Graph Continuous Thought Machines: A Dispositional Neural Architecture with Simulated Prefrontal Cortex for Adaptive Problem Solving

International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 288 Graph Continuous Thought Machines: A Dispositional Neural Architecture with Simulated Prefrontal Cortex for Adaptive Problem Solving Tofara Moyo∗ ∗Mazusa AI Bulawayo, Zimbabwe Email: [email protected] Panashe Chiurunge† †DeepAnalytics Harare, Zimbabwe Email: [email protected] Abstract—This paper introduces the Graph Continuous Thought Machine (Graph CTM), a novel neuro-inspired computational architecture that emulates biological cognition through dynamic graph-based representations and dispositional neural connectivity. Unlike conventional neural networks with static topologies, Graph CTM employs a three-dimensional dispositional neural tensor from which context-specific subgraphs are instantiated at each processing step (or ”tick”). Each node within this architecture maintains a learnable property vector that encodes both accumulated knowledge and dispositional weights that determine activation probabilities for downstream nodes. Importantly the nodes of the GNN ARE a subset of the neurons in the dispositional neural tensor, instantiating just those that are currently firing. Since the GNN outputs graphs it means that currently firing nodes cause the next nodes by the effect they have on the graph neural network’s (GNN’s) output and so may be seen to be connected to them in some sense. A neural synchronization mechanism dynamically forms and dissolves these connections in this dispositional model, based on activation covariance, effectively implementing Hebbian-like plasticity to minimize prediction loss and building brain like connectivity. The architecture incorporates a simulated prefrontal cortex module that regulates information flow through reinforcement learning and employs harmonic constraints derived from musical consonance in the cyclic group Z12 to determine solution convergence. We formalize the mathematical foundations of Graph CTM, including the synchronization dynamics, dispositional connectivity, and prefrontal regulation mechanisms. Experimental evaluation on the Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI 2) demonstrates the architecture’s capacity for adaptive problem-solving, albeit with limited success compared to human performance. This limitation is theoretically expected given the vast scale disparity: the human brain employs approximately 80 billion neurons, while our implementation utilizes merely 10,000 nodes. Nevertheless, Graph CTM represents a significant step toward biologically plausible AI by modeling how neural synchronization dynamically creates and breaks dispositional connections to optimize information processing, mirroring fundamental mechanisms observed in biological neural systems. Index Terms—Continuous Thought Machines, Graph Neural Networks, Dispositional Representations, Neural Synchronization, Prefrontal Cortex Simulation, ARC-AGI, Reinforcement Learning, Harmonic Constraints, Neuro-inspired Computing I. INTRODUCTION Contemporary artificial intelligence systems predominantly rely on static computational graphs with fixed connectivity patterns, processing discrete inputs to produce discrete outputs through predetermined pathways. This paradigm fundamentally differs from biological cognition, which exhibits continuous, adaptive processing with dynamic neural pathway formation and dissolution based on context, attention, and task requirements. Bridging this gap requires architectures that can model the fluid, context-dependent nature of neural processing while maintaining computational tractability. The Continuous Thought Machine (CTM) framework was introduced as a step toward this goal, featuring uninterrupted processing streams, internal state evolution, and adaptive termination based on confidence metrics. However, conventional CTMs still operate within relatively rigid structural constraints, limiting their capacity to model the dynamic reconfiguration of neural pathways observed in biological systems. In this work, we present the Graph Continuous Thought Machine (Graph CTM), a significant architectural evolution that replaces traditional synapse and neuron models with a graph-based dispositional representation system. The core innovation lies in conceptualizing neural processing as the traversal of a high-dimensional dispositional space, where only contextually relevant portions of this space are instantiated at any given processing step. This approach draws inspiration from neuroscientific evidence suggesting that the brain dynamically configures functional networks rather than relying solely on fixed