Understanding customers’ experience can help business owners make decisions to improve their service. Existing techniques such as surveys, interviews and interactive screens are commonly used to get direct feedback from customers after they experience the service, but it seems humanly impossible in larger areas such as airport which has more problems and complexity.
Human emotion can be seen as implicit feedback. It is non-verbal communication showing how a human reacts to the surrounding event. For example: when we have a delayed flight, we tend to show sad expression. Facial expression can contribute 55 per cent to the message’s effect.
Artificial intelligence (AI) enables us to have a smart model to recognise emotions quickly regardless of gender, age and race, and can even surpass human performance. Using deep learning technique, a deep learning (DL) model is trained by learning from millions of facial images. However, such emerging surveillance technology still lacks people’s trust to adopt it as it is considered as a technology that intrudes people’s privacy, regardless of the benefits.
Trust in adopting AI solution for real business
AI’s decision-making processes are not always transparent. When people do not really understand how AI makes decisions, it is human nature that many people will not easily trust AI, nor will they fully adopt it or make use of its positive capabilities. We tend to focus on what AI can offer but forget to gain trust from people, expecting them to immediately fully trust the AI solution. Seeing benefits does not guarantee that people will trust the AI solution.
Availability of data and maintaining models’ performance over time
Existing generic DL models are not accurate enough to solve real-world problems (e.g. head pose and occlusion variations introduced over time). Existing datasets, such as AffectNet which has a half million facial images, are still limited in representing the real-world’s complexity. Therefore, more data must be collected continuously from cameras to update the DL models over time in order to adapt to the cameras’ characteristics. Hence, DL models’ performance can be improved and maintained over time.
Dealing with sensitive data like people’s faces can raise serious data privacy issues. Data misuse and data theft is possible when data is retained for a long period of time.
To gain people’s trust, it must be built and maintained gradually throughout the process by involving all stakeholders. To improve the DL models’ performance, ‘continuous learning without forgetting’ approach enables DL models to update their knowledge continuously without old data and long-term data retention. Federated learning approach enables DL models to update their knowledge within camera, making data more secure as no sensitive data is allowed to leave the device.
The proposed solution is currently being implemented in the context of a local airport while showing how those challenges are gradually addressed. The solution provides automatic, prompt feedback in real-time, hence reducing human involvement and helping people make well-informed decisions without intruding people’s privacy. In the long term, this analysis will be applicable to any public surveillance and home surveillance systems.
The study has been very challenging yet interesting as it goes beyond an IT project, to managing technology so industries can fully gain the AI solution’s benefits, while ensuring people are comfortable using it. I am truly grateful to my both supervisors, Professor Dian Tjondronegoro and Professor Rosemary Stockdale for believing me and giving me ongoing support!
Nehemia Sugianto is a PhD candidate at the Department of Business Strategy and Innovation, conducting research to develop a privacy-preserving facial analytic surveillance solution built on responsible AI technology.