Machine Learning - the current state
By Loraine Nijhuis & Vera Engelbertink
The biggest buzz in data and technology: Machine learning (ML). What is the current state of this development, and what are its implications for the world as we know it today?
Concluding of our visit to the Papi’s (the real world stories ML event) in London in the first week of April, ML applications has grown. Last year we mainly heard the brilliant ideas of ML scientists, this year we are talking proof of concepts of applied ML techniques. With, sometimes overwhelming, pioneering implementations.
But first, let’s start at the beginning, what is machine learning?
Basically, ML is statistical computer model that is capable to interpret data and learn from it. The model is taught to recognize patterns. In the early days of ML this mainly concerned the ability to recognize very basic images or text.
A ML model is a network of neurons, similar to the human brain. The human brain also consists out of a lot of neurons (neural pathways). The brain is able to function through the connections between these neurons and it will give output as emotions, recognition, etc.
This is not so different for a machine learning model. The model could be described as complicated decision tree based on neurons that can be linked together in many ways. Depending on the input its given, the model delivers an output. For ML it becomes more complicated when it is supposed to recognize sentiment. This is a complex matter that is too difficult to teach in our present time.
Enabling a model to learn, we need extensive computer power. You can imagine that the input is hundreds of thousands of rows of data, it will take a considerable amount of time before the model is trained enough to generate output that actually makes sense. ML models use GPU (Graphics Processing Unit). Exactly, the same GPU as used in game consoles. The reason GPU is used, is because it has much more computing power than CPU. However, due to inventions as the BlockChain which also require massive amounts of computing power, there is a strong competition between the data and gaming industry for GPUs. To the annoyance of parties like Sony who is responsible for the Playstation.
Until we are able to crack the computing power problem (read: produce more GPUs for an affordable price), training large and complex ML models will continue to cost a lot of time ánd money.
Where are we, truly, with ML solutions?
At conferences like the Papi’s the machine learning and technical society showcase their solutions and groundbreaking possibilities and results. There are many revolutionary initiatives at different domains. Like we mentioned at the start of this article, this year is “proof of concepts” year. Models that have been deployed successfully in a startup-like setting and are contributing to solutions for our societies’ main problems.
Some great examples are Antiverse and GTN, both startups using ML to improve healthcare.
Antiverse created a models that is contributing to the prediction of antibody-antigen bindings and antibody drugs. The computational model is able to do this within a day, thereby dramatically decreasing timings where the average lies between 3 and 18 months.
GTN combines advance generative ML models with quantum physics to speed up the process of drug discover. Continuously looking for new chemical compositions the model is able to find the right medication for a disease much faster than any human being.
These two examples show how much our society can benefit from ML tackling major issues in the healthcare domain.
Machine learning in the creative domain
At first glance combining ML and creativity seems a bit off. Machines can not be creative, can they? Well, machines are very well capable of showing creative behavior to support the human creative process. A really nice example of this is Folk-RNN, a model that was fed tens of thousands of Irish Folklore melodies. Training the model with these melodies enabled the system to actually compose melodies on its own. Hence the melodie, which is something completely different than producing music which is a human proces.
A lot of designers use ML (or AI) in their daily job. Whether they know it or not. All Adobe programs, i.e. InDesign or PhotoShop, are equipped with some form of machine learning application. Like the “liquify” tool in Photoshop. So there is much more AI in our daily live then we actually realise.
BigML, one of the biggest and best known ML platforms available, is a really good example of how to downscale ML’s complexity for the rest of the world to understand. BigML was able to predict the Oscar winners in 2018. They got 6 out of 6 right. Which of course is an awesome job! One could discuss the meaning of this for the technical domain, but the fact of the matter is, they contributed a lot to the public discussion.
The ethics of it all
Although we have seen so many proof of concepts that work, we face problems with computing power, a sceptic society and the ethical questions machine learning, artificial intelligence and robotics raise. We have stressed it many times before, but in our opinion the technical domain is responsible for the choices that are made whilst developing, training and deploying ML. Responsibilities such as transparency, the use of unbiased data, sharing knowledge, explaining to the public are with us. We have the obligation to protect humanity against wrong models or solutions that do not benefit our society.
Is it responsible to teach machines what humans are capable of by nature? If you train a model with open source data, data that is free and available online, who is the owner of the output that is generated? How does the model handle biased data, data that basically is insufficient or where the expectation of the output is out of line with the actual output. This could be devastating for models that contribute to i.e. healthcare.
Imagine a model that predicts the cure for certain diseases but was trained with biased data, missing data for a specific blood type. The output could suggest a cure for this group that is not a match at all, maybe even aggravating the disease instead of curing it.
All ethically raised questions should be clear the to technical society. Especially for delicate models, the gap for error should be 0.
Machine Learning in the future
What do all these ML related accomplishments and developments mean to us, the life that we live, now ánd in the future?
In time, ML will replace the repetitive tasks of our daily activities. In our professional or our private lives. Imagine, i.e. a purchasing process or the delivery of relevant content to online users. Ordering groceries, done by your fridge, automatically based on your purchase behaviour and what you consume over time.
Does this also mean that human employment will disappear, or that jobs will be lost? Honestly, it remains to be seen, but the focus of human action will definitely change. The repetitive tasks in our professional lives will be done by machines, sooner rather than later. But this does not necessarily mean jobs will disappear, it means professionals can spend time on tasks and new things and focus on what is really important in a job.
Current and new generations will grow up in a world where technological developments are part of everyday life. Self-driving cars are on the road, machines recognize objects, images, animals, discover tumors and predict cures. Algorithmes know better when you need to buy something then you do.
Digitally the world is at a huge turning point. For digital services it means we’ll perform based on facts, data, models, etcetera, instead of our gut feeling. Machine learning helps to predict what you buy or where you want to go on your next family trip.
Also from a security perspective ML enables the prediction of storms and disasters, coordinate help and maybe even prevent a disaster from happening in the first place.
Maybe we are not ready yet, but there is no way back. ML is a fact; it is up to us to lead it where it is supposed to go and take fitting responsibilities to this innovations that eventually will benefit us all!
