What are the main drivers in machine learning? How do you tell promising technologies from those born to be forgotten soon?
My recent sci-fi readings and their (failed) forecastings
When I was reading the Hyperion Cantos by Dan Simmons, I became convinced that … Steve Jobs had read it too! The thing that Simmons called commlog obviously resembles an iPhone or iPad.
Yet throughout the four books of this brilliant sci-fi cycle, you notice technologies here and there that probably were a hype somehow in the past, but nowadays seem rather clumsy and of a little use, like a levitation barge.
As I moved to a later sci-fi generation, it started to look like a sequel to what used to be hot on TV two decades ago, when I still was a youngster. For instance, 2003 everybody mourned the death of the first ever living clone, a sheep called Dolly, and 2008 the novel House of Suns by Alastair Reynolds continued the cloning topic, and predicted it a big future of leading the space colonization.
How often do we hear about this domain now?
What both sci-fi readings have in common, is their certainty about clever machines and their strong link with the humanity. A typical situation from a sci-fi book would be a hero calling his or her space ship to come, talking to it, and even building a strong emotional connection to it.
The present world is the past of the sci-fi future
Recent developments have their own logics and not always follow what sci-fi literature forecasts. I checked, how the most relevant — in my very subjective opinion! — fields are doing and what keeps them going.
My check list included technologies that have to do with machines being able to move and decide independently, as well as to communicate and understand human communication. This what a machine needs to be able to do to fit into the picture that most of the science fiction books draw for us.
I am not saying these technologies will remain leading in the next two hundred years, since some of them are too primitive in comparison with what a future machine can. But I do believe that a couple of these (mainly machine learning) technologies play a role of a starting point in a long journey to the space ships behaving like pets or even humans.
Moving around requires for a single machine an ability to classify visual images and a coordinating infrastructure that would prevent a traffic mess. Currently, the Internet of Thing are computer vision seem to be the responsible subfields.
Decline of IoT and CV?
It would not be an exaggeration to say that internet of things (IoT) or computer vision (CV) are still kept in a cradle. The first one remains dependent on the cellular networks. The biggest players in this industry do not show a lot of interest for messing up with IoT companies. The IoT startups — most of such companies are rather small ones — spend a lot of their time analyzing connectivity issues rather then working on something substantional.
The computer vision seemed very promising just a few years ago. Picture and defect recognition, then picture prettifying quickly reached a sufficient perfection level. They gave up being a source of inspiration for computer visionists.
Car manufacturers do not need a more profound technology: the existing autonomous cars only need to be able to read road signs. Due to a legislation gap, they are not allowed to make decisions imposing potential risks for human beings. This makes autonomous cars too much an uncertain domain to invest into.
The same industry used to finance a lot of IoT projects. Nowadays, rental cars can even send notifications if the tank is getting empty, or some other problem is detected. But that would be all for now.
Energy and defense may turn their faces towards IoT, but their hands are bound due to the network instability. As soon as the next generation of mobile networks will be there, IoT may well make another advance.
What is selling best?
While the high-end technologies seem to slide into a recession, the low-end ones look more vivid. The whole new approach has emerged in this regard: it is called tinyML.
In fact, machine learning and robotics never managed to consort. However, less clever and cheaper devices with embedded systems, for instance, surveillance cameras and speech decoders, started to get equipped with primitive AI algorithms.
This is where we come to the first success factor. Easy to guess, it’s a commercial usage. It should have a potential, a market niche that is still hungry. Now, not in the distance future.
But what about NLP? What’s its main commercial purpose apart from text generation?
Indeed, NLP has got a bad reputation as a content supplier for online shops. We should not forget that e-commerce industry is huge and nobody dares to predict the end to its expansion. Imagine, that NLP would replace all content mills!
Online content writing did evolve into a solid economic domain. Writing itself is seldom well-paid. Most of the costs were caused by the hours that content managers spent with coordinating the work of single writers.
Last but least, the recent pandemic may contribute to a new upturn in the number of NLP research projects. Medical documentation as it tends to be well-structured and often standardized, is a perfect candidate for becoming another application area for NLP. Automated translations may save a lot of lives and help pharmaceutical corporations make some more money by shipping their medical products to a big variety of countries on short notice.
Making or saving money
This time I am not going to say: “It’s not all about money!” Because industries that are able to pay make machine learning move forward.
For instance, a new ML subfield, reinforcement learning, was partially inspired by autonomous cars. Depending on the type of a vehicle and on the traffic or other situation, the car must be able to make a several decisions and prioritize them properly. A truck or a cabriolet will behave differently due to its nature.
Medical companies come in play again. Reinforcement learning is being used to train artificial limbs to adapt to the walking style of the owner and make his experience more comfortable.
Another catalyzer in the recent ML developments does not refer to making money, but, on the contrary, to the lack of money. Algorithms grow their wisdoms if they are fed with new data. But where and how to get new datasets without spending billions of dollars?
Let the algorithm extract the meta-data from the learning process. It can now not only learn some applied stuff, but it learns how to learn in general. This new subfield is also known as AutoML.
All ML subfields are fascinating and attract most talented programmers and mathematicians. Still, the economic factor often decides not only about resources, but also provides a direction for future work, an idea, where to move next. Subfields that do not manage to find a practical implication will stay in the shadow of NLP, IoT and the others.