Название: A Geometric Approach to the Unification of Symbolic Structures and Neural Networks Автор: Tiansi Dong
Издательство: Springer
Год: 2020 (2021 Edition)
Страниц: 155
Язык: английский
Формат: pdf (true), epub
Размер: 29.2 MB
The methodology in the research of Artificial Intelligence (AI) consists of two competing paradigms, namely symbolic approach and connectionist approach. The symbolic approach is based on symbolic structures and rules, in which thinking is reviewed as symbolic manipulation. Associated with this paradigm are features such as logical, serial, discrete, localized, left-brained. The connectionist approach is inspired by the physiology of the mind, in which thinking is reviewed as information fusion and transfer of a large network of neurons. Associated with this paradigm are features such as analogical, parallel, continuous, distributed, right-brained.
Neural-networks (Deep learning) are robust to noisy inputs, able to learn at a level of approximation from enough high-qualified data, but lack of explainability. In contrast, symbolic systems have a set of manually designed rules. This makes outputs explainable and the results guaranteed, but inputs are not robust to noisy inputs. Do the two kinds of systems talk about the same thing (“human intelligence”)? Neural people hope this, struggle for decades to design elegant neural-networks that can reach symbolic levels of reasoning, and land at a certain level of approximation.
Symbolic approaches deal with (1) the representation and construction of symbols, (2) programs that manipulate these symbols, and (3) descriptions of intelligent behaviors. A typical symbolic system consists of three components as follows:
- symbols: some are primitives, others are constructed using primitive symbols;
- combinatorial semantics: the meaning of constructed symbols can be interpreted by the meaning of primitive symbols and the way of the construction;
- reasoning, which is conducted as symbol manipulation.
Different from symbolic approach, the connectionist approach was initially inspired by the observation that the functioning of the mind is a network of neurons. The connectionist approach views intelligence as activities of interconnected neurons. Connectionists, from the very beginning, have been aiming at understanding how the mind works through the simulation of networks, e.g., how memories are established? how associations are stored for patterns?
Скачать A Geometric Approach to the Unification of Symbolic Structures and Neural Networks