Thread by @IntuitMachine

There is a fundamental flaw that exists in almost every branch of science that is a consequence of the requirement for objectivity.

The most obvious manifestation of this the need to disentangle parts from their whole to gain understanding. In many complex systems (i.e. biology) this approach of reductionism is problematic.

However, I want to focus on something less obvious. What is less obvious is that there exists a notion of global truth.

We already know from physics that truth can only be measured in a relativistic sense. As an example, a clock on earth ticks more slowly than a clock in space. A measurement in the quantum realm affects the observation of reality.

In deep learning systems are trained on an objective function or a reward function which is based on an underlying assumption that there is a global measure as to what is good (or valuable).

This is problematic from the perspective of a real organism that grows. That is because the information to what is good is unavailable from the perspective of an individual. Only local information is available and at best there is a proxy for global information available.

All living things create themselves from within. But they do so by sampling the environment, making adjustments as they grow. This implies some intrinsic mechanism that drives towards some goal that has no measure.

Said differently, there exists intrinsic behavior that drives towards a goal that the organism itself invents for itself.

Every living thing plays a game that it invents for itself. It is successful when it plays a game that approximates the strategies that prosper in its environment.

This is one of the blind spots of the current generation deep learning. We mistake adversarial inputs for the kinds of misclassification we see in visual illusions. We don't take into account that biological systems look at the world very differently from the objective world.

This is because, organisms have innate sensory and interaction mechanisms that have been fined tuned over billions of years. They are able to act fast and frugal because of the affordances that they are tuned for. They are blind to actual reality.

This is no difference in humans being unaware of the blindspot in their eyes. We hallucinate a continuous visual space but in reality, we actually do not see everything. We fill in the blanks.

We fill in blindspots with what exactly? With the information we habitually expect to see. Thus, in our gaps of perception or in our gaps of awareness, we fill it up using models of intuition. But we don't actually see anything, our expectations delude ourselves.

Every living organism lives in a constructed virtual bubble of expectations of the world. Expectations that have either been forged by evolution or amortized through experience.

How then do we see outside our virtual bubble? We pose new questions and seek out a hypothesis to confirm these questions. But this only works when our questions are not biased from our experiences. But we cannot create a question that is not grounded in experience.

To see beyond our self created bubbles requires a stance that seeks discovery. A person with a passion for photography seeks new creative angles on a scene. Such a person actually discovers more about a setting than the casual visitor.

It is only through active questioning and testing that we see beyond our hallucinations of this world. There is thus no global goal or reward function. There is only the stance towards seeking greater understanding.