Happy to announce a new and exciting collaborative effort between the Kietzmann lab at Radboud AI and the ROSEN group. Energy asset inspection shares many aspects with problems that the brain faces: Indirect measurements, sparse ground truth data, large volumes of data with few relevant events and strong energy constraints. Together with the ROSEN group, we aim to develop and test neuro-inspired AI systems that take inspiration from neuroscience to address these issues.
Energy asset inspections face several key challenges that are a direct match to the benefits and promises of neuro-inspired AI. First, the vast amount of collected sensor information contains few relevant events that need to be identified with few false alarms. This requirement finds direct correspondence in the brain’s ability to learn from sparse data and in its ability to filter and compress sensory information, for example by unsupervised learning mechanisms based on predictive coding. Second, the interpretation of sensor data into actionable information depends on contextual information that changes on multiple timescales. Again, brain-inspired AI systems hold promise for a solution. Recurrent neural network architectures, which take direct inspiration from largely recurrent connectivity found in the cortex, are capable of considering context information while maintaining a relatively small memory footprint. Energy asset inspection thus allows rigorous real-world testing and further developments in the domain of neuro-inspired AI.