On June 15, 2017, we hosted an after-work event dedicated to "Artificial Intelligence – The Technology Of The Future”
We do realize that sometimes the terminology and key concepts around AI are hard to understand for those who lack a technical background in mathematics and computer science. Research rapidly advances, which makes these concepts increasingly complex and reliant on a scientific vocabulary that keeps expanding.
The aim of this article is to take a step back and provide the reader with a simple understanding of what makes today’s advances in Artificial Intelligence (AI) possible. Moreover, we provide additional explanations of technical concepts in simple words to make them further accessible.
With this, we also want to reiterate why our strong convictions, a key feature in all of our investments, are first a function of what’s happening in the real world, and second followed by what’s happening in the financial world (macroeconomics, interest rates, sentiment, volume, newsflow, liquidity, etc).
We are firmly convinced that the right mix between the tangible world (industrial) and the intangible world (finance) will yield the best results.
Strong convictions within a rightly but not overly diversified portfolio (over diversification kills diversification benefits) and the application of very simple financial valuation models (discounted cash flows and bull/bear analysis) provide investors with above-average financial returns in the long run.
We are also confident that in a world where passive investments, indexing and benchmarking is becoming the norm, theme-investing based on an industry angle and strong knowledge will capture most of the investment money at the expense of “in-the-box” strategies in the next few years.
The goal of our event was to show that:
Machine Learning is the general concept: training a machine to perform like a human
Machine Learning (ML) is a specific type of AI technology, enabling the machine to learn complex tasks which are non-linear such as cognitive functions (knowledge acquisition, decision-making etc.). The machine searches through large quantities of data to detect patterns that allow it to adjust its actions accordingly.
The first step in machine learning is to have the right algorithms and parameters in place. An algorithm is simply a mathematical formula describing a function, or more accurately, a set of rules that precisely defines a sequence of operations. Algorithms are essential to the way computers process data.
Many experts believe that in order to achieve Artificial General Intelligence (AGI) or, simply stated a machine/robot more intelligent than a human, more of those algorithms are needed and are still to be invented. Such breakthroughs are to come out from the human brain in our view, even though 2nd generation software programs (that do not require any in-advance training such as the 1st generation innovation programs) are solving concrete problems.
We can summarize this by saying that with more and more knowledge available (no matter if it comes out from a human or a machine), some exceptional and very skilled humans are for sure to conceive new formulas (and algorithms), which will make further advances in AI possible.
After all, all the great geniuses of the past and the present are like everybody else, they share the same information, which is today easily accessible through the internet.
What makes them different from all the others is their deep knowledge, curiosity, interest in solving hard problems, intuition (overturning common sense and ideas which are thought to be correct), focus, idea sharing (others are needed to prove themselves right) and non-conformism (never satisfied until their own satisfaction is reached).
As far as we know, no machine nowadays has all of the above-mentioned qualities but there are humans with all of the listed features and, as such, major advances in AI are to be expected rather from humans than from machines.
When we talk about a machine that learns, there are several methods of learning and the most successful and widely used for the moment is called supervised learning.
With supervised learning, the model is given a lot of input data it can learn from and produces a certain output. The output is then compared to the targeted output in order for the data scientist to optimize the model efficiently in case the model makes a mistake.
With supervised learning, the second important step in machine learning - labeled data - is introduced.
The key is to have as much data available as possible. This data needs to be labeled in order for machines to understand it. Labeling means that the data (could be a number, picture, pixel, etc.) is tagged.
A good example is the picture of a cat or a dog. We humans know whether it is a cat or a dog. But for a machine, to understand whether it is a cat or a dog requires the picture but also the description (the label or the tag) of what the picture is about. With such labeled data, a machine model can be trained.
Later on, we will talk about other methods of learning that are less common, but that do not rely on labeled data.
Computing power is core in Machine Learning advances
The 3rd step in machine learning is to have massive computing power. Time and again, we wrote in our research reports that Graphic Processing Units (GPUs), or more simply stated the graphic cards you can find in your laptop or PC, do play a huge role in machine training.
Much of today’s advances in AI are made possible thanks to high-performance computing. Without such advances on the computing side, the pace of innovation in AI would be much slower, not to say impossible.
