Dean Anthony Gratton
Artificial Intelligence: Deep Learning
In last month’s column, I tackled the subject of machine learning (ML) in Artificial intelligence: machine learning and I suppose one obvious question would be “What’s the difference between machine and deep learning?” Well, both terms are subsets of artificial intelligence (AI), although deep learning is a subset of machine learning. But more about this later …
Developing our empathic agent
In the meantime, let’s touch upon machine learning so as to re-establish what we have already understood and, that is, ML is a broad field of study, where the development of software (a computer program) automatically improves through experience, based on the data that it has received. We nowadays associate the term with “big data,” “data modelling” or “data science,” where retailers, for example, like to collect information about your shopping habits so, in turn, they can more accurately target their advertising. And then there are the likes of Amazon and Netflix, who use data analytics to predict or suggest what you might be interested in viewing or purchasing next.
What’s more, it seems the head-scratching gray matter folk – you know people such as academics, analysts, researchers and the like, have yet to agree upon a definitive statement as to what machine learning actually is. Anyway, for me, this is where I take my soapbox and stand proudly to proclaim that this whole AI-malarkey is nothing more than clever programming and smart technology. You see, when you have a software program that has an adaptive set of processes, combined with experience captured through the data it has received, whether that’s previously or currently, then perhaps we can liken it to an “empathic” agent.
When is human intervention needed?
I mentioned last month that such software has an ability to “see” our world through data and, as such, from the software’s perspective, it can manifest a representation and an understanding of “its” world – something our software creates through data modelling, in turn, leveraging its experience with data. More so, the software or algorithm can then make empathic and educated decisions based upon what it understands presently and what it has “learned” in the past.
Now, deep learning, is not too dissimilar, albeit slightly different, as it’s regarded as a progressive evolution to machine learning; if you like, it’s “machine learning+” and, on reflection, it should have been named as such since, arguably, both machine and deep learning are often used interchangeably. Nevertheless, there are some differences and most notable is whether or not human intervention is needed. For example, a machine learning algorithm might be capable of performing a task to an extent and may, at some point, become confused or offer a prediction that is inaccurate or, at least, wasn’t expected.
Deep learning uses artificial neural networks to mimic the ability to learn - just like the synaptic mechanics of the human brain.
Software or algorithms must remain adaptive
In this instance, human intervention is required to correct or possibly resolve the open issue. On the other hand, deep learning will work through, with the use of better algorithms and “training” datasets, which it has acquired over a period to systematically resolve any issues without the need for human intervention. Furthermore, deep learning also uses artificial neural networks (ANNs) as a mechanism to “learn” and understand. ANNs, in fact, mimic the human ability to learn, where patterns in the artificial neural network, that is, akin to the biological neural network of the human brain, are established linking routine patterns just like the synaptic mechanics of the human brain.
Over a period of time, “neurons” will become established as routine or known behaviors, of sorts, to predict behaviors that have defined events and outcomes, although these networks remain adaptive so as to allow them to learn new things when new data is received, for example. However, most importantly, the algorithms must be sufficiently robust in nature to allow them to establish new neural pathways and patterns when they experience something new.
Developing advanced software and algorithm techniques
Anyway, back to the names, “machine and deep learning.” This does leave me wondering as to why there should be a distinction made between the two, as they model or behave moderately the same with some notable or minor differences. In short, this would explain why I consider this to be “machine learning+” for want of a better term. After all, we are just referring to better software or algorithms.
As I mentioned earlier, both terms are subsets of artificial intelligence, where deep learning is a further subset of machine learning, but I’m a little lost as to why they have been separated into two seemingly unique methods of learning. Fundamentally, I’m confident that, over time, we will inevitably develop advanced and stronger software techniques that will enrich the “holistic” essence of a machine that’s capable of learning, vis-à-vis “machine learning.”
Until next time …
For me, it is important to make this distinction of “machine learning” irrespective of the software or algorithm techniques used to bestow its capability. As such, instead of creating two domains of “learning,” we should perhaps just leave it as one concept that is “machine learning” as a subset of artificial intelligence, where we offer a granularity versioning or revision set to the evolutionary advancement of machine learning and ultimately a distinction might be made based on what it’s ultimately designed to do. I would suggest that a taxonomy of classification is developed to more accurately portray ML functionality and its inherent characteristics and qualities, all captured through software!
So, this is where your “AI psychologist who’s always deeply learning” Dr. G signs off
Originally published in Technically Speaking.