For me, AI is a very interesting topic - I studied it at university, then spent the first decade of my career realising that AI was treated as a joke, the old cliches about automated phone menus where you have to speak words, and then the computer does a terrible job of recognising those words etc, had you screaming for a real person.
Yet, here we are, and seemingly out of the blue AI has now become a buzz word. Everyone is scrambling now to put the word AI into everything - if you haven't go AI in whatever you do, you're behind the curve.
Reading through this thread highlights the problem. There's a lot of 'unknowns' and we are at serious risk of silly legislation being introduced and so on. And what does all this mean for investors as well?
So the first thing is being realistic. Why is AI suddenly such a buzz word, after decades upon decades of hollow promises?History
Pre-noughties, AI had a problem. There were lots of interesting ideas, and AI could do some tasks very well. Back in the 20th century AI had already beaten chess grandmasters.
The problem with that
AI was that it was just number crunching. Follow simple rules, and throw enough computing power, and in a constrained world like chess - composed of very straightforward rules of the game - crunching numbers allowed computers to outperform the best chess players in the world.
The problem was, when the same approach was tried with things like computer vision, it failed miserably. Computer vision seemed just so hard.
Back in the 1990's neural networks ('connectionist AI') were struggling to take off. Perceptrons (a popular type of neural network) were taught from the theory that anything that could be done in 4 or more layers could be theoretically shown to be do-able in 3 layers of neurons. Further more, each neuron in the input layer tended to connect to all neurons in the subsequent layers ('fully connected networks').
When I studied it, it felt like if you submitted work with more than 4 layers, your tutor would be wondering whether you'd understood the course material, with all the meticulous proofs that it could all be done with 3 layers, and so on.The Quiet Revolution - Convolutional Neural Networks
The current fashion for AI is almost solely based on these. These are also more coloquially termed 'deep learning'.
Strictly speaking, these aren't completely new. In fact, people have toyed around with them since the 1950s!
But a number of things and realisations have recently come together.
(1) The 'convolutional' aspect is important - instead of a 'fully connected' network with neurons looking, say, at the top left of a picture being trained independently of neurons at the bottom right of the picture, a convolution approach takes the view that we don't know where our 'cat' (or whatever) might be in an image, so the same 'neuron weighting' is applied - 'convolved' across the entire image, at least in the first few layers of the network. In the first layers of the network, instead of many different sets of neuron weightings, a single set of weights is repeatedly applied (convolved) across the entire layer (start at top left, apply it, shift it one to the right, apply it, shift it one to the right, apply it ... and so on).
(2) Depth Matters. AI programmers have finally broken free of the rigid idea that everything can - or rather 'should' - be reduced to 3 layers. Going against what I was taught in the 90s (no matter how theoreticaly 'provable' it might be), the realisation today is that in practice, 10 or even 20 (or more) layers of neurons can potentially arrive at a solution far quicker, and better, than a 'theoretical' 3 layer solution. Theorists haven't yet determined the science behind 'why' this is the case, nor how to find the 'best' architecture for a particular problem. But real world experiments have left no doubt! The theory can come later.
(3) ImageNet. This is the game changer. ImageNet is a massive database of images that have been meticulously labelled, and can be used for testing machine vision. There have been yearly challenges for quite a while now. Initially no-one seriously used neural networks. Initially they used more 'traditional' techniques where programmers carefully programmed their software to look for pre-determined patterns, etc.
For a couple of years, the resultant programs performed noticeably worse than a human. And any improvements from the previous years, were small tiny incremental improvements. Certainly, reliable computer vision still seemed an age away.
But then,...https://qz.com/1034972/the-data-that-ch ... the-world/
Two years after the first ImageNet competition, in 2012, something even bigger happened. Indeed, if the artificial intelligence boom we see today could be attributed to a single event, it would be the announcement of the 2012 ImageNet challenge results.
Geoffrey Hinton, Ilya Sutskever, and Alex Krizhevsky from the University of Toronto submitted a deep convolutional neural network architecture called AlexNet—still used in research to this day—which beat the field by a whopping 10.8 percentage point margin, which was 41% better than the next best.
Suddenly everyone took notice - this was a revolutionary change in computer vision performance on general image recognition tasks. Suddenly computers were now in the ballpark of matching humans.
Since then, the majority of entries to the imagenet challenge have all switched to convolutional neural nets of some variant or other, and their capability is now actually - just - able to surpass that of human beings.Where we are now...
All of a sudden, using computer vision to identify pedestrians, cyclists, cars, lorries, buses, street signs (including reading them!), traffic lights, dogs, cats, etc..... all these are now very possible with around the same level of reliability as a human being.
This was the pivotal moment that meant fully autonomous cars went from being something from distant science fiction, to becoming a very, very real probability
It is now just a race to be the first to - safely - get autonomous cars to the market.
Crucially, from an investors point of view, without any major catastrophe that completely undermines public perception of the tehnology. No mean feat, when the end goal isn't perfection
. They just need to be better than the average human driver.