anatomical connections. Formally, we model cognition as a trajectory through a latent graph space G = (V, E, Φ) , where V represents the complete set of potential neurons (nodes), Edenotes the complete set of potential connections (edges), and Φ : V → R d maps each node to a property vector in R dencoding its dispositional characteristics. At each processing tick t , only a subset V t ⊂ V is active, forming the instantiated subgraph G t = (V t , E t ) that processes the current input in the context of accumulated internal state. Each graph, G t = (V t , E t ), has a subset of the dispositional International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 289 neural models neurons as nodes. These change dynamically as new graphs are derived from the output of the gnn, instantiating different nodes/neurons according to their virtual connections. A central mechanism in Graph CTM is the neural synchronization matrix St ∈ R| V t|×| V t|, computed as the outer product of node activation histories. This matrix dynamically modulates ”weights” and connection strengths between nodes, effectively implementing a form of Hebbian plasticity where ”neurons that fire together wire together.” Crucially, this synchronization process continuously forms and breaks dispositional connections to minimize prediction loss, mirroring synaptic pruning and strengthening observed in biological neural development. The synchronization process also introduces regions of specialization within the dispositional neural model, using its effect on the node property vectors to achieve this. Further inspired by prefrontal cortical function in biological systems, we introduce a meta-regulatory module that governs the exploration-exploitation trade-off during graph space navigation. This simulated prefrontal cortex employs reinforcement learning to optimize the traversal policy π:G → G, determining which nodes should activate at the next tick based on expected reward. Additionally, it implements a novel termination criterion based on harmonic convergence in the cyclic group Z12, where solution states correlate with high consonance among the property vectors of active prefrontal nodes. This induces a symmetry within the dispositional neural models organisation that is exploited by the system for better computation. We evaluate Graph CTM on the challenging ARC-AGI 2 benchmark, which tests abstract reasoning capabilities through visual pattern completion tasks. While our implementation (constrained to 10,000 nodes) achieves modest performance compared to human solvers (who leverage approximately 80 billion neurons), the results validate the architecture’s capacity for adaptive problem-solving through dispositional connectivity. The similarities with the brain are not merely quantitative but qualitative, as human cognition benefits from hierarchical organization across multiple spatial and temporal scales. This quality emerges as a function of the effect attending to the graphs nodes while simultaneously to the input, causing these structures to emerge. For instance simultaneous attention to both input and the gnn would induce aspects akin to short term and long term memory. With the feautures in the environment acting as keys and the parts of the gnn as values.In fact the system designs connectivity and specialization of regions within the dispositional model. The contributions of this work are threefold: 1) We formalize the mathematical foundations of dispositional neural representations as traversals through a latent graph space, with dynamic node instantiation based on context and attention. 2) We introduce a neural synchronization mechanism that forms and dissolves connections between nodes to minimize prediction loss, modeling a key aspect of biological neural plasticity. 3) We implement a simulated prefrontal cortex that regulates information flow through reinforcement learning and harmonic convergence criteria, providing a biologically plausible termination mechanism. The remainder of this paper is structured as follows: Section II reviews related work in neuro-inspired computing and graph neural networks. Section III presents the formal mathematical framework of Graph CTM. Section IV details the architectural components and their interactions. Section V describes our experimental methodology and results on the ARC-AGI 2 corpus. Section VI discusses the biological plausibility of our approach and limitations due to scale constraints. Finally, Section VII concludes and outlines directions for future research. II. RELATED WORK A. Continuous Thought Machines The original Continuous Thought Machine (CTM) architecture introduced the concept of uninterrupted neural processing with adaptive termination. Traditional