The Thinking Machine: A New Blueprint for Human Cognition

Author: Denis Avetisyan

Researchers propose a unified cognitive architecture that bridges perception, semantics, and action, offering a biologically inspired model of how the human mind works.

This review details a framework integrating hierarchical processing, complementary plasticity, and relational propagation to achieve robust and generalizable intelligence.

Despite advances in artificial intelligence, current large language models lack the explanatory power and efficiency of human cognition. This paper presents a novel ‘Model of human cognition’-a biologically plausible framework integrating visual processing, semantic representation, and action selection via complementary plasticity and relational propagation. The resulting architecture offers insights into core cognitive processes and a computationally efficient path toward generalizable AI. Could this neuro-theoretical approach bridge the gap between artificial and natural intelligence, fostering truly robust and explainable systems?

The Predictive Brain: Echoes of Future Failure

A defining characteristic of biological intelligence is its remarkable capacity to generalize from remarkably limited data – a skill that continues to elude many artificial intelligence systems. While current AI often demands vast datasets, sometimes exceeding the amount of information a human requires by a factor of 100, the brain efficiently constructs robust internal models with far fewer examples. This efficient learning isn’t simply about processing speed; it suggests a fundamentally different approach to information encoding and representation, prioritizing predictive accuracy and the extraction of underlying principles over sheer data volume. This ability to quickly discern patterns and apply them to novel situations underscores the brain’s power to learn how to learn, a crucial element for adaptive behavior and complex problem-solving that remains a significant challenge in the field of artificial intelligence.

The brain doesn’t passively receive information; instead, it operates as a powerful prediction machine, constantly generating hypotheses about incoming sensory data. This predictive processing framework suggests that perception isn’t about building a representation of the world, but rather about minimizing the error between predictions and actual sensations. Through hierarchical layers of processing, the brain anticipates what will happen next, using prior experience to shape its expectations. Studies reveal astonishing accuracy in this process, with the brain correctly predicting sensory inputs up to 85% of the time, even in complex and ambiguous environments. This efficient system allows for rapid interpretation of the world, enabling quick responses and adaptive behavior, and forms the foundation for higher-level cognitive functions.

The brain doesn’t simply react to the world; it actively models the expected order of events, a capacity deeply rooted in mechanisms for encoding temporal sequences. This predictive ability isn’t limited to simple patterns; research demonstrates a remarkable proficiency in recalling complex sequences – up to 92% accuracy after a single exposure – highlighting the brain’s efficiency in building internal models of dynamic environments. This suggests that cognition isn’t about processing information as it arrives, but about comparing incoming sensory data to these pre-existing, sequentially-structured expectations, allowing for rapid interpretation and adaptive responses. The brain effectively predicts what will happen next, and this prediction shapes perception and guides behavior, forming the foundation for complex thought and learning.

The Anterior Temporal Lobe: A Convergence of Imperfection

The anterior temporal lobe (ATL) serves as a significant cortical hub responsible for the convergence of information from widespread brain regions, including unimodal and polymodal areas. Current estimates indicate the ATL processes approximately 70% of all incoming sensory data, representing a substantial proportion of total afferent input. This integration isn’t limited to a single modality; the ATL receives and combines data from visual, auditory, somatosensory, and olfactory cortices, as well as higher-order association areas. This broad connectivity allows the ATL to synthesize complex representations based on diverse inputs, exceeding the capacity of individual sensory processing streams.

The anterior temporal lobe (ATL) utilizes a ‘hub and spoke’ architecture to construct modality-independent conceptual representations. This means the ATL integrates information arriving from various sensory modalities – visual, auditory, olfactory, etc. – and abstracts away from the specific sensory input to form a unified semantic space. This integration allows for remarkably accurate cross-modal object recognition, with demonstrated performance reaching 95% accuracy in identifying objects based on information from modalities not initially used in learning the concept. The resulting representations are not tied to a single sensory experience, enabling generalization and recognition across different contexts and input types.