But since the average human driver makes occasional mistakes, autonomous cars can still be massively beneficial even if they have the occasional bump. Nobody is, or can, claim they will be perfect.
The problem for the techonology - and investors - is that any occasional bump will be jumped on by the world's media like a pack of hyenas.
Google (Waymo) seem to recognise this, and - in my view - are taking a very prudent, careful approach. They realise one big cock-up could set back their attempts quite literally years, and encomber the industry with draconian, restrictive legislation, that could potentially prevent the technology ever being allowed to reach its potential.
Uber on the other hand - in my view - seem to be taking quite a cavalier approach, and rushing cars onto the streets without first gaining public approval - not just from the authorities, but also the approval of general public opinion. And by doing so they pose not just a real risk to their own attempts, but to the more sensible attempts of others as well. If they have one bad accident, it could seriously turn public opinion against driverless technology and AI in general.But Get Real...
been any massive jump towards sentient AI.
That is still yonks away.
There hasn't even been any massive jump towards any form of general intelligence.
That is still a while away - though there are some small scale, rough attempts. But like the neural nets of the 1990s, todays general AI are still at the noddy stage. They are still waiting for their revolution, and there are no signs that it is at all imminent. Contrary to the impression the current AI buzz might give.
Nothing in the current AI revolution is putting us at risk of computers becoming sentient, and 'breaking their programming' and taking over the world.
There's no threat from the current state of the art technology - at least not from the technology itself turning against us.
As always, there may be threats from how humans
might apply the technology, but be under no illusion - any bad effects from the current AI revolution will be entirely due to the will and labour of the humans behind it - just like any technology.Quite Simply..
The current revolution is based on a revolutionary change in one particular aspect - convolutional neural networks. And in particular towards visual processing tasks.
These superbly fill in a number of 'black boxes' in engineering terms. They provide a functionality as one part of a modular system, that was not available before. Suddenly we have a few more boxes that provide limited - but incredibly useful - functionality that we didn't have before.
You can now have your £40 digital camera identify your face, and recognise when you are smiling. Even 15yrs ago, that was still complete science fiction with the state of the art back then.
It isn't an understatement to call convolutional neural networks a revoution.
But don't worry - they haven't turned your compact camera into a sentient being working out ways to kill you off.
All it has done, is provided the engineers who made your camera with a module (may be implemented in software, or may be implemented in a dedicated chip) that they can incorporate into the camera that can take the input from the camera pixels, and output where it thinks faces are in that picture, and whether each of those faces are smiling.
It is simply then up to the engineers to decide how they incorporate that into a product. They can simply use the output from that module in their main camera program, programmed using regular techniques, to decide when to trigger the shutter.
Similary, that is how self driving cars will work, albeit there will be much more 'engineering' work to build a reliable system which can also be programmed to take into account changes in the rules of the road, car handling, etc. But that is more straightfoward engineering / programming, than simply builds on top of the image reconition AI.
That's not to say that other AI techniques won't be brought into play. For example, for planning paths, etc. But these aren't quite so revolutionary, more evolutionary (and I don't mean in the terms evolutional algorithms - there's been no revolution there). So where does that leave us as investors
Well, the obvious application of this AI revolution is autonomous cars - these were clear science fiction before. Now all the parts are in place. The race is now on to make it happen.
Waymo would be my obvious candidate, but they aren't making the cars themselves.
And it seems that all the car manufacturers recognise that self driving is now a matter of when rather than if, and they are all now ploughing billions into it.
I can see two probable outomes...
1. Waymo becomes the standard - they develop the technology and licence it to the rest. They are certainly going about autonomy the right way in my view. They fully recognise the potential risks - both in terms of public opinion, and also the risk to life of users of their technology - and seem to be following a very pragmatic approach.
2. Each car manufacturer manages to develop their own - which they all seem to be trying to do - and actually, autonomous car technology from an investor point of view, would then just be an investment in the car companies themselves .... Ford, Toyota, and so on. It's just another aspect of technology that goes into cars.
I can see investment risks with each....
Waymo seem to be doing brilliantly with object recognition, place finding in the real world (GPS is useless for local lane navigation), and path planning around what other road users are doing. Where I think they may struggle is in car handling - the actual driving. Waymo are looking for a technology they can sell to other manufacturers. But half of driving is about how the car handles. How quick to you turn the steering wheel, what's its turning circle, at what point will it lose grip on a wet, snowy road.
Other car manufacturers, have been developing traction control, assisted steering, ABS braking systems for a while now.
The question now is which is going to be the bigger challenge. Google seems to have got the high level pedestrian recognition and object avoidance reasonably well sorted already.
My gut feeling is that the remaning challenge, particularly for cars in the UK and other colder wetter regions - away from California - is going to be the automated *driving* aspect - the actual 'hands on' control of the vehicle. And that might actually be to the advantage of the existing main car manufacturers, who already have a lot of experience developing (safety) technology related to those aspects.And the Hype
Technology is always improving. Other AI techniques are making incremental improvements.
