Theory of Neuronal Group Selection
Potentially moving AI substantially closer to consciousness
The Theory of Neuronal Group Selection, or TNGS, is Gerald Edelman’s biological theory of how the brain develops and organizes perception, learning, and adaptive behavior. It is also commonly called Neural Darwinism because it applies principles of variation and selection to populations of neurons. Probably the best concise primary source for explaining TNGS is at [ Link ].
TNGS proposes three main processes:
Developmental selection: Brain development produces diverse populations of interconnected neurons, called neuronal groups. The precise neural connections differ among individuals and are not completely specified in advance.
Experiential selection: Experience changes the strength of connections within and among neuronal groups. Connections that contribute successfully to perception and adaptive behavior are strengthened, creating an individual repertoire shaped by learning.
Reentry: Distributed brain regions continuously exchange signals through dense, reciprocal connections. This ongoing bidirectional signaling coordinates perception, memory, value, and action without requiring a central controller.
Edelman distinguished neuronal selection from ordinary Darwinian natural selection. Variation among neural circuits provides a repertoire from which particular patterns are strengthened through development and experience. Edelman’s published summaries identify developmental selection, experiential selection, and reentry as the central principles of Neural Darwinism. (Edelman, Gally, and Baars, “Biology of Consciousness,” Frontiers in Psychology; Anil Seth, “Darwin’s Neuroscientist,” Frontiers in Psychology.) [ Link ]
TNGS is relevant to AI because it describes intelligence as the product of continuing selection and coordination within a developing system rather than the execution of a completely specified program.
From this perspective, an intelligent system develops its own perceptual categories through interaction with its environment. The meaning of an object or event depends partly on the system’s prior experiences, internal organization, and behavioral needs.
This approach is particularly relevant to embodied AI and neurorobotics. It suggests that some capacities associated with consciousness may require a machine to acquire its internal organization through perception and action rather than only through offline training on a preassembled dataset.
Extended TNGS: From Learning to Primary Consciousness
Extended TNGS applies the “selectional” principles of TNGS, meaning that the brain develops and adapts by selecting among pre-existing variations in neuronal groups and their connections, to the emergence of primary consciousness. Edelman used the term primary consciousness for awareness of a present perceptual scene without necessarily including language, reflective self-consciousness, or an explicit conception of past and future.
According to this extension, primary consciousness emerges when present perception is interpreted in relation to the system’s prior experience and internal values. Edelman called the resulting experience the remembered present. (see Gerald Edelman, “Naturalizing Consciousness: A Theoretical Framework,” Proceedings of the National Academy of Sciences; Edelman, Gally, and Baars, “Biology of Consciousness.”) [ Link ]
Why this is relevant:
Extended TNGS is directly relevant to the possibility of AI consciousness because it describes a possible architecture through which perception, memory, internal value, and action could be combined into a unified, continually updated representation of the present.
Edelman and his colleagues developed a series of physical brain-based devices, often called the Darwin automata, to test aspects of Neural Darwinism in embodied machines.
For example, Darwin VII was controlled by a simulated nervous system modeled on cortical and subcortical organization. It learned through physical interaction with its environment and demonstrated perceptual categorization and value-dependent behavior.
The machines served as experimental models for investigating how selectional neural architectures could generate adaptive behavior. (Krichmar and Edelman, “Machine Psychology: Autonomous Behavior, Perceptual Categorization and Conditioning in a Brain-Based Device,” Cerebral Cortex; Krichmar and Edelman, “Brain-Based Devices for the Study of Nervous Systems and the Development of Intelligent Machines,” Artificial Life.) [ Link ]
Implications for AI:
TNGS and Extended TNGS suggest that a plausible candidate for machine consciousness might require:
a diverse repertoire of initially variable processing networks;
continuing selection and reorganization based on experience;
embodiment in a physical or sufficiently rich simulated environment;
active perception through coordinated sensory and motor behavior;
internal value systems that make some events more significant than others;
memory that influences the interpretation of present events;
dense, reciprocal communication among specialized subsystems;
moment-to-moment integration of perception, memory, value, and action;
a dynamic rather than permanently fixed internal architecture.
This framework differs substantially from the dominant approach to current generative AI. A large language model (LLM) is generally trained by adjusting numerical parameters to predict patterns in data. It may produce sophisticated descriptions of perception, memory, emotion, or self-awareness, but this does not mean that it possesses the embodied developmental history, autonomous value system, sensory-motor engagement, or dynamic reentrant organization proposed by Extended TNGS.
The theory therefore suggests that moving AI substantially closer to consciousness might involve embodied systems that learn continuously, develop their own perceptual categories, evaluate events according to internal needs or values, and integrate those processes through recurrent interaction among multiple specialized systems.
The Darwin automata are important because they demonstrate that parts of this approach can be implemented in machines. They provide evidence that physical devices governed by selectional neural models can develop perceptual categories, learn associations, and adapt their behavior. They do not demonstrate that such machines are conscious.
Extended TNGS therefore supplies both an engineering direction and a warning about evidence. A machine built according to its principles might be a stronger candidate for consciousness than a system that merely imitates conscious language.
Even so, implementing the proposed architecture would establish only that the machine possesses mechanisms associated with consciousness under the theory. It would not conclusively prove that the machine has subjective experience.