3: Key ML Terminology: Are You 'Sorting Ingredients' or 'Baking Cakes'?
Part 4: ML Lessons from Lion Vision. AI Failures, and ‘Sensing Like A Robot’.
Part 5: Getting Started with Jetson Nano / Orin. And Why Octopus Brains ML Marvels.
Part 6: A *Shiny* Idea, Whilst at Lion Vision: “Hi Vis Batteries”. And Why You Need Underdog Engineers.
Are you looking for additional information on Lion Vision?
Comments