It seems like every week now we hear new stories about how some form of AI, or more precisely a specific type called Deep Learning (DL) will revolutionise the world around us in the form of self-driving cars, healthcare diagnosis, automated trading or hyper personalised marketing. There is no question that we are in the midst of a fast-moving technological revolution which presents a host of new opportunities for those able to act fast but also, beyond the bright headlines, a set of challenges which must be overcome if its promise is to be upheld and realised.

At their simplest, Deep Neural Networks (DNN’s) which realise Deep Learning advances do one thing well: they approximate or model the relationship between a very large number of inputs to a typically smaller number of possible outputs in an efficient way. Often that relationship is complex or non-linear which makes it impossible to model using traditional algorithmic approaches. Mapping large numbers of inputs to possible outputs allows us to capture the relationship between items of interest we may find in digital content and some form of meaningful classification. For example, we might wish to classify objects we find in digital images, so in that case, the inputs are pixels in the image and the outputs are classes of things – e.g. person, cat, dog, table etc.

 

Unrivalled Ability to Learn

The key thing is that DNNs typically learn this relation or mapping between pixels and possible classes from a large set of labelled, image examples. If trained well, the network can then generalise what it has learnt and predict the content classes of images it has never seen before with a high level of accuracy. This ability to learn from large volumes of labelled data and generalise well with images “in the wild” is what can set DNN’s apart from other computer vision techniques and previous incarnations of neural networks. This is interesting when you know that in 2017 alone approximately 1.2 trillion images are estimated to have been produced which is emblematic of the vast quantities of digital data being generated more broadly. It would be impossible to write a rule or standard computer vision-based set of programmes to robustly label these images at this scale given the variety diversity of their content. It is clearly inefficient for human labellers do to such a task given that it may take several minutes per image and consistency is extremely difficult if not impossible to achieve.

What is also interesting is that DNNs have been shown in recent years to outperform human labellers in terms of accuracy on certain tasks such as image classification. Not long ago, the prevailing view was that such tasks required unique human skills which could not be matched by computer techniques. This is of course a narrow achievement as there are many more vision challenges to be addressed but coupled with the fact that DNN’s can process data at orders of magnitude faster than human labellers and with greater consistency raises the prospect of a huge advance. We are now able to draw up to real-time insight from the world around us be it physical or virtual. Going further, the quality of the outputs gets better as we obtain more data and retrain or fine-tune our DNN models. Critically, we can also uncover previously unmapped relationships between classes in our data which can provide a whole new level of actionable insight.

“In addition to straight-up task productivity, consistency and efficiency gains there is a potentially more important outcome and that is in unlocking the know-how of our human capital.”

The Real Opportunity: Unlocking Human Know-how

Taking a step back this has the potential to make a profound impact if we can learn to fully exploit this opportunity. For example, consider a single DNN based MRI assessor which consistently provides first-tier analysis or anomaly detection outcomes for all MRI’s in the nationwide NHS network in near real-time. Its not hard to imagine the positive impact this could have on diagnostic variability, improving the consistency of outcomes and overall NHS efficiency and cost.

In addition to straight-up task productivity, consistency and efficiency gains there is a potentially more important outcome and that is in unlocking the know-how of our human capital. Consider for example a recent development SCISYS have been working on where we have produced a system which can classify the content of high resolution terrain maps. Currently, a handful of experts spend months on a first-tier analysis task which is extremely challenging given the scale, number of resources available and hard deadlines which often apply.

It also acts as a bottleneck which can delay higher level strategic analysis phases. The results from this first-tier analysis can now be available in minutes depending on the configurations we use. This allows these experts to focus on their core competency and interest which in this instance is understanding Mars and guiding its exploration. By using the results from the analysis output they can gain and act fast on strategic insight in ways which would not have been possible previously.

In another example SCISYS provides DL based systems which can label defects from tunnel environments in fraction of the time it takes a team of human experts to carry out. This could be carried out at a network level – simultaneously generating condition reports at large scales in orders of magnitude less time. Once again it brings a level of consistency that can be applied to all instances and filters out the variation in performance. This allows the expert team to focus on strategic insight and planning and ultimately to achieve significant performance improvements.

Such examples highlight key advantages of AI namely – augmenting the capability of human capital and repurposing those experts to address higher order tasks. Staff Repurposing to insight centric work, cognitive Augmentation coupled with Productivity at a task level through automation are at the heart of the DL revolution and its potential to bring about significant change.

“…key advantages of AI namely – augmenting the capability of human capital and repurposing those experts to address higher order tasks. Staff Repurposing to insight centric work, cognitive Augmentation coupled with Productivity at a task level through automation are at the heart of the DL revolution.”

Deep Learning (DL): a Confined Dragon?

However, using such powerful algorithms for high-level functions, where they impact or disrupt existing operations and strategy can be like introducing a technological dragon to a confined space. There are many roadblocks in achieving such gains which also makes it difficult for those who are entering the field and hoping to exploit its potential. Foremost among this is the need to educate and gain acceptance with those impacted by it, establish/ensure trust in the technology, and access appropriate but scarce expertise. Recent incidents in the automotive sector have highlighted the human and potential reputational/brand risks associated with deploying such complex technologies in shared environments. Thankfully though we can take advantage of lessons learnt and best practice from early use and adoption of AI in other industries.

For over 15 years the SCISYS autonomy and robotics group has led the introduction and development of AI based technologies for the European Space Agency’s ExoMars rover mission to Mars. This robot will autonomously explore the surface of Mars whilst searching for signs of life.

In recent years we have successfully transferred this know-how and our IP to industrial inspection applications in other extreme environments here on Earth. For space missions in particular and remote inspection in general the trust and confidence bar has been set extremely high. Onsite maintenance and repair is not an option on Mars or some remote environments should we have problems with our autonomous control systems. It is therefore critical that the entire range of mission stakeholders have trust in the technology from mission directors through to tactical operations staff responsible for rover well-being during live ops. We have been able to engender trust by creating and executing dedicated but extremely challenging technology trials accompanied by hands-on educational activities which allowed stakeholders to become comfortable and take advantage of what the technology had to offer. As one trial director noted:

I was very concerned about the inclusion of the AI Autonomous Navigation Component fearing that it would add unnecessary operational complexity, increase our workload and potentially endanger the mission … but it after a few days we realised that it did exactly what is was supposed to do allowing us to focus on the strategic planning, extracting insight and make much faster progress. In the end we just forgot about it as it worked in the background.

In addition to the educational and trust work, we have created a new approach to the validation and verification of autonomous systems to help minimise the risk of the kind of incidents seen in the automotive industry. Given the fast-paced nature of the AI revolution there is a high degree of constraint free development being adopted at present, but AI is a powerful tool and if not developed and deployed correctly it has the potential to inflict serious damage – either physical or reputational or both. Fortunately, techniques and expertise from mission critical industries such as Space can be adopted to address this challenge as part of an overall customer led process for deployment.

Conclusion

We are at the cusp of an exciting future enabled by AI and robotics technologies. Significant gains await in the form of Repurposing, Augmentation and Productivity (RAP) for those enterprises who are able to carefully introduce these complex technologies in a meaningful and managed way. Engaging and collaborating with established and experienced entities such as SCISYS who provide the requisite expertise and production level diligence specific to AI solutions. This will both enable and “chaperone” the introduction of this level of change whilst minimising adverse consequences allowing the smart enterprise to ride the dragon to a whole new level of success.