CTMs process inputs through an encoder, attention mechanism, synapse model (typically a U-Net), and neuron-level models with memory augmentation. A key innovation was the neural synchronization matrix S = ZZ⊤, computed from historical outputs Z, which captures temporal dependencies and contextual relationships. However, conventional CTMs still operate within fixed architectural constraints, limiting their capacity to model the dynamic reorganization of neural pathways observed in biological systems. Graph CTM addresses this limitation by replacing static components with a dispositional graph representation that dynamically instantiates context-relevant subgraphs. B. Graph Neural Networks Graph Neural Networks (GNNs) have emerged as powerful tools for processing relational data. The Graph Convolutional Network (GCN) architecture propagates information through graph structures using spectral filtering operations. Recent advances in attention-based GNNs enable adaptive weighting of neighborhood influences based on node features. Unlike conventional GNNs that operate on fixed input graphs, Graph CTM treats the graph structure itself as a dynamic, learnable representation that evolves during processing. This aligns with recent work on differentiable graph generation but extends it with dispositional representations and neural synchronization dynamics. C. Neural Synchronization and Plasticity Neural synchronization has been extensively studied in computational neuroscience as a mechanism for information binding and communication through coherence. Models of spike-timing-dependent plasticity (STDP) formalize how correlated firing patterns strengthen synaptic connections. Our neural synchronization mechanism draws inspiration from these principles but implements them in a continuous, differentiable framework suitable for gradient-based optimization. The synchronization matrix Stmodulates both virtual International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 290 v v ij i j Σ i k v v t v v i ij v v N(v) j ”weights” and dispositional connectivity, effectively implementing a form of supervised Hebbian learning where connections are strengthened or weakened to minimize prediction loss. D. Prefrontal Cortex Modeling B. Node Processing Dynamics Each active node v ∈ V tprocesses input features through a differentiable function f v : R din → R dout parameterized by its property vector ϕ v: h (t)=fv( x (t), h (t − 1);ϕv) where x (t) is the external input to node v at tick t , h (t − 1) Computational models of prefrontal cortex function (pfc) ofv N (v) ten focus on working memory maintenance, cognitive control, and goal-directed behavior. Reinforcement learning frameworks have been particularly successful in modeling prefrontal function. represents the hidden states of neighboring nodes from the previous tick, and h(t) is the updated hidden state. The node output z (t) is computed as: z(t)=gv(h(t);θv) Our simulated prefrontal cortex extends these approaches by incorporating harmonic constraints as a convergence criterion. This builds upon research linking musical perception to neural synchrony and the role of oscillatory dynamics in cognitive state transitions. Within this model there is a subset of the dispositional nodes that are chosen to be pfc neurons. These nodes learn their property vectors in such a way that if you map the dimensions of their property vectors to a musical keyboard, the property where gvis an output function parameterized by node-specific parameters θv. C. Neural Synchronization Matrix We maintain a history matrix Zv ∈ R d z × Tv for each node v, where T v is the number of ticks since node vwas last activated, and dzis the dimensionality of node outputs. The neural synchronization matrix St∈R|Vt|×|Vt|is computed as: S= σ Σ Z Z ⊤ ! (1) exactly those pfc nodes that form polyphonic harmonization. What this does is that it utilizes the symmetries of the cyclic group Z12 to organize the information within the dispositional model. III. MATHEMATICAL FORMULATION A. Dispositional Neural Tensor We formalize the dispositional neural representation as a three-dimensional tensor T ∈ R N × N × D , where Nrepresents the maximum number of potential nodes along each spatial dimension, and Ddenotes the dimensionality of node property vectors. Each position (i, j, k)in this tensor corresponds to a potential neuron with property vector ϕi,j,k ∈ R D . The property vector ϕi,j,k encodes two critical aspects of dispositional representation: 1) Accumulated knowledge: Features learned from previous experiences 2) Activation disposition: Parameters governing probability of activating downstream nodes (This also induces implied connectivity with the downstream nodes) where