I thought it time I returned to some of my other areas of interest prior to discovering cryptocurrencies. And I thought I’d kick it off by doing a round up of what the current state of machine and AI research is. Before diving in, let’s start with setting the background.
Machine learning is an area of computing in which outputs or actions are derived from a set of inputs through a learning function. At its heart it is just based on mathematical formulas, just like in statistics. So take some inputs, apply some mathematical transformations, to get to an output. Statistics relies on a combination of practioner knowledge and established statistical methods, to formulate the right equation. This can be what inputs to use and how to transform them (aka data preprocessing and feature engineering), the mathematical model to use (e.g. the linear regression formula), and the statistical method for deriving the co-efficients for the model.
Machine learning is essentially in the same field, but where we try to reduce the amount of human involvement and get the computer to find the best inputs, formula, and coefficients by itself. This is done through a learning function. That is to say, it is fed lots of historical input and output data, and it attempts to mould the mathematical model to fit the data such that providing a set of inputs tends to give you the correct output.
This method of learning the coefficients from historical data and using it in conjunction with classical mathematical models such as linear regression is the most popular form of machine learning and often used via the scikit-learn library in Python.
A more advanced mathematical model is the artificial neural network. This was invented in the 1950’s and is meant to be a (hugely) simplistic representation of how the brain is thought to work.
It was followed by decades of research but was met with limited success. With AI summers where some breakthrough is made, followed by AI winters when the mainstream computer science community shunned artificial neural networks.
The reason why machine learning has seen a resurgence in recent years is down to the persistence of Prof. Geoffrey Hinton and one of his students who had the idea to try to use GPUs instead of CPUs to Hinton’s multi-layered neural network idea. GPUs are normally used for gaming, and do matrix multiplication really well, which is the heart of how machine learning algorithms work.
As a result of the faster technology, decades worth of AI research that previously struggled to bear fruit suddenly started to work well. This was the dawn of Deep Learning and it began only in 2017.
Since then there has been a new AI summer and significant investment in machine learning research from the worlds top technology companies from IBM to Facebook to Google to Microsoft.
Still the holy grail of AI research remained – that of Artificial Generalized Intelligence or AGI for short. There are two fundamental problems with current artificial neural network methods. One is the vast amount of data required in order for the model to learn (and even then it can be flawed) and the other is adaptability to problems that are similar but not the same as the original dataset. Humans on the other hand learn in a much more efficient manner and can apply reasoning gained from one area to another.
We are still many years away from achieving AGI (a lot of pundits feel it will be at least 2050 before we do so), but deep learning has nevertheless progressed a lot in the last decade. In part II of this post I will delve further into deep learning and provide a rundown of cutting edge methods in the pursuit of AGI.