Do you need a drone that can identify human action in the wild using machine vision to alert authorities to the presence of poachers? If you think this is still a couple years or even decades away you’d be wrong – this is now.
Do you need a mobile app that can identify any make and model of car simply by using the camera? Practically impossible using traditional programming techniques – now a reality!
Do you need to increase the accuracy and speed of tumor classification for your oncology lab? Reduce processing times, cut costs, save lives.
Here at DataProphet we have noticed that there is a real lack of human-friendly information on the amazing capabilities of machine learning and artificial intelligence in the market. Leave all the jargon to us – this is what you need to know!
Machine learning does not compete with BI departments. ML techniques are complementary and do not produce descriptive statistics – i.e. graphs and things that are ‘nice to know’. Rather ML produces actionable insights – who to sell what product to, which agent to route the call to, what stock to buy when and so on. ML is not something you can do in Excel and the skills are currently very scarce – the pioneering Stanford machine learning course started just a few years ago!
Machine learning was first meaningfully postulated by Alan Turing but has only recently come to the fore due to the exponential advances in computational power achieved since the 1940s. In 1959, Arthur Samuel defined machine learning as a “field of study that gives computers the ability to learn without being explicitly programmed”. It exists at the intersection of several fields – statistics, computer science, engineering and mathematics.
In the past 5 years, new breakthroughs in the use of GPUs (graphics processing units) have yielded speed increases of 10-100x over traditional CPU (central processing unit) processing when using these algorithms. This has made the use of these incredibly powerful techniques possible for entities smaller than a nation-state with a supercomputer.
Self-driving cars. Applications able to correctly classify tumors with greater accuracy than a team of oncologists with decades of experience between them. Prediction of stocks, bonds and derivatives. Web search results. Product recommendation engines like those in use on Amazon. Email spam filtering. Fraud detection. E-discovery systems for legal practices. Credit scoring. Natural language processing – actually understanding the intent behind a sentence and being able to glean culture and language from name. Identifying an object from a photo such as a car’s make and model.
Machine learning is in use in a myriad of industries and has allowed us to quickly and easily increase key metrics such as productivity and efficiency – even number of sales – by as much as 20% for numerous processes. Not only has machine learning allowed us to do this – but it has also opened up the possibility for doing things that were never thought possible before, such as the self-driving cars from Tesla, Google and others that are on the road today!
There are two main limitations to what is typically required for a machine learning process – a large dataset and the requirement that the model be human interpretable. There are ways to minimize or circumvent the second limitation but the first is immutable.
Machine learning is best applied to large datasets. No human being can look at a dataset of ten million rows and find a complex relation between variables just by eyeballing it – but a machine can. This is where the true power of machine learning is found. ML has the ability to find patterns that may have been hiding in plain sight for decades – and these patterns may just make or save your company millions.
The dataset does not necessarily have to be owned by the company. It can, for example, be in the public domain. The data to train a model to analyze sentiment in news articles or on social media such as Twitter is out there in the wild and free to use.
Machine learning models may not be very easily interpretable by a human – that is to say that we might not know exactly why a certain output has been given – but the results speak for themselves. These types of models are termed ‘black box’ models and the general trend is that as a model tends towards being a true ‘black box’ in terms of human interpretability so does the power of the model increase. Industries where the reasons for submitting a particular output are required to be known exactly – such as credit scoring, where several reasons are required for why an applicant was declined – are not good fits for the application of ML techniques.
ML is great at predicting future trends and probabilities by using past data. Essentially with enough of the right data any predictive question could be answered. Even events that may seem unpredictable are more reliable than you think – top machine learning algorithms had accuracy rates of over 90% in correctly predicting the results of the matches in the 2015 Rugby World Cup!
Data is critical when considering applying ML techniques – to take advantage the data needs to be deep. The deeper the data is, the more complexity the model can discover in the data. As you can imagine, to discover nuanced relationships in noisy data of the real world, the models requires substantial evidence. In addition, ML techniques are typically very well suited to modeling wide data e.g. many columns of variables – traditional statistical techniques find it rather difficult to use all the available variables but ML techniques can.
Machine learning returns actual predictions and so system integration is usually required to make best use of it. The returned prediction might be what object exists in an image or which customer to approach with which product.
Do you need a serious machine learning company to unlock the value in your data and build a bespoke machine learning solution for your business?