Setting the course for Machine Learning
Artificial Intelligence (AI), Machine Learning and Deep Learning are all terms that are coming to our attention ever-increasingly in industry — particularly for me in transport in relation to the rapid development of the systems that will support the roll-out of fully autonomous vehicles.
While AI encompasses the broader goal of computers that can learn and act, machine learning is much more specific sub-set of AI which can be used for solving well-defined problems. Deep learning is a further extension of machine learning, which expands the concept of neural networks (which are inspired by the functionality of the human brain).
Unlike usual algorithms used to perform specific tasks, machine learning methods are employed to learn how to perform a specific task — learning as more data is provided. Just as we have different learning styles, there are (quite a few) different ways which a machine can learn. These methods can be categorised into either supervised learning (where the algorithms have a training dataset to learn from) or unsupervised learning (where we are interested more in discovering underlying patterns and structure in data).
As our computing processing power increases, storage becomes cheaper and data sources richer there is an increasing demand to the develop methods and skills to solve problems with machine learning in many domains — which is potentially any that involves data and identifying patterns.
The Course
Last week, a bunch of Arupians participated in a three-day crash course in Machine Learning, developed in partnership with the University of Technology Sydney (UTS) Advanced Analytics Institute. The course was run by Dr. Richard Xu who directs the university’s Machine Learning and Data Analytics Lab (and who is not shy to include a few complementary ‘big data’ jokes).
Our team’s individual backgrounds were diverse — ranging from spatial sciences, software development, geology, transportation, structural engineering and building physics. Although from different fields, we all had a common interest and purpose in learning the skills to apply and scope out Machine Learning and Artificial Intelligence projects within the built environment.
The course focused on learning the fundamentals of machine learning, while gaining knowledge of applications of current tools and techniques available. In particular, we focused on understanding the statistical learning techniques that allow automatic identification of patterns in data. An outline of the topics can be found clustered below:
The Tools
While we focused on Matlab and Octave there are also many other options such as Python’s Scikit-learn, Azure Machine Learning Studio and R. All of these have a great set of tutorials available to hit the ground running on Machine Learning.
Current applications of machine learning
Machine learning applications are already ubiquitous in our everyday life.
When you log into Facebook and someone has tagged you in a photo, that is a prime example of the roots of machine learning, which reside in image and facial recognition. Not only does it recognise that it is your face, but also that you have a human face based on the features and relationship between your pixels and all other pixels in the image.
When you speak to Siri on an iPhone, it recognizes your words through speech recognition. When you use Google Translate , the sequence of words you used is likely being translated now by something called a recurrent neural network.
When you open your email (mostly) free of unwanted messages, you can thank machine learning for the spam filter — which is likely powered by a technique which has classified junk from non-junk based on the nuanced features of many millions of spam-classified emails.
When online shopping, or browsing Netflix, recommendations are are given to us on what we are likely to watch from an algorithm of people who are likely similar to us, and have made similar choices to us.
So away from use cases that we can mainly find on our mobile phones and the internet what are some potential applications in the physical, built environment?
Built environment applications
1) Intelligent Transportation
Analysing data to identify patterns and trends is a key component of the transportation industry, which relies on making routes more efficient, meeting travel demand and predicting potential problems.
One of the main draws for me to the transport field is the sheer amount of data that is being generated by the movements of people and vehicles , and how important these movements are to such a wide variety of groups — from retailers, to delivery companies, to public transport operators, to property developers, advertisers, employers, planners, economists and toll road operators.
Data such as the soon-to-be released Uber Movement, Strava, GTFS Real-time and even Geolocated Tweets are both contesting and complementing our traditionally-used methods. As these technologies are continuously collected they have an advantage over traditional travel data collection methods, which are generally cross-sectional in nature and generally, due to costs, sampled at a much less frequent rate.
With more data, we are much better able to infer behaviour before and after a specific event (such as disaster or disruption), as well as over a long-period of time from continuous data coverage. As machine learning is a data-hungry discipline, this places many transportation data related tasks well to be optimised.
Autonomous vehicles
Machine learning is well-known to be integral to driverless cars, which use complex image and spatial recognition to identify road features, pedestrians and other vehicles in order to provide seamless, automated travel in cities. Using these sensors and onboard analytics, cars are able to recognise objects and react appropriately using Deep Learning. MIT’s Moral Machine is an experiment on how machines might decide in crash scenarios, where human moral decisions are collected and analysed against the same decisions machines would need to make.
Transport services on-demand
Whether empty-running or on fixed timetables, bus routes could be dynamically altered to meet passenger needs. When the weather is bad, buses could be put to use to keep up to pace with increase in ridership. Routes could be adjusted dynamically to better fit with door-to-door demand for the users who have opted into that service.
Personalised trip-planning
Mobile phone applications could review the travel options available to you and make personalised recommendations using machine learning to account for preferences such as lifestyle choices, fitness levels, previous locations visited, amenity along the way, budget and dynamic predictions of congestion en-route.
Crash and congestion monitoring and response
Image recognition could monitor and recognise both congestion and road accidents before, during and after they happen. These would allow systems to adjust the road conditions (such as variable message signs and speed limit) systems accordingly, as well as notify traffic control and emergency services.
2) Energy & Utilities
The optimal, sustainable use of our energy systems both at the supply-side and the user-side is critical to sustainable future. Being able to better predict accurate energy consumption forecasts can help implement better energy-saving policies in cities. A number of machine learning applications could benefit further investigation in this sector. For example:
Optimising consumer energy use
The Nest Thermostat uses machine learning to learn a homeowner’s preferences and schedules to optimise heating and cooling.
Creating a smart energy grid
The energy sector in Germany has employed machine learning to optimise the power grid and manage the maximisation of renewable versus non-renewable energy.
3) Environment and waste management
Automated waste collection and tailored collection services
Several years ago, Volvo announced that it was developing robots to replace the physically-demanding, sometimes dangerous task of garbage collection with a more automated, robotic system. To add to this, machine learning could track and predict waste levels in a city’s bins and manage demand (and charge users) accordingly.
4) Architecture and urban design
Informing design through crowd-sourced perception of place
MIT Media lab have collated a large set of data on people’s perceptions of safety to feed a machine-learning algorithm which determines how safe a street may look to the human eye. This kind of research could alter the way we design spaces to suit the potential emotional goals or needs of a space, and help us understand which features we ought to include or exclude from design.
Next steps..
It is clear that machine learning has many current and potential applications in the future for the built environment. For many of us on this course it was the first step in our machine learning journey.
The fact that we could get together and gain this deeper knowledge, apply faces to names of those around Australasia and form a common network could not have been possible without the generous support of Arup University and the Digital Services Skills Network.
I’m looking forward to seeing where our new network can go in machine learning and how we can collaborate the with industry, government and academic partners in tackling issues in cities.
