How do computers learn? An overview of patent activity in machine learning
If you were asked the rather philosophical question of “how do humans learn?”, your answer would probably include “from experience”. The positive and negative feedback we receive, and our observations of others performing similar tasks, helps us improve at whatever task we set our mind to, whether you are a pathologist reviewing tissue samples for signs of cancer, or a financial expert reviewing transactions for fraudulent activity, or a interpreter translating from one language to another. From seeing previous cancerous and non-cancerous slides, fraudulent and non-fraudulent transactions, and good and bad translations, we learn how to make the relevant distinction.
The answer to the question of “how do computers learn?” is surprisingly similar. It would be almost impossible to create an algorithm by hand that captures the nuance of what makes a tissue sample cancerous or a financial transaction fraudulent. However, provided with sufficient labelled data (i.e. experience), computers can generate a statistical function (i.e. learn) that maps input data to a category (e.g. cancerous or non-cancerous). The classification and clustering of data using this approach is referred to as machine learning.
We live in a world where a vast quantity of digital data is generated on a daily basis – an often-quoted statistic is that 90% of the world’s data was created in the last two years. The prevalence of these massive data sets (so-called “big data”) has opened the flood gates for the application of machine learning to solve an enormous variety of problems, and the sheer quantity of data means that the learning algorithms make better generalisations over the data, avoiding over-fitting and improving accuracy.
This surge in the application and sophistication of machine learning has been accompanied by a corresponding surge in patent filings over the last 5 or 6 years, as those innovators developing better algorithms, or applying them to new problems, seek to protect their inventions. As the graph below shows, the number of applications filed each year more than doubled from 2011 to 2015.
So what are the key problems this technology is being applied to? The data shows that a significant component of those machine learning filings (around 40% per year) relate to bioinformatics. Bioinformatics is the application of computational techniques to interpret and analyse biological data. This encompasses a wide range of biological data processing, from identifying cancerous cells, to exploring gene-protein interaction for drug discovery, to mining the enormous back-catalogue of academic publications in the biomedical domain.
Given the huge expense involved in traditional drug discovery, it is no wonder that computational methods offer an attractive alternative. However, it’s not necessarily traditional big pharma and universities that are paving the way in terms of bioinformatics IP – big tech companies are getting involved too, with IBM recently obtaining granted patents for discovering potential side effects of drugs and generating textual description of medical images using machine learning techniques.
There’s also been a real growth in the past 5 or 6 years in machine learning filings that relate to FinTech – the emerging field that relies on cutting-edge computing to support banking and financial services. For example, machine learning can help to detect fraud, or more accurately derive credit scores. Some of these applications will skirt the boundaries of what can and cannot be patented in Europe, the US and elsewhere, but that has not stopped interested parties seeking to obtain protection.
What about the actual detail of the computational and mathematical approaches that are being used? The landscape below, itself generated using unsupervised machine learning to cluster concepts in patent data, gives an indication of the trends. There are a multitude of techniques being employed across the board – support vector machines (SVMs), neural networks (including convolutional neural networks and the deep neural networks that underpin deep learning), bespoke statistical algorithms, and genetic algorithms all form significant features in the machine learning patent landscape.
© Questel 2017, courtesy of Patent Seekers Ltd
As we look to the future, there is no reason to think that this growth in machine learning IP will slow down. Innovators will continue to explore the ways to exploit the quintillions of bytes of data that are generated each day, developing better algorithms and applying them to new problems.
At Appleyard Lees, we have attorneys with first-hand experience in the development and application of machine learning algorithms, as well as extensive experience of drafting and prosecuting machine learning related applications. For more information, please contact Andrew McKinlay or Julia Gwilt.
Return to Articles