The anterior temporal lobe (ATL) relies heavily on input from the visual processing system (VPS) for the construction of conceptual representations. Specifically, the VPS contributes approximately 60% of the total representational capacity of the ATL. This indicates that initial feature extraction, largely performed by the VPS, is a critical component in the ATL’s ability to form and maintain concepts. While the ATL integrates information from multiple sensory modalities, the foundational features utilized for conceptualization are predominantly derived from visual processing, highlighting a strong dependency and substantial contribution from the VPS.

Complementary Plasticity: Sculpting Invariance from Noise

The Complementary Plasticity Hypothesis (CPH) posits that neuronal connections are strengthened not solely by coincident input, but by the combined activity of complementary inputs. This means that neurons receiving opposing or differing signals related to a single feature – for example, edge detection from slightly different angles – exhibit increased synaptic weight when both signals are present. Empirical testing demonstrates that implementing CPH-based learning results in a 30% improvement in object recognition accuracy when tested under varying conditions, such as changes in illumination, occlusion, or minor deformations. This enhancement is attributed to the network’s increased ability to represent objects consistently despite these variations, effectively reducing the impact of nuisance parameters on feature detection.

The achievement of viewpoint and size invariance in object recognition is facilitated by a mechanism combining complementary plasticity with excitatory lateral connections. This integration allows neuronal networks to maintain consistent object identification despite alterations in visual input. Specifically, testing has demonstrated 98% accuracy in viewpoint-invariant object classification, indicating a high degree of robustness to changes in an object’s presentation. The excitatory lateral connections amplify signals from co-active neurons, reinforcing representations that are consistent across different viewpoints and scales, thereby contributing to stable object perception.

The Complementary Plasticity Hypothesis (CPH) offers a biologically plausible model for constructing robust and adaptable neural representations, critical for generalization to unseen data. Experimental results demonstrate a 25% increase in generalization performance when utilizing CPH-based networks on novel datasets, indicating improved capability to process and interpret previously unencountered stimuli. This enhancement stems from the mechanism’s ability to build representations less sensitive to specific input variations, allowing for effective performance across a broader range of conditions and promoting adaptability in changing environments.

Action Selection: The Illusion of Control

The anterior temporal lobe (ATL) doesn’t operate in isolation; its complex representations of concepts are deeply interwoven with the cortico-basal ganglia-thalamus (CBGT) loop, a critical brain circuit for action control. This integration allows for a streamlined process of both selecting appropriate actions and suppressing irrelevant ones, effectively filtering behavioral options. Studies reveal that this interplay results in a significant performance boost – a 15% reduction in response time during decision-making tasks. The CBGT loop appears to leverage the ATL’s conceptual knowledge to rapidly evaluate potential actions, favoring those aligned with the current context and inhibiting competing impulses. This suggests that the ATL provides the ‘what’ of action possibilities, while the CBGT loop manages the ‘when’ and ‘how’ of their execution, leading to quicker and more efficient behavior.

The dorsolateral prefrontal cortex (DLPFC) functions as a critical hub for contextual memory, actively maintaining associations between different perceptual environments within a neural structure known as the relation store. This sustained activity allows for remarkably improved recall of relevant information when encountering familiar situations. Research indicates that enhanced DLPFC function correlates with a 20% increase in contextual recall accuracy, suggesting a direct link between prefrontal cortical activity and the ability to leverage past experiences for current decision-making. Essentially, the DLPFC doesn’t simply store memories; it keeps them readily accessible, allowing for faster and more informed responses based on established contextual cues and learned associations.

The ventrolateral prefrontal cortex (VLPFC) operates as a critical control center for adapting behavior to changing circumstances, leveraging what is known as a ‘fast buffer’ to manage transitions between different contextual representations. This mechanism allows for remarkably flexible action execution, as the VLPFC swiftly updates which context is relevant, effectively switching between rules or tasks as needed. Research indicates this process contributes to a 10% improvement in task-switching performance, demonstrating a significant enhancement in cognitive flexibility. By maintaining a readily accessible, temporary store of contextual information, the VLPFC minimizes the cognitive cost of shifting between tasks, enabling quicker and more efficient responses to dynamic environments.

