Autonomous Vehicles 2021: Thoughts on Architecture

Thought:  Autonomous driving fails when a plastic bag flies in front of the car and is recognized as a deer.

One solution that has been touted is a “cheat” by looking at the recorded data and therefore knowing where the bag ends up in the future and basing the reaction/calculated act on that.  This is not possible in the real world - or is it?

How to we achieve this sort of forward in time calculation/reaction? One way is to return to the three screens, or in this case three brains. There is the left brain, calculating based on the existing model. This identifies and reacts to normal stimuli that has been learned from the past. There is the right brain, which continues to adjust the model and is still learning as it goes. Then there is the lizard brain - or right brain reaction. The right brain sees a new situation - something that it has not seen before - and it reacts immediately. The lizard brain reacts based on the immediate best guess from what that object is, what the situation is and what it should do. Then there is the overarching observer, the true right brain of judgement, that figures out after the fact (and during, but slower than initial reaction), does this make sense?  

So there is a part1 (right brain, Master) that is continuously learning and creating the model. There is the part2 (left brain, Emissary) that is employing the existing model. It is algorithmic, has a checklist of things to do and a model to follow. It has no idea what to do when this model fails - no idea. This is where part1 takes over. Part1 can 1) identify something is different 2) cause instant reaction 3)exercise judgment - does this make sense. [Models to look at for thought: System 1 and System 2 - Kahneman, or Master and Emissary - McGilchrist, then Kant after that on Judgement and Reason]

In a computer system this may boil down to three systems however. The judgment of part1, the instant recognition of something new and reaction of part1, and the standard operations checklist and algorithmic work of part2.

The instant recognition of something new and reaction will be the “lizard brain.” It’s new - react. This must could be a sparse matrix (D4M) type setup to see instantly if it has been seen before or is it something new? This system should be a high speed latched processor like the DEC Alpha and not a flip/flop based like most other processor. (Using mutually exclusive clocks to avoid race conditions.) 

Can there be a system similar to the Plastic Cell Architecture with NeuronMOS transistors from NTT for the instant “judgement piece?” Decisions are based on multiple weighted inputs leading to a tanh of sigmoid that rails the decision in each part of the tree very fast (depending on the speed of movement). There could also be a bit of pre-charge logic and gradient descent similar to sense amplifier use in RAM. The rails are pre-charged and whichever drops first determines whether there was a 0 or 1 stored. The interest is speed of instant judgment followed by instant reaction - which can then be tamed by slower judgment. [This pattern of actions are taken in our brain, and in our immune system. The lizard brain reaction of the right brain spots something new and immediately reacts, then the conscious left brain does once it understands, and the right brain takes back over with “does this make sense?” and “when do we return to normal(steady state)?” The immune system is thought to react to a new virus with T-cell production and then toning down or cleaving in the thymus.]