Artificial intelligence is trending high in the hype cycle. It seems as if every company has a product claiming to have artificial intelligence built in, but that is not quite how it works. We asked David Thogmartin, Head of the Deloitte aiStudio and “Analytics Data and AI” practice in Deloitte Germany, about artificial intelligence in the real world.
Q: Could you give our readers a brief snapshot of what is currently happening across the artificial intelligence (AI) landscape?
David Thogmartin: AI is a hot-topic and a rapidly expanding field on a global scale. A lot is happening! Two of my favourite topics are explainability and efficiency of AI models, because both well represent the trade-offs that we – as data scientists, business people, and society – must contend with…
More complex (AI) models tend to better capture the dynamics of the systems they seek to approximate, i.e. there is a trade-off between accuracy and transparency. Deep neural networks count among the most accurate AI models and are notoriously black boxes. This is disconcerting to the human need to understand how and why a conclusion was reached or a decision made. Regulated industries in particular demand such transparency, for example, why someone has been refused credit or a mortgage.
So, how do we explain the workings of machines to humans? Given only inputs and outputs, how can one tell how many hidden layers, how many nodes, whether they are recurrent, convolutional, or fully connected… all daisy-chained together to form a complete model? A tough problem to solve, but an important one, which explains the high degree of activity in explainable AI (XAI).
Another area of concern is model efficiency. The highest performing models are computationally expensive and thus leave behind a substantial carbon footprint. Sometimes simpler models, while less accurate, are “good enough”, presenting a reasonable trade-off between performance and environmental cost.
Concern about global warming has cast a spotlight on energy-intensive, cloud-based deep neural networks, encouraging research into optimisation and efficiency, an effort combining new methods, better software implementations, and tighter hardware integration. This, in turn, paves the way to faster models that can be implemented on local “edge” devices vs requiring a constant connection to “the cloud,” a boon to truly “smart” devices that not only sense their surroundings, but also alter their behaviour in response.
Q: What type of projects are Deloitte’s clients, particularly those in the financial services and retail sectors, looking for support with around AI at the moment?
David Thogmartin: This is also a wide spectrum with a vast range of use cases. Off the top of my mind: fraud detection, credit decisioning, ensuring fairness/ethics in AI models, transparency (XAI), shortening of cycle times, assisting with data migration, building data lakes using AI, automation of labour-intensive processes, enhanced cyber-security, and building more useful models to predict potential outcomes – to name a few, in finance, manufacturing, medicine and sustainability.
Q: AI and machine learning have become buzzwords. Every second vendor claims to be using these technologies in their solutions. How do you help clients evaluate the various options currently open to them?
David Thogmartin: I ask questions. What problems does the client face? It is not constructive to ramble on about great technologies if the use cases are not relevant to the client. Dialogue is key.
At the same time, innovative clients want to see what is possible; they want to ignite their imaginations. They look to us to bring both the breadth and the depth, the combination of functional expertise, industry experience and technological know-how.
Q: How do businesses prepare themselves for an AI-driven future and the radical change that will result?
David Thogmartin: Start with the mind-set. Be agile. Be daring. Take calculated risks. Launch and learn. AI will require iterations to get right. Don’t be dismayed if the first pass underperforms. That is normal. We see banks doing this when helping them re-tool and re-organising their teams to break old moulds and usher in the flexibility of more nimble start-ups.
Secondly, understand your data, its riches, and how to extract value out of it. AI is an incredible tool to do exactly that, but must be applied correctly. The big tech companies of today realised that early on. We are helping our clients realise similar advantages to differentiate themselves from their competition.
Q: What is the future of AI (trends, technological advancements, take-up etc.) over the short-, medium- and longer-term?
David Thogmartin: In the short-term, we will continue to refine “narrow AI” (data-trained algorithms applied to very specific tasks): new methods will be developed, offering greater accuracy, greater transparency, greater efficiency. Short-term, we will see more devices that can interact with us on human terms, reducing barriers to technology and effectively augmenting our intelligence. Mid-term, we will see autonomous robots and vehicles that can “reason” the best course of action in unknown situations.
Over the long-term, who knows? Many technologists and pundits muse about an AI-driven future: some beneficial, some dystopian. One thing they all have in common: they see AI as an unstoppable force. I would argue it is actually humans who are the unstoppable force, powered by strong feelings and the belief that they can make a difference – to whatever end – using this technology. Machines don’t have such urges.
We continue to see advances in artificial general intelligence (AGI), for instance with reinforcement learning: training models on “10.000 years of experience” within hours through simulation. Open AI taught a (single) robotic hand to solve a Rubik’s cube. DeepMind moved beyond board games into the realm of open-world role-playing games like Starcraft, able to beat the best human players within days of training. This is equipping machines with the capacity to understand or learn any intellectual task that a human being can.
Perhaps the greatest breakthrough of all: AI has escaped the lab and established itself as a strong economic contributor. Yet artificial general intelligence remains a long way off. Are we nonetheless heading toward the “singularity”, where computers comprehend their existence, wake up and question whether they want to do our bidding? Yuval Noah Harari states in Homo Deus that where we’ve seen great advances in artificial intelligence, we have witnessed no such advances in artificial consciousness. On the flip side, Geoffrey Hinton contemplates that there is no reason why we could not, someday, replicate emotions in machines. I personally believe we will indeed one day reach the “singularity”, simply because humans won’t stop until we do.
This interview was originally published in the exclusive RiskConnect 2019 Magazine and updated for 2020 by David Thogmartin.
RiskConnect returns this year in virtual form! Find out more about the conference here.