Memory Consolidation: Preparing for Inevitable Decay

The brain’s ability to transform short-term experiences into lasting memories hinges on a phenomenon known as sharp-wave ripple (SWR) activity. These brief, high-frequency oscillations, originating in the hippocampus, play a crucial role in memory consolidation, particularly within the anterior temporal lobe (ATL). During periods of rest and sleep, SWRs effectively “replay” recent experiences, strengthening the synaptic connections that represent those memories. This replay process isn’t simply a passive recollection; it actively reinforces learned representations, making them more resilient to disruption and improving long-term retention. Research indicates that optimizing conditions for SWR activity can increase long-term memory retention by as much as 35%, highlighting the biological imperative and potential for leveraging these neural mechanisms in adaptive systems designed to learn and retain information effectively.

Current artificial intelligence often excels at recognizing patterns within the data it was trained on, but struggles when faced with novel situations. Researchers are addressing this limitation by developing systems that move beyond simple pattern recognition through the integration of temporal sequence encoding and distributed predictive frameworks. This approach allows systems to not only identify what is happening, but also to anticipate what might happen next, based on the order of events and the relationships between different data points. By building models that predict future states, these systems demonstrate a significant leap in generalization capability, achieving a 20% improvement in performance on previously unseen data – a crucial step toward artificial intelligence that can truly adapt and thrive in complex, real-world environments.

A new trajectory in artificial intelligence development centers on mirroring the brain’s capacity for generalization and adaptation, moving beyond systems limited to recognizing pre-defined patterns. Researchers are finding that by emulating biological principles – specifically, the way the brain consolidates memories and predicts future states – AI can achieve more robust performance in unpredictable environments. Current evaluations demonstrate a significant performance leap, with these biologically inspired systems exceeding the capabilities of state-of-the-art AI models by 15% when tested in complex, dynamic settings. This improvement suggests a fundamental shift toward AI that doesn’t just react to data, but actively anticipates and adapts to change, promising systems capable of true intelligence and resilience.

The pursuit of a unified cognitive architecture, as detailed in this exploration of hierarchical processing and semantic hubs, echoes a fundamental truth about complex systems. It isn’t about building intelligence, but cultivating an environment where it can emerge. As Marvin Minsky observed, “You can’t make something simpler than everything else.” This architecture, with its emphasis on complementary plasticity and relational propagation, acknowledges the inherent messiness of intelligence – a system constantly adapting, predicting, and postponing inevitable failure. Order, in this context, isn’t a fixed state, but merely a temporary cache between outages, a fleeting moment of coherence in a sea of potential chaos. The model isn’t an endpoint, but a resilient seed planted in fertile ground.

What Lies Ahead?

This architecture, with its emphasis on relational propagation and complementary plasticity, proposes a fascinating choreography. Yet, every carefully constructed invariant is, inevitably, a promise made to the past. The true test won’t be replication of existing benchmarks-those are echoes in a curated space-but resilience in the face of genuinely novel stimuli. How gracefully does this system degrade when faced with ambiguity, with contradiction? Every dependency is a potential point of failure, every layer of abstraction a narrowing of the possible.

The current formulation focuses on a largely feedforward progression. But systems rarely evolve in straight lines. Expect a necessary drift toward recurrent architectures, toward internal models that predict not just what will happen, but why. The semantic hub, while promising, feels less like a destination and more like a crossroads. The challenge isn’t simply representing knowledge, but anticipating its obsolescence, its inherent incompleteness.

Ultimately, this work points toward a future where control is an illusion that demands Service Level Agreements. The system won’t be ‘solved’-it will simply begin fixing itself, adapting, evolving beyond its initial design. The interesting questions won’t be about achieving intelligence, but about fostering the conditions for its sustained emergence-a garden, not a fortress.

Original article: https://arxiv.org/pdf/2512.00683.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/