Had it not been for the meteor that altered Earth’s fate or the ice age that devoured everything, could small mammals like humans, facing these immensely powerful and intelligent giants, truly have established themselves in this jungle?
However, history is always full of dramatic twists. Humans took the stage and competed with the giants not through brute strength, but through the unique intelligence of tool use. It is this intelligence that has allowed us to evolve from vulnerable survivors to the dominant species on Earth, even holding the potential to realize a stunning concept: resurrecting the long-extinct dinosaurs.
▷ Figure 1. Deep Time Cognition Team Homepage. Source: Deep Time Cognition Official Website
Besides dinosaur enthusiasts, there is a group of researchers interested in the relationship between humans and dinosaurs. They belong to DeepTimeCognition, a research team composed of cognitive ethologists, neuroanatomists, and paleontologists, who are dedicated to exploring the development of cognition from an evolutionary perspective.
▷ Osvath M, Němec P, Brusatte SL, Witmer LM. Thought for food: the endothermic brain hypothesis. Trends Cogn Sci. 2024;28(11):998-1010. doi:10.1016/j.tics.2024.08.002.
In November 2024, the team published a study in the journal Trends in Cognitive Sciences, exploring the relationship between “whole-body endothermy” and the evolution of “consciousness.” They observed that dinosaurs and mