"Demystifying Machine Learning"

By Graham Wilkinson, Global Head of Product

Before we jump into the definition of machine learning we first have to clarify its context in relation to a term much more common in popular culture today, artificial intelligence (AI). AI was a term first coined in 1956 by John McCarthy at Dartmouth College at a conference where the topic was first discussed. Put simply, artificial intelligence is the ability of a machine to perform tasks that are characteristic of human intelligence.

With that definition in mind, how does a machine go about performing these kinds of tasks, which are so closely associated with human intelligence; planning, object and sound recognition, language understanding, problem solving and learning? Well, one of the ways a machine can go about doing this is through machine learning.

Machine learning can be defined as the field of computer science that deals with algorithms that learn from, and make predictions on, data without needing to be explicitly programmed.

How does a machine learn?

So now we have established that machine learning is one of the means by which a machine can try and achieve a form of artificial intelligence. But how might a machine learn? Just like human beings learn in many different ways, there are also several types of learning techniques which can be employed for machines.

  • Supervised – An algorithm is presented with ‘labelled data’ which highlights inputs and desired outputs. The algorithm learns the rules which connect the inputs and outputs, and can then use these general rules to predict future outputs, using only the input data.
  • Unsupervised – In this instance, the algorithm is given no data labels whatsoever and is therefore left to find structures and patterns in data. This type of learning is obviously less structured than the supervised type, and therefore is less focused on the output data and more focused on any hidden patterns which it may well be able to discern. This is important when considering that data labelling often imparts conscious or unconscious human bias into the learning process.
  • Semi-supervised – Uses a combination of both labelled and unlabeled data, usually a greater amount of the latter. Researchers working in the field of machine learning have found that this type greatly increases the algorithm’s learning accuracy.
  • Reinforcement – With reinforcement learning an algorithm is given a specific goal which could be anything from selling a product online or playing the popular game Go. Each step that the algorithm takes towards the goal is either rewarded or punished, forcing it to find the most efficient path towards the objective.

Not to complicate anything unnecessarily, it’s worthwhile talking briefly about some very closely related topics: neural networks and deep learning.

Neural networking is focused on processing data in a similar fashion to the way the human brain works. There several types of neural networks, but the easiest way to think of them is that any point (or node) in the network that the machine learning algorithm learns is also connected with other nodes whose learnings in turn may be dependent upon the outcome of the connected nodes. This simply allows neural networks to learn from observational data.

Deep learning can be thought of as a powerful set of techniques by which learning can take place within a neural network. The reference to ‘depth’ in this term comes from the multiple inter-dependent layers which may exist within a network, whereby the output from one layer may well be the input to another layer.

What can we do with machine learning?

With machine learning being an ever-growing field of study, it’s fair to assume that the associated list of capabilities are also growing. But there are some which are more common and familiar to us.

  • Knowledge representation – Information about the world (anything) is represented in a way that a computer system can use it to perform complex tasks. These tasks could be varied, from diagnosing disease to holding a conversation with a human being. In the latter example, this is used is to understand the complex semantic relationships between concepts so that a computer can formulate an appropriate response to a human in a conversation. This is used in search engines to determine the difference between a nail (finger) or nail (construction) in a sentence and provide relevant results.
  • Natural language processing – Computers process human (natural) languages in order to recognize speech, conduct translations and understand sentiment. This is used in conversational voice devices such as Google Home, Amazon Echo to recognize speech and commands.
  • Speech to text – Neural networks convert audio signals to text in a variety of languages. This is used for translation purposes, voice command technology and audio transcription, and mobile devices to dictate direct to messaging apps.
  • Expert systems – This a system which incorporates specific knowledge (medicine, legal, etc.) and combines it with a rules engine which determines how the knowledge is applied. This system can be improved both by adding more knowledge, or updating the rules. Used in NASA mission control for interpreting predicting, repairing and monitoring system behaviors.
  • Predictive systems – A system which finds relationships between variables in historical data sets and their subsequent outcomes. Once identified, these relationships are used to develop models which can be used to predict future outcomes. Used in stock market trading to determine when and which stocks to buy or sell.
  • Optimization – An approach to problem solving where a variety of possible solutions are searched through to find the optimal solution. Used by Amazon in their approach to warehouse management and package delivery.
  • Audio signal processing – Machine learning that analyses audio and other digital signals, in particular in high-noise environments. Used in selective noise-cancelling in your headphones.
  • Computer vision – Computers learn to identify, categorize and understand the content within images and video. Used to identify unsafe branded content on the likes of YouTube, Facebook, etc.

Where can you start with Machine learning?

There are numerous ways to start using something powered by machine learning but the chances are that you’re already interacting with several forms of it. From checking the weather forecast in the morning, to looking for the best route to your destination on your mobile, through to something as simple as checking Facebook. All these things have elements of machine learning powering them.

But if you really want to get your hands dirty, then one of the easiest ways is to activate Smart Bidding on your Google campaigns. I assure you that this is not a shameless plug for Google. The reality is that this feature gives us (as digital marketers) access to a very powerful Machine Learning bidding algorithm. For those of you who don’t know, this feature offers:

  • True auction-time bidding solution
  • Adaptive learning at a query level
  • Richer user signals and cross-signal analysis

You can learn more about this feature here, but it is relatively easy to activate and you should ask your local Google representative about it if you’d like to know more. Even if you have no further interest in machine learning beyond this article I would encourage you to activate this feature as it is a core pillar of our best practice methodology.

Finally, for the more adventurous of you interested in playing around with some machine learning-based APIs, there are several cloud-based solutions that will let you analyze image, video, speech and language. If you do go and visit any of these sites and come up with a great idea for us to implement into our Reprise Product suite, feel free to email me at graham.wilkinson@reprisedigital.com. Here are some links below for the APIs I mentioned above: