Understanding Machine Learning

Machine learning (also popularly called “Deep Learning”), cognitive computing and AI are words that now feature in our daily vocabulary. Yet understanding what they actually mean is not quite as common as you might think.

The closest misapprehension we get is that Deep Learning is some form of artificial intelligence and computers are networked in ways that are like a human brain. Deep Learning is a very weak form of artificial intelligence where a complex algorithm is trained to recognize patterns during a supervised stage which then leads to greater confidence of accuracy in its unsupervised stage.

Considering that I just introduced two new terms trying to explain the first one it will probably be best if we take things from the beginning. The real breakthrough in machine learning has been in the use of convolutional neural networks (CNNs) which are networks of processors (the neural nets in question) structured in an overlapping way where processing redundancies are introduced in an attempt to correct pattern recognition mistakes on the fly and save computing cycle time.

Convolutional neural networks were initially used, with great success, in visual image recognition and have been inspired by the way the human brain processes visual information, but they have since found uses elsewhere, anywhere as a matter of fact where patterns emerge that can lead to a greater understanding of what is being indexed.

When we talk about machine learning we basically imply the use of a convolutional neural network of some kind and an algorithm that goes through two stages: a supervised stage where a human operator helps the algorithm go through masses of data and adjust its understanding of it and then an unsupervised stage where the algorithm is left to work on its own with occasional quality control intervention by a human, as warranted.

If this sounds familiar consider semantic search and Google’s human rater guidelines used to train its algorithms. Similar logic applies to Google’s Voice Search, YouTube recommendation engine and Google’s Image Search.

Deep Learning is not really ‘Deep’

With all these terms out of the way the question is do machines exhibit learning behavior and are they truly intelligent? These are two very different sets of questions. The term “Deep” as applied to “Deep Learning” has often been misinterpreted to mean learning of the same type that the human brain exhibits and that is not what is happening there.

Deep Learning refers to the convolutional neural net architecture where more than two layers of neural networks are involved. Neural nets these days can go pretty deep and what sets them apart from an ordinary computer network is that each stage of the neural net can actually be trained. The intelligence aspect of the behavior is also the result of the network’s architecture where a memory is attached that remembers the results obtained which means the neural network apportions resources to ‘learning’ new stuff by recognizing new patterns. Where the new patterns are derivatives of the patterns it already knows. An example of this is teaching a neural network what a dog is by looking at a few thousand pictures of Alsatians, Labradors and Poodles and then allowing it to understand that “dog” refers also to Pit Bulls, Dobermen and Chihuahuas.

This is not true intelligence in the way humans exhibit it, though within very narrow contexts it’s close enough, especially if a recursive neural network is created where the computational outcomes are fed back into the network (which is why Google Voice Search can hold a conversation with us on Barrack Obama’s marital status and children, without our having to specify the context every time).

Machine Learning Everywhere

Moving forward we will see the use of machine learning everywhere. The uptake is driven by the same reasons that made semantic search a no-brainer: lower cost implementation, greater savings and greater reliability in complex results.

And Google have just opened sourced too:


This will mean that our devices and apps will become ‘smarter’ in the sense that they will now be more responsive to the environment and more focused in the context of our needs when we interact with them. Smarter devices means we can now make better use of information and, as neural networks spread everywhere, better informed decisions and choices.

It is our ability to use information in such a smart way that actually makes us, humans, intelligent.