Synthetic Intelligence from Zero Principles

Human-level intelligence is still a castle in the sky for any established domain of research. A typical answer to how it could be done comes down to a combination of some current approaches after their significant development and implementing them at a much larger scale, which often means billions of investments and many years of work.

However, another way is possible and, arguably, much more feasible: focusing on the core problems and finding a holistic approach to solving them. In other words, reasoning from first principles, or, if you code, from zero principles.

1. Base hypothesis. Considering the pace of neuroscience development, there is a much higher chance to synthesize a biologically plausible model of intelligence, than to induce it from complete brain analysis.
2. Biological plausibility is not an arbitrary restriction for the creation of Synthetic Intelligence (SI), but the most promising strategy of research. Even though it’s theoretically possible to get other types of general intelligence, it’s much less probable to find other than the evolutionary proven one. At the same time, the problem should be solved at the abstract level of analysis, so biological plausibility doesn’t imply the recreation of the brain-like structure.
3. Mountcastle principal. Just as the neocortex has a quite homogeneous organization and, therefore, the same computational principles for any processing, SI should use the same basic algorithm for all kinds of tasks like visual perception, voluntary movements, language, or anything else.
4. True invariance is the cornerstone of SI, and everything else should be derived from it. SI should be based on a representation that covers all possible modifications and transformations, contrasting to the pseudo-invariant approach to learning based on many different representations of the same pattern.
5. Model of the world (representation of the environment and itself) is an inherent component of SI.
6. Hardware capability. Even the currently available technologies are good enough to get the competitive or even higher performance comparing to the brain, due to a much higher speed of in silico computations in contrast to the slow biological processes, and leveraging algorithms that are not possible in the wetware.
7. Reachability of the SI mostly depends on the starting point and research direction, and with the right ones can be achieved by a single lab or even an individual researcher within a reasonable time.