But if it weren't for the convolutional neural nets, I don't believe we would be seeing this current AI euphoria.
In other words, I do believe that the current convolutional neural networks will have a transformation change.
But it will be limited, in the sense that there won't one or two companies holding all the patents.
In fact, most of the actual AI stuff, is in the public domain and free. The patents and protections are coming from associated technologies (like the Lidar systems in waymo cars, etc), not the AI itself.
We currently have Microsoft, IBM, and Google, etc, all trying to sell AI services.
The intersting question is how much of a monopoly will they have - how much will people need to use AI services, compared to how much the AI will instead be incorporated, e.g. into integrated circuit boards that can be embedded into other electronics.
The 'value' from the AI services is not from the "AI" technology itself. It isn't from the convolutional neural network architecture.
The value comes from 'training' of it. The value is from havin a trained network that recognise objects.
But google, etc, can only really make that general purpose. Specialists might want a dedicated convolutional neural network trained to recognise e.g. brain tumours. But then it is going to be the specialist that is going to need to train that network. A general network brilliant at recognising cats, or models of car, isn't going to be much use identifying a brain tumor in a scan.
And this might be the achilles heel of the AI technology from an investors point of view.
Yes it's likely to be revolutionary. But that revolution is likely to be from a broad use across all industries. I'm not sure that there is going to end up a single commercial entity that owns and controls access to a single massively intelligent AI to which everyone will necessarily need to connect to.
I supposed after all that, as an investor it gives me hope for the general future that there is a lot of scope for companies to massively innovate and provide new features, technology and do things in a far more efficient and effective way.
So there is potentially a lot of scope for general economic growth.
But I'm not sure that there are any clear indiviaul winners in terms of being the controlling owner acting as a gatekeep to such technology.
From a worker's perspective, an AI future isn't something that should be feared. AI will be just another tool we all use. It will be valuable, but fragmented. It benefits being realised in many areas but through the work of many people crucially being aided - not replaced - by it.
The fears of it taking over and wiping out mankind, are massively overblown. That really is simply fear arising out of ignorance.
In terms of society, it is potentially going to be transformative. Easily trainable image recognition has potentially enormous numbers of beneficial applications. Medical image diagnosis, face recognition, automated monitoring - defect monitoring in factories, etc. More complex OCR - recognising not just text, but potentially diagrams, and drawings as well.
The same technology behind it has been adapted to other image processing functions as well.
There are various examples of work involving depth estimation from a single, monocular image. I can't find it now, but a while ago I saw a video of a remote controlled toy car self driving around a campus, avoiding obstacles, solely using a single 2d video camera, with a convolutional neural net estimating the distance of obstacles on a single frame-by-single-frame basis, etc.
Similarly there are other examples of 3d scene reconstruction from a single 2d image. And other examples of 3d models of faces being generated from a single 2d image ... you can even try it yourself! http://cvl-demos.cs.nott.ac.uk/vrn/
Adobe is using image recognition technology to isolate entities in videos, removing the need for laboriously having to do this by hand.
This will open up a whole new world in movie visual effects - or at least, make a previously niche, expensive world, available to even the lowest of budget film producers. Easy to add people in, take them out, change their clothing, etc.
Once you can have the computer easily and automatically recognise independent objects, a whole plethora of opportunity opens up in graphics and video packages. And consequently whole avenues open up to advertisers, graphic designers, etc.
There is even work that attempts to use machine learning (convolutional neural networks) to generate predicted motion from a 2d image... https://www.theverge.com/2016/9/12/1288 ... iction-mit
The scope for economic growth out of this is huge.But is it all what it seems?
So the technology is accessible to all.
However, there may be one saving grace for investors.
Convolutional Neural Networks are relatively fast and cheap computationally when you have them already trained up. That's why your compact camera can find your face without needing a supercomputer.
But training the network in the first place is a whole different ball game. And this may be of use to investors.
Although the algorithms are freely available, if you actually want to create your own networks, you potentially need a lot of computing power to train them up. Programming the underlying network itself is the easy bit - training the network is the hard bit.
And this is basically where the big IT companies are pitching it. They may superficially sell "AI" services. But the reality is, the AI is actually relatively free and easy to implement, and for the most part isn't protected by patents, etc.
The reality is, the services are just a front end for their cloud computing. What you are actually paying for is the CPU time.
For example when google boasted about AlphaZero beating the worlds best with only 4hrs training from nothing (https://www.theguardian.com/technology/ ... four-hours
), it wasn't really the AI algorithm they were showing off.
They were really showing off the immense brute processing capability of their cloud computing platform.
Cloud computing predates the current AI revolution.
Cloud computing almost felt like a solution looking for a problem.
A cynical person might wonder if the current buzz around AI might potentially just be a way of making you think you have a 'problem' for which you might need a processor hungry, cloud computing solution.
Sorry. I hope that isn't too long. It just seemed a good point to try and collect my own thoughts having been keeping a close eye on the recent AI revolution quite closely myself from a technical perspective.