Why AVs May Need Lessons From a Human Brain?

Humans are the most intelligent species on this planet. Humans have travelled to space, climbed the highest mountains on Earth, and flown faster than the speed of sound. Yet Autonomous Vehicles fail to even adapt a fragment of human intelligence, though they were initially thought to be smarter than humans with their 360-degree superhuman vision and penetrative waves that can see up to 300m in line of sight. Why1?

Because it isn’t just about sensor performance. For instance, consider how a mantis shrimp1 has 10 times better eyes than humans yet very poor navigation skills due to the lack of associated brain matter. Developing a basic human-like intelligence is essential for building safe and effective autonomous driving systems as we can’t just rely on the capabilities of sensor physics to cover for the shortcomings of existing AI approaches.

Humans are really good drivers. An average human can drive almost perfectly in 99.999897%2 of scenarios. To drive safely in a human world, these robots must learn to think like us — or at least understand how humans think. But how will they learn? Maybe they can take some inspiration from the human Learning Process.

The Learning Process:

As humans what we are really good at is learning - learning anything. Our brains never cease to learn - from learning to crawl as an infant or to flying an airplane, we are learning constantly. But if you closely follow, our brain is doing the same process perpetually through infancy, adolescence, and adulthood.

This process is a continuous cycle of having a basic understanding of the world → reasoning to predict & assess consequences → making a decision → ingesting information from our senses → updating our prior understanding of the world, if needed.

From Generalization to Specialization:

Unlike current task-driven AI models specialised for a specific use, our brain’s process of evolving our understanding, often presumed to be a “world model” is contrarian. We develop a generalised understanding of the world and evolve our reasoning capabilities, and eventually use that knowledge to specialise in one domain quickly. And that prior understanding allows us to specialise in multiple domains quickly, making us a multi-task species. Our array of domains of specialisation keeps changing with time and our interests, yet each contributes to evolving our “world model” to be more rational, comprehensive, and reliable.

Whether it's selecting a career path in engineering or psychology or excelling in specific skills like driving, singing, or playing football, all of these represent acts of specialisation built up on our generalised intelligence.

Our understanding evolves over four stages in our life:

? Foundational Understanding

? Intellectual Understanding

? Specialisation

? Refinement

And the interesting part is that The Learning Process remains the same across all stages. Let’s illustrate our learning process across different stages of our life with an example of how humans learn to drive.

How do we learn to drive?

? Foundational Understanding

This phase primarily encompasses the early years of human life, during which an infant endeavour to comprehend a completely new world through his senses & neural systems that have matured over millions of years of human evolution.

Gradually, we construct our foundational understanding of the world by assimilating information from physical senses such as sound, vision, motion, pain, etc. This understanding enables us to perceive obstacles, comprehend the surrounding environment, respond to stimuli like a mother's voice, and coordinate our limbs to navigate without collisions. This is the first version of the world model that we build in our brains.

? Intellectual Understanding

Through the Learning Process, we refine our Foundational Understanding, acquiring information to construct an intellectual grasp of the world. Our initial understanding provides a foundation for knowledge, enabling reasoned evolution. This intellectual understanding evolves through a continuous Learning Process through an acquired sense of opinions (that can be a mixture of supervised, self-supervised, and reinforced). That includes encompassing fundamental concepts like distinguishing between a car and a bike, understanding rules and regulations for walking/driving, etc. This ongoing process, whether conscious or unconscious, persists until we adapt to our growing environment, acquiring essential decision-making capabilities.

? Specialisation

Guided by our interests, beliefs, and values shaped through intellectual understanding, we consciously utilise the Learning Process to refine skills in specific tasks. Whether it's driving a bicycle, cooking, pursuing medicine, or operating a microwave, these specialised tasks evolve through the conscious use of the Learning Process over the years.

First-time drivers often fall in this category, spending a few hours in curriculum-based learning from a “trusted driver”. We use our inherent understanding of the world to tune our decision-making to real-world interactions with a vehicle, including learning to control the vehicle actuators and applying human intuition to adapt to different scenarios. The universality of the Learning Process is evident from the fact that in the past, people used to train and specialise their skills in riding horses and in the future, might adapt to using air taxis as a means of transport. Human adaptability stems from our continuous Learning Process and ability to refine our understanding over and over again.

That makes us a decent driver over time with consistent practice and refinement of our skills.

? Refinement

Humans refine their chosen specialisations through continuous practice, applying the Learning Process persistently to enhance their existing understanding. This constant refinement helps in building confidence in specific tasks - such as learning to drive over a 7–day course, and then practicing in real-life situations, and helps us become confident drivers. Beyond this stage, the progression and pace of improvement become subjective, which results in good, decent, or bad drivers.

Over time, we can transfer the understanding we have developed from driving one vehicle form factor to others, from one geography to another, and adapting to adversarial scenarios. Each new driving experience involves a self-supervised methodology to consistently refine our driving skills.

What’s next?

The contrarian approach to solving for intelligence has led us to Moravec’s paradox3 . While AI is extremely good at specialised tasks with deterministic constraints, it fails miserably at primitive tasks like locomotion in the real world, which even under-evolved species like cats and dogs have mastered. Driving is essentially a derivative of a similar locomotion problem where you are actuating a machine instead. The world we navigate in, not being a rule-based system, requires a generalised understanding of the world and the ability to reason to account for the unpredictability involved.

Though we present a very simplified overview of the intricate operations of the human brain through an approach of developing a Foundational Understanding and refining that through a continuous The Learning Process as the brain is such a complex organ, it's really hard to decipher each and every part of our learning process with an accurate understanding. This is an ongoing research topic for many researchers worldwide but combining neuroscience and AI definitely seems to be a way ahead.

Check out our next article to learn more about how we are taking inspiration from the human brain using our Nature-inspired AI approach.

References:

1. https://thewire.in/science/the-mantis-shrimps-eyesight-is-even-more-extraordinary-than-we-thought
2. https://www.claytex.com/tech-blog/avs-4-measuring-safety-the-99-reliable-fallacy/
3. https://www.aiplusinfo.com/blog/moravecs-paradox-what-is-it-and-what-does-it-mean-for-ai/