Abstract
This paper describes a neuronal device that embodies the computational principles of the "paper-and-pencil" machine envisioned by Alan Turing. Based on cortical organization, the model implements pointers and investigates how neuronal circuits instantiate basic operations like variable assignment (e.g., $x=5$), value comparison, and retrieval. The network's collective function is demonstrated through simplified arithmetic and "blocks-world" problem-solving tasks.
Introduction
While neuroscience has successfully mapped "atomic" operations—such as how individual neurons respond to visual stimuli or how simple decisions are reached—a significant gap remains in understanding how the brain organizes these atoms into complex, multi-step "programs." In computer science, complex tasks are handled by algorithms involving variables, pointers, and loops. However, the brain is a massively parallel, analog system. This research addresses a fundamental question: How can a network of neurons instantiate the symbolic operations of a computer to perform sequential computations?
The Architecture of Sequential Thought
The authors propose a framework based on "production systems," a concept from cognitive architectures like ACT-R. The model relies on three key mechanisms to bridge the gap between parallel neural activity and serial logic.
1. Seriality from Competition
Based on electrophysiological studies of the lateral intraparietal (LIP) area, the model uses ramp-to-threshold dynamics. Neurons accumulate evidence until they hit a threshold, triggering a "production" (an action or state change). While the brain processes information in parallel, seriality emerges because only one production can "win" the competition at a time.
As shown in the figure above, the completion of one diffusion process collapses the distribution into a few symbols, which are then passed to the next step. This chaining allows "if-then" programs to arise from sequential diffusion processes without the need for a central digital clock.
2. Neuronal Pointers and Variable Binding
To implement variables (e.g., $x = 5$), the model uses transient synaptic plasticity. When a "pointer" neuron and a "value" neuron are active simultaneously, their connection is temporarily strengthened. This allows the brain to flexibly re-use the same neuronal territory for different variables across different tasks.
To retrieve a value, the pointer is activated within a competitive network, biasing the system to "recall" the specific value it was bound to through a Winner-Take-All (WTA) competition. This mechanism effectively serves as the "memory" in the neuronal Turing machine.
Computational Benchmarks
The architecture was evaluated using two classic benchmarks:
- Arithmetic Counter: A "while-loop" task where the network counts from an initial number to a target. The model successfully cycled through distinct productions (e.g.,
PrepareNext,IncrementCount,CheckEqual) while maintaining distinct variables for the current and target counts. - Blocks-World: A paradigmatic AI problem requiring the system to move colored blocks to match a goal configuration. The model integrated visual search, attentional cuing, and motor planning to solve the puzzle.
Discussion and Future Directions
The model demonstrates that the brain can manage noise by using thresholding to "clean up" analog evidence into discrete, reliable steps. However, in this study, the production rules were "hand-wired." A major next step is explaining how the brain learns these programs, likely through reinforcement learning. Additionally, while the model is inspired by regions like the LIP and Intra-Parietal Sulcus, the exact biological identity of "production neurons" remains a hypothesis requiring further experimental validation.
Summary updated