Computer Vision News - December 2022

15 Eleonora Giunchiglia “ I think what was interesting for them was thatwehadaclearapplicationdomain, ”she responds. “ These neuro-symbolic methods are powerful, but they’re still applied to toy datasets, while our work bridged the gap between theory and practice. That’s all it did. We said let’s port this neuro-symbolic world into the real world and see how it performs on a real-world, safety-critical task . ” Eleonora continues to work on this task in Vienna. She says the problem of incorporating hard logical constraints or requirements is understudied, and there is still much to do. “ Inthefuture,Ihoperequirementsolicitation and specification for deep learningwill beas normal as it is for software development, ” she tells us. “ If this message gets through, we’ll have much safer models, and nobody will lose anything. ” necessarily, says Eleonora, but they were built on a different assumption. Looking backat thehistoryof AI, in the ‘80s and ‘90s, people thought reasoning was enough . Then neural networks came along, and people thought data was enough . Now, it has reached the phase where people realize we need reasoning and learning capabilities . “ There is this whole neuro-symbolic wave happening, ” she continues. “ I hope my tiny contribution will be to say we need the reasoning not only to improve performance or to learn from fewer data, but we also need it for safety reasons. We can learn lessons from standard software engineering and logic and apply them in neural networks and deep learning. We don’t need to reinvent the wheel! ” Despite her modesty, this work is potentially game-changing. Does Eleonora think that is why the jury at IJCLR 2022 recognized it with a top award ? Predictions made by the 3D-RetinaNet model for the same traffic light in two consecutive frames.