GPUs are essential in machine training. They perform multiple tasks simultaneously, referred to as parallel computation. On the other side, Central Processing Units (CPUs), more commonly known as processors, are simply the Intel chips (and to a lesser extent AMD) everyone has in his laptop or PC and are best suited for sequential computing.
CPUs handle all the instructions (both the hardware and the software) they receive from the computer. More simply stated, a CPU performs mathematical, logical, and decision operations that direct all of the processors operations.
It is quite understandable that performing multiple tasks at the same time, vs. one after the other, yields much faster results and reduces costs. As machine training requires billions math operations and connections in parallel, GPUs are a must-have. During the last three years, Nvidia GPUs sped up by a factor of 50 times the training of deep neural networks and the speed is expected to increase by 10 times in the coming years.
This is the reason why since the inception of our Artificial Intelligence & Robotics certificate in October 2015, Nvidia and AMD have been strong convictions in the portfolio. We believe that valuations still leave room for significant upside as just every industry and every company seeks to become more intelligent in order to improve its efficiency and create new products and services.
There are many advances currently taking place in the hardware/software industry, which are likely to further increase the speed and power of computing going forward and we are currently exposed to these trends in the different portfolios we manage.
Deep learning, reinforcement learning, unsupervised learning: don’t get lost in translation
In the data science space, advances in different types of learning are helping such machine learning models to become more powerful. We commented about the supervised learning method above and will now introduce a class of machine learning models and other learning methods that we see as very important.
“Deep Neural Networks” or simply “Deep Learning” models are among the most powerful machine learning models, which have successfully rivaled human performance in domains such as conversational speech recognition, object recognition and game playing.
Deep Learning models allow a machine to learn a hierarchical feature representation of an object and imply numerous data processing layers (for example, it will search for four legs and then for a tail for a dog), leading to a more accurate outcome. Moving hierarchically from inferior to superior levels, the model interconnects those layers and completes the output. That is why these models are compared to neural networks, as the process is as complex and similar as the neural networks interactions.
Despite major advances, the possibilities offered by Deep Learning are still limited. These methods, which rely heavily on the presence of labeled data in order to be trained, learn very specific functions and are unable to transfer learned knowledge across domains. Therefore, one is required to train the model for each domain separately, which can only be done if enough labeled data is available.
In addition, Deep neural networks are “black-box models”, meaning that it is difficult to evaluate what the model has learned, which in turn makes difficult to obtain guarantees on performance. This may hinder application in domains where decisions are critical (medical, self-driving cars etc.).
Large-scale research in deep learning models will mitigate these practical disadvantages in the near future (labeled data notably). For example, one can split the function to be learned into several parts to increase robustness, as opposed to learning the entire function from input to output (end-to-end).
Another example of how the dependence on labeled data can be mitigated is by considering different learning paradigms.
Reinforcement Learning is one. In reinforcement learning, a model observes an environment, interacts with it by means of actions and gets some feedback in the form of rewards (the target being to maximize this reward).
A chess game is a perfect illustration: the model receives information about whether a game played was won or lost. While the model only has the result of the whole game and does not have every move in the game labeled as successful or not, it will play a lot of games and will give bigger “weights” to those moves that resulted in a game win.
Unsupervised Learning is another one. It’s a hot topic of today and this method of learning will allow achievements in domains where limited labeled data is available. In unsupervised learning, the model has access to a large amount of data and characteristics of each piece of data and will be able for instance to separate pictures of dogs and cats in two groups based on some features of the pictures.
A promising outlook, the only challenge will be to align the goals of AI with ours
We are convinced that Artificial Intelligence (and intelligent robots) is just getting started and that the trend will last more than anyone can imagine today.
While AI already has a broad practical impact in many industries (from finance to health care or robotics), it is expected to become much more powerful in the future thanks to these new ML models which are less dependent on labeled data, with Artificial General Intelligence or Superintelligence being the ultimate goal.
Whether we’re just a couple of decades or centuries away from Superintelligence remains an open question as most AI experts disagree on the timeline.
As you have probably noticed, recent advances in AI have raised concerns about AI safety and its impact on the job market. Jobs that are well-defined or have an exact outcome and for which a lot of data is available to be learned from are the most at risk: radiologists, taxi-drivers, logistics managers…
So, the challenge will be to align the goals of AI with ours before it becomes super intelligent. It implies developing methods to evaluate what an AI has learned, being able to test its goals and learning how to raise an AI that shares the same objectives with humanity.