ABSTRACT
Before we decide upon the AI approach to be used for our HAI, we need to look into the challenges different methods have to face.
Neurologism: Whole Brain Emulation approach must meet two challenges: (1) Learning enough of the details of human neural networks, which may not happen in the foreseeable future, (2) How to embody the machine after knowledge is uploaded; important because some knowledge becomes knowledge only with associated embodiment. This can be virtual and/or accomplished through some kind of nervous system. Now, having an athlete's brain will not make the lame run fast. Imagine what would happen if you uploaded knowledge of color to a color-blind person.
Logicism: Not all knowledge can be expressed in logic. Most commonsense knowledge probably is not logically expressive. Emotions cannot usually be expressed by mathematical or logical reasoning. We cannot use logic to express a situation where we purposely speak the opposite of the fact for the purpose of amusement. We make irrational choices or feel happy by some feat of our imagination, these are beyond what logicism can accomplish. A human becomes human not only because of his intelligence but also his unintelligence.
Behaviorism: Here the agent's action is mainly a commonsense-based and goal-driven one. The challenges are: (1) How to build a comprehensive commonsense knowledge base for uploading that can grow constantly and still ensure quick responses. Even when such a knowledge base is available and loaded into the agent's brain, we are making a God-like person, not a human. (2) The same embodiment issues as in the approach of neurologism: embodiment requirements, (3) How an agent can identify goals, even when he is in uncharted territory. This is achievable in our HAI architecture, as we'll show later.
Constructivism: This approach has difficulties in the following aspects: (1) Determining the smallest units of actions, say, one inch of movement, the fabrication of a 0.1-second-long sound, etc. (2) How an agent determines the list of action options the agent can take in real time. (3) How an agent can learn complex concepts and responses within a reasonable amount of time.
