well. A Finite Automaton (FA) is a base net for many sophisticated probability-based systems of artificial intelligence, for state-based abstraction. However, an FA processes symbols, instead of images that the brain senses and produces (e.g., sensory images and effector images). This paper informally introduces recent advances along the line of a new type of, brain-anatomy inspired, neural networks —Developmental Networks (DNs). The new theoretical results discussed here include: (1) From any complex FA that demonstrates human knowledge through its sequence of the symbolic inputs-outputs, the Developmental Program (DP) of DN incrementally develops a corresponding DN through the image codes of the symbolic inputs-outputs of the FA. The DN learning from the FA is incremental, immediate and errorfree. (2) After learning the FA, if the DN freezes its learning but runs, it generalizes optimally for infinitely many image inputs and actions based on the embedded inner-product distance, state equivalence, and the principle of maximum likelihood. (3) After learning the FA, if the DN continues to learn and run, it ―thinks‖ optimally in the sense of maximum likelihood based on its past experience. These three theoretical results have also been supported by experimental results using real images and text of natural languages. Together, they seem to argue that the neural networks as a class of methods has passed ―neural networks do not abstract well‖.

Abstract—This is a theoretical, modeling, and algorithmic paper about the spatial aspect of brain-like information processing, modeled by the Developmental Network (DN) model. The new brain architecture allows the external environment (including teachers) to interact with the sensory ends S and the motor ends M of the skull-closed brain B through development. It does not allow the human programmer to hand-pick extra-body concepts or to handcraft the concept boundaries inside the brain B. Mathematically, the brain spatial processing performs real-time mapping from S(t)×B(t)×M(t) to S(t+1)×B(t+1)×M(t+1), through network updates, where the contents of S, B, M all emerge from experience. Using its limited resource, the brain does increasingly better through experience. A new principle is that the effector ends in M serve as hubs for concept learning and abstraction. The effector ends B serve also as input and the sensory ends S serve also as output. As DN embodiments, the Where-What Networks (WWNs) present three major function novelties — new concept abstraction, concept as emergent goals, and goal-directed perception. The WWN series appears to be the first general purpose emergent systems for detecting and recognizing multiple objects in complex backgrounds. Among others, the most significant new mechanism is general-purpose top-down attention. Index Terms—Mental architecture, cortical representation,

Abstract—Although a caregiver tries to be consistent while she teaches a baby, it is not guaranteed that she never makes errors. This situation is also true with a digital game, during which the human player needs to teach a non-player character (NPC). In this work, we report how a teacher can successfully train the Developmental Networks (DNs) while she cannot guarantee an error-free sequence of motor-supervised teaching. We establish that, under certain conditions, a DN tolerates a significant number of errors in a teaching sequence as long as the errors do not overwhelm the correct motor supervisions in terms of the Z-normalized frequency. We also provide theoretical arguments why task-nonspecific agents like DNs create a new dimension for the play values of future digital games. The emergent representations in the DN can not only abstract well like a symbolic representation (e.g., Finite Automaton) but also deal with the problem of exponential complexity with the traditional symbolic representations currently prevailing in the artificial intelligence (AI) field and in the digital gaming field. The experimental results showed that the speed of convergence to correct actions depends on the error rates in training. I.