How the human perception system can improve AI

 Since October 2015, when Tesla introduced their semi-autonomous driving system and the technology in their cars to recognize pedestrians, other vehicles, and everything else on the road, almost every other company in the car industry have followed suit. Recently, with the almost exponential rise of AI in the past year or so, this technology has the potential to become extremely advanced and can possibly become standard in every vehicle in the near future. However, AI and these recognition technologies have their limitations. The camera's that aid in the recognition system are not perfect and can easily mistake a stopped vehicle as an obstacle on the road and trigger an erratic lane change or sudden braking, which can possibly be detrimental to the passengers and others on the road.

    Enter Nicholas Baker's research article titled Constant Curvature Segments as Building Blocks of 2D Shape Representation. Dr. Baker and his team investigate how the human visual system represents shapes using constant curvature segments. The researchers explain that human perception is adept at recognizing and encoding contour shapes by breaking them down into curved segments, which are essential building blocks of shape representation. The brain can then easily piece these shapes together, even when faced with a complex or partially completed object, thanks to specialized neural systems that process curvature and boundaries. 

    Now, as great as AI is, the technology is not as perfect as the human brain. Philip Kelman at UCLA joined "The Science Show" on ABC and discussed the topic of robot and AI visual limitations. The conversation was dictated in the article titled "Challenges for AI visual recognition". Dr. Kelman discusses how technology systems, "when probed for the shapes of objects in the absence of other cues, these devices fail miserably", despite advancements in image recognition abilities. Dr. Kelman explains that AI struggles with shape recognition of abstract objects and will usually mistake them for simpler objects. He further goes on to talk about how AI lacks the symbolic coding that human perception is exemplary at. Human perception mechanisms naturally break down objects into understandable parts and shapes (as we stated earlier); this is something AI systems are unable to do yet. Kelman then specifically talked about driverless cars, where shape recognition is crucial. Because of the problems current AI systems have in perception, Kelman sees driverless cars to be a risk. He gives an example of the technology not being able to recognize a stop sign covered in graffiti, making them unreliable in real-world situations. 

    Kelman underscores how vast the differences in AI and human perception are. The human brain has evolved to process intricate shapes through constant curvature, as Dr. Baker explained in his paper, a skill AI has been unable to achieve. However, with technology constantly advancing, AI developers and engineers can surely benefit from studying the human brain's skilled way in handling perception through the break down of curvature. Developing algorithms that mimic the brain's ability to decompose shapes into simple pieces could greatly improve AI's ability to recognize objects in real-world situations, where visual cues and context can be scarce. Though it may seem scary to us, providing AI with a "brain" like system where "neural networks" are used to focus on curvature processing could greatly improve the ability of these AI systems. Although it is easier said than done, as driverless cars and AI technology's implication in everyday life seem inevitable, improvements in this specific aspect of the technology can prove to be extremely crucial as lives could be saved and improved.

    Bridging the gap between AI and human perception skills can significantly enhance the efficacy of AI systems. The connection between neuroscience and artificial intelligence is a relation with limitless potential. As researchers continue to explore these connections, the future for AI is bright. 

References

Baker, N., Garrigan, P., & Kellman, P. J. (2020, December 17). Constant Curvature Segments as Building Blocks of 2D Shape Representation. Journal of Experimental Psychology: General. Advance online publication. http://dx.doi.org/10.1037/xge0001007 

Williams, Robyn. “Challenges for AI Visual Recognition.” UCLA Human Perception Laboratory, UCLA Human Perception Lab Department of Psychology, 1 Apr. 2019, kellmanlab.psych.ucla.edu/post-challenges-for-ai-visual-recognition.php.