σ(·)is a non-linear activation function (e.g., softmax) ensuring proper normalization. St [i, j] represents the synchronization strength between nodes v i and v j, which dynamically modulates attention weights and connection probabilities. This implements a form of Hebbian plasticity where: ∆w ij ∝ St[i, j] · η (2) where ηis a learning rate parameter, and wij is the connection weight between nodes i and j. The synchronization process continuously forms and breaks dispositional connections to minimize the prediction loss Lt. By backpropagating gradients through St, the system learns which connections should be strengthened or weakened given the current context and task requirements. This mirrors biological processes where neural pathways are pruned or reinforced based on their contribution to successful outcomes. D. Graph Propagation with Synchronization-Modulated Attention Information propagation through the graph is governed by a synchronization-modulated attention mechanism. For each node v i ∈Vt, the aggregated message from its neighborhood At processing tick t, a subset of positions Ωt ⊆ { 1, . . . , N } 3 N (v i ) is: m (t) = Σ α(t) · W(t)h(t − 1) (3) vj∈N (vi) V t = { v i,j,k | (i, j, k) ∈ Ω t } where W (t) is a learnable weight matrix, and attention coefficients α (t) are computed as: E t = { (v p , v q ) | v p , v q ∈ V t ,(i p , j p , k p ) − (i q , j q , k q ) 1 ≤ r } (t) exp St [i, j] · LeakyReLU a⊤ [ W(t)h(t) W (t) h (t)] Here, r defines the maximum Manhattan distance for direct α ij = k ∈N (v i ) exp St [i, k] · LeakyReLU a ⊤ [W (t) h (t) W (t) h(t)] connectivity between nodes. (4) vectors describe meaningful chords. It then follows that states of high confidence align with v ∈ Vt is active, defining the current graph instantiation Gt= (Vt, Et) where: International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 291 12 p " Σ i i t v v where ais a learnable attention vector and denotes concatenation. The key innovation is that attention weights are modulated by St[i, j], ensuring that nodes with historically correlated activations receive stronger attention. B. Graph Processing Loop For each subsequent tick t > 0: 1) Synchronization Matrix Update: Compute St Equation 1 based on node activation history. using The updated node representation is then: h (t+1) = σ m (t)+ b (t) (5) 2) Message Passing: Propagate information through the graph using Equations 3-5, with attention weights modwhere σ is a non-linear activation function and b(t) is a bias term. E. Simulated Prefrontal Cortex The simulated prefrontal cortex is implemented as a specialized subset of nodes P ⊂ Vthat regulate the activation of other nodes. These PFC nodes have property vectors constrained to represent musical chords in the cyclic group Z12. Specifically, each dimension of ϕpfor p∈ P is mapped to a pitch class in Z12, and we define a consonance function 3) PFC Evaluation: Compute the consonance C of active PFC nodes and confidence max y P(y | h(t)) to determine if the termination condition in Equation 6 is satisfied. 4) Node Selection: If not terminating, select new nodes to activate using the RL policy in Equation 7. Prune nodes with low activation or minimal contribution to reducing prediction loss. 5) Parameter Update: Update node property vectors and network parameters through backpropagation to minimize prediction loss Lt. C: Z k →[0,1] that measures harmonic coherence. C. Output Generation The termination condition is determined by: Upon termination, the final output is generated by decoding τ t = I C { ϕ (t) mod 12 | p ∈ P t } > γ ∧ max P(y | h (t) ) > th δ e hidden states of active nodes: where y are active PFC nodes at tick , (6) y = D Σ α h(T ) ! δare thresholds for consonance and confidence respectively, and I[·]is an indicator function. The PFC also guides node selection through a reinforcement learning policy π:G → ∆(V)that maps the current graph state to a probability distribution over potential next nodes: π(v next | G t ) ∝ exp (Q(G t , v next ;ψ)/β) (7) where Qis a value function parameterized by ψ, and βis a temperature parameter. The value function is trained to maximize expected reward: v ∈ V T where D is a decoder network, α vare attention weights based on node importance, and T is the termination tick. V. EXPERIMENTS A. Experimental Setup We evaluated Graph CTM on the Abstraction and Reasoning Corpus for Artificial General Intelligence 2 (ARC-AGI 2), a benchmark designed to test fluid intelligence and abstract reasoning capabilities. The dataset comprises 800 tasks, each consisting of 3-5 input-output example pairs and a test input J(ψ) = E τ∼π fi T t=0 r t | τ 0 = 1 # (8) requiring the correct output generation. Tasks involve visual pattern recognition, object