Artificial intelligence (AI) has made much progress, but the existing paradigm for AI is still basically pattern recognition representations, including various neural networks. This work proposes an emergent turing machine (ETM)-based machine thinking model as a new paradigm to address Alan Turing's machine thinking question. The main practical motivation is to fundamentally raise the power of generalization. This is a journal version of <xref ref-type="bibr" rid="ref1">[1]</xref> from which the major conceptual clarification is the skull-closure, which means that the representations inside the closed-skull is totally off-limit to human teachers in the skull-external environment. The machine learning inside the skull, like an animal, is fully autonomous throughout the lifetime during which a grand emergent TM is learnt by automatically integrating many incrementally learnt sub-TMs, as a process called scaffolding in developmental psychology. The model inside the closed-skull is our proposed developmental network (DN) that has been mathematically proven to be optimal in the sense of maximum likelihood. Therefore, the postselection problem in deep learning <xref ref-type="bibr" rid="ref2">[2]</xref> is avoided since we develop only a single network for each life. Experiments in simulated mazes are conducted. With new mechanisms (e.g., <inline-formula><tex-math notation="LaTeX">$Z$</tex-math></inline-formula>-to-<inline-formula><tex-math notation="LaTeX">$Z$</tex-math></inline-formula> connections that learn “Where–What” concepts) added to this journal version, the success rate of disjoint tests increases from 35% to 62.5%, showing a high generalization power of “Where–What” abstraction to new settings. Instead of real-world scenes in our prior work <xref ref-type="bibr" rid="ref3">[3]</xref>, simulations are necessary for this work because they provide arbitrary scenes and ground truths for precise and quantitative error measurements.
Paper
Full text
On Skull-Closed Machine Thinking Based on Emergent Turing Machines
Semantic Scholar · Computer Science · 2024
Abstract
Artificial intelligence (AI) has made much progress, but the existing paradigm for AI is still basically pattern recognition representations, including various neural networks. This work proposes an emergent turing machine (ETM)-based machine thinking model as a new paradigm to address Alan Turing's machine thinking question. The main practical motivation is to fundamentally raise the power of generalization. This is a journal version of <xref ref-type="bibr" rid="ref1">[1]</xref> from which the major conceptual clarification is the skull-closure, which means that the representations inside the closed-skull is totally off-limit to human teachers in the skull-external environment. The machine learning inside the skull, like an animal, is fully autonomous throughout the lifetime during which a grand emergent TM is learnt by automatically integrating many incrementally learnt sub-TMs, as a process called scaffolding in developmental psychology. The model inside the closed-skull is our proposed developmental network (DN) that has been mathematically proven to be optimal in the sense of maximum likelihood. Therefore, the postselection problem in deep learning <xref ref-type="bibr" rid="ref2">[2]</xref> is avoided since we develop only a single network for each life. Experiments in simulated mazes are conducted. With new mechanisms (e.g., <inline-formula><tex-math notation="LaTeX">$Z$</tex-math></inline-formula>-to-<inline-formula><tex-math notation="LaTeX">$Z$</tex-math></inline-formula> connections that learn “Where–What” concepts) added to this journal version, the success rate of disjoint tests increases from 35% to 62.5%, showing a high generalization power of “Where–What” abstraction to new settings. Instead of real-world scenes in our prior work <xref ref-type="bibr" rid="ref3">[3]</xref>, simulations are necessary for this work because they provide arbitrary scenes and ground truths for precise and quantitative error measurements.