counting, spatial reasoning, and other abstract transformations. where r tis the reward at tick t , and T is determined by the termination condition in Equation 6. IV. ARCHITECTURE A. Input Encoding and Initialization At tick t = 0, the input xis encoded through a feature extractor E : X → R d , producing an initial representation e 0 = E(x). A random subset of knodes is activated from the dispositional neural tensor, forming the initial graph G0. The property vectors of these initial nodes are initialized to maximize diversity while maintaining feasibility constraints. It is also important to attend to a n step size history of graphs combined with the current graph in the gnn as well as keep a record of the node output values in order to calculate the values in the synchronization matrix. Our implementation used 10,000 nodes in the dispositional neural tensor, with each node’s property vector having dimensionality D = 64 . The maximum number of active nodes per tick was limited to 500, and the maximum number of ticks per problem was capped at 50. Training was performed using proximal policy optimization (PPO) with a learning rate of 3×10−4. B. Results Table I summarizes our experimental results. Graph CTM achieved a success rate of 0.7% on the ARC-AGI 2 test set, substantially lower than human performance (approximately 80%). However, this outcome was theoretically expected given the enormous disparity between our implementation (10,000 nodes) and the human brain’s approximately 80 billion neurons. ulated by St. P t = P ∩ V t γ and International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 292 TABLE I PERFORMANCE COMPARISON ON ARC-AGI 2 TEST SET Model Success Rate (%) Parameters Human solvers 80.3 ∼80 billion neurons Graph CTM (ours) 0.7 10,000 nodes Current SOTA neural model 23.5 100+ million VI. DISCUSSION A. Biological Plausibility Graph CTM represents a significant step toward biologically plausible AI by modeling several key aspects of neural computation: 1) Dynamic connectivity: Like biological neural systems, Graph CTM dynamically forms and breaks connections between nodes based on their synchronization patterns, rather than relying on fixed architectures. This dispositional connectivity mechanism mirrors how the brain continuously rewires synaptic connections through Hebbian plasticity. 2) Sparse activation: Only a small subset of potential neurons is active at any given time, reflecting the brain’s energy-efficient sparse coding strategy. 3) Hierarchical organization: The dispositional neural tensor provides a latent space from which contextspecific subgraphs are instantiated, analogous to how the brain recruits specialized neural circuits for different cognitive tasks. 4) Meta-cognition: The simulated prefrontal cortex regulates information flow and determines solution convergence, capturing essential aspects of executive function and metacognitive monitoring. The neural synchronization mechanism is particularly noteworthy as it implements a biologically grounded approach to connection formation. In biological systems, neurons that consistently fire together strengthen their synaptic connections through long-term potentiation (LTP), while those with uncorrelated activity undergo synaptic pruning. Graph CTM’s synchronization matrix Stformalizes this process mathematically, allowing the system to dynamically adjust connection strengths to minimize prediction error—precisely as observed in biological neural development and learning. Also as it attends to both the internal and external representations, the system forms long and short term memory. B. Limitations and Scale Considerations The primary limitation of our current implementation is scale. With only 10,000 nodes, Graph CTM cannot match human performance on complex reasoning tasks. The human brain’s 80 billion neurons are organized hierarchically across multiple regions specialized for different functions, with intricate feedback loops and parallel processing pathways. C. Broader Implications Graph CTM offers a novel framework for understanding how dispositional neural representations enable adaptive problem-solving. By formalizing cognition as traversal through a latent graph space, where neural synchronization dynamically shapes connectivity patterns, we provide a mathematical foundation for studying the emergence of intelligent behavior from basic neurocomputational principles. This architecture suggests that intelligence may not primarily depend on the number of neurons or connections, but rather on the flexibility of neural connectivity patterns and the ability to dynamically reconfigure information pathways based on task demands—a principle that could guide the development of more efficient and adaptable AI systems. VII. CONCLUSION We have presented Graph Continuous Thought Machine (Graph CTM), a novel neuro-inspired architecture that models cognition as dynamic traversal through a dispositional graph space. By integrating neural synchronization, reinforcement learning, and harmonic constraints within a graph-based framework, Graph CTM captures essential aspects of biological neural computation while maintaining mathematical rigor and computational tractability. Experimental evaluation on ARC-AGI 2 was poor though this was expected given the large disparity of the human brain and this system in neuron counts. Future work will pursue key enhancements to improve biological plausibility and performance: multi-task training with language modeling objectives,leading to the integration of neuro scientifically-grounded architectural constraints, through specialized prompting to guide the formation of brain-like topological organizations within the dispositional connectivity patterns.This would be optimised by including neuroscience research with the language modeling objectives. REFERENCES [1] L. Darlow, C. Regan, S. Risi, J. Seely, and L. Jones, ”Continuous Thought Machines,” arXiv preprint arXiv:2303.13439, 2023. [2] T. N. Kipf and M. Welling, ”Semi-supervised classification with graph convolutional networks,” in International Conference on Learning Representations (ICLR), 2017. [3] P. Velicˇkovic´, G. Cucurull, A. Casanova, A. Romero, P. Lio`, and Y. Bengio, ”Graph attention networks,” arXiv preprint arXiv:1710.10903, 2018. [4] P. Fries, ”A mechanism for cognitive dynamics: neuronal communication through neuronal coherence,” Trends in Cognitive Sciences, vol. 9, no. 10, pp. 474–480, 2005. [5] H. Markram, J. Lu¨bke, M. Frotscher, and B. Sakmann, ”Regulation of synaptic efficacy by coincidence of postsynaptic APs and EPSPs,” Science, vol. 275, no. 5297, pp. 213–215, 1997. [6] G. Buzsa´ki, Rhythms of the Brain, Oxford University Press, 2006. [7] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, ”Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017. [8] F. Chollet, ”The ARC challenge: Rethinking general intelligence in artificial intelligence,” arXiv preprint arXiv:2105.05480, 2021. [9] K. Friston, ”The free-energy principle: a unified brain theory?” Nature Reviews Neuroscience, vol. 11, no. 2, pp. 127–138, 2010. International Journal of Computer Science Engineering Techniques – Volume 9 Issue 6, November - December - 2025 ISSN: 2455-135X https://www.ijcsejournal.org/ Page 293 [10] S. Dehaene, Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts, Viking Press, 2014. [11] G. Tononi, ”Consciousness as integrated information: a provisional manifesto,” Biological Bulletin, vol. 215, no. 3, pp. 216–242, 2008. [12] D. S. Bassett and O. Sporns, ”Network neuroscience,” Nature Neuroscience, vol. 20, no. 3, pp. 353–364, 2017. [13] O. Sporns, Networks of the Brain, MIT Press, 2010. [14] M. Bar and A. Baram, ”Neural dynamics of the prefrontal cortex in decision making,” Trends in Cognitive Sciences, vol. 23, no. 10, pp. 851–864, 2019. [15] X. J. Wang, ”Synaptic reverberation underlying mnemonic persistent activity,” Trends in Neurosciences, vol. 24, no. 8, pp. 455–463, 2001. [16] J. D. Cohen, M. Botvinick, and D. M. Barch, ”Computational studies of prefrontal development and function,” Developmental Science, vol. 3, no. 4, pp. 401–410, 2000. [17] M. J. Frank, B. Loughry, and R. C. O’Reilly, ”Interactions between frontal cortex and basal ganglia in working memory: a computational model,” Journal of Cognitive Neuroscience, vol. 13, no. 5, pp. 663–686, 2001. [18] P. Dayan and J. J. Yu, ”Decision making and sequential sampling from memory,” Neuron, vol. 58, no. 4, pp. 465–468, 2008. [19] M. Thalmann, M. M. Schlaghecken, and M. A. L. Nicolelis, ”Neural mechanisms of musical perception and memory,” Frontiers in Neuroscience, vol. 12, p. 516, 2018. [20] P. Lakatos, J. Gross, and G. Thut, ”A new unifying account of the roles of neuronal oscillations,” Current Biology, vol. 29, no. 4, pp. R94–R103, 2019. [21] S. Guo, Y. Zhou, and Y. Wang, ”Systematic review of graph neural networks and their applications,” IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 6, pp. 2285–2303, 2020. [22] S. Herculano-Houzel, ”The human brain in numbers: a linearly scaled-up primate brain,” Frontiers in Human Neuroscience, vol. 3, p. 31, 2009. [23] E. K. Miller and J. D. Cohen, ”An integrative theory of prefrontal cortex function,” Annual Review of Neuroscience, vol. 24, no. 1, pp. 167–202, 2001. [24] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, ”Asynchronous methods for deep reinforcement learning,” in International Conference on Machine Learning, PMLR, 2016, pp. 1928–1937. [25] O. Ronneberger, P. Fischer, and T. Brox, ”U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention, Springer, 2015, pp. 234–241. [26] D. E. Scott, ”Mathematical foundations of harmony theory,” Journal of Music Theory, vol. 42, no. 2, pp. 249–271, 1998. [27] D. Tymoczko, A Geometry of Music: Harmony and Counterpoint in the Extended Common Practice, Oxford University Press, 2011. [28] M. Bader, ”Computational music analysis,” Springer Handbook of Systematic Musicology, pp. 277–290, 2019. [29] J. Hawkins and S. Ahmad, ”A thousand brains theory of intelligence,” Frontiers in Neural Circuits, vol. 15, p. 626292, 2021. [30] J. L. McClelland, D. E. Rumelhart, and the PDP Research Group, Parallel Distributed Processing: Explorations in the Microstructure of Cognition, Vol. 2, MIT Press, 1986. [31] D. O. Hebb, The Organization of Behavior: A Neuropsychological Theory, Wiley, 1949. [32] T. J. Sejnowski, ”Synaptic plasticity,” Neural Information Processing Systems, pp. 303–308, 1987. [33] G. Bi and M. Poo, ”Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and postsynaptic cell type,” Journal of Neuroscience, vol. 18, no. 24, pp. 10464–10472, 1998. [34] W. Singer, ”Neuronal synchrony: a versatile code for the definition of relations?” Neuron, vol. 24, no. 1, pp. 49–65, 1999. [35] P. Fries, ”Neuronal gamma-band synchronization as a fundamental process in cortical computation,” Annual Review of Neuroscience, vol. 32, pp. 209–224, 2009. [36] R. J. Douglas and K. A. Martin, ”Neuronal circuits of the neocortex,” Annual Review of Neuroscience, vol. 27, pp. 419–451, 2004. [37] M. Carandini and D. J. Heeger, ”Normalization as a canonical neural computation,” Nature Reviews Neuroscience, vol. 13, no. 1, pp. 51–62, 2012. [38] G. Buzsa´ki and K. Mizuseki, ”The log-dynamic brain: how skewed distributions affect network operations,” Nature Reviews Neuroscience, vol. 15, no. 4, pp. 264–278, 2014. [39] G. E. Hinton, ”Learning distributed representations of concepts,” in Proceedings of the Eighth Annual Conference of the Cognitive Science Society, vol. 1, p. 12, 1986. [40] Y. LeCun, Y. Bengio, and G. Hinton, ”Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015. [41] A. M. Zador, ”A critique of pure learning and what artificial neural networks can learn from animal brains,” Nature Communications, vol. 10, no. 1, p. 3770, 2019. [42] B. A. Richards, T. P. Lillicrap, P. Beaudoin, Y. Bengio, R. Bogacz, A. Christensen, C. Clopath, R. P. Costa, A. de Berker, S. Ganguli, C. C. Gillon, D. F. Hassabis, C. Latham, and K. P. Miller, ”A deep learning framework for neuroscience,” Nature Neuroscience, vol. 22, no. 11, pp. 1761–1770, 2019. [43] J. Schulman, S. Levine, P. Abbeel, M. Jordan, and P. Moritz, ”Trust region policy optimization,” in International Conference on Machine Learning, PMLR, 2015, pp. 1889–1897. [44] J. Schmidhuber, ”Deep learning in neural networks: an overview,” Neural Networks, vol. 61, pp. 85–117, 2015. [45] G. A. Carpenter and S. Grossberg, ”A massively parallel architecture for a self-organizing neural pattern recognition machine,” Computer Vision, Graphics, and Image Processing, vol. 37, no. 1, pp. 54–115, 1987. [46] J. Cohen and R. C. Servan-Schreiber, ”A neural network model of attention and control functions of the prefrontal cortex,” in Modeling the Mind, S. P. Stich and T. A. Warfield, Eds. Oxford University Press, 1995, pp. 13–56. [47] E. K. Miller and P. S. Goldman-Rakic, ”Prefrontal cortical representations and functions in working memory,” Journal of Neuroscience, vol. 14, no. 11, pp. 7369–7385, 1994. [48] K. Doya, K. Samejima, K. Katagiri, and M. Kawato, ”Multiple modelbased reinforcement learning,” Neural Computation, vol. 14, no. 6, pp. 1347–1369, 2002. [49] A. G. Barto and R. S. Sutton, Reinforcement Learning: An Introduction, MIT Press, 1998. [50] M. S. Schroeder, Number Theory in Science and Communication, Springer, 1999. [51] J. J. Bharucha, ”Musical cognition,” in Concepts and Processes in Music, J. L. Pressing, Ed. Oxford University Press, 1988, pp. 140–168. [52] S. Koelsch and W. Siebel, ”Towards a neural basis of processing musical semantics,” Trends in Cognitive Sciences, vol. 9, no. 12, pp. 578–584, 2005. [53] A. S. Bregman, Auditory Scene Analysis: The Perceptual Organization of Sound, MIT Press, 1994. [54] J. A. Fodor, The Mind Doesn’t Work That Way: The Scope and Limits of Computational Psychology, MIT Press, 2000. [55] R. Penrose, Shadows of the Mind: A Search for the Missing Science of Consciousness, Oxford University Press, 1994. [56] A. M. Turing, ”Computing machinery and intelligence,” Mind, vol. 59, no. 236, pp. 433–460, 1950. [57] J. R. Searle, ”Minds, brains, and programs,” Behavioral and Brain Sciences, vol. 3, no. 3, pp. 417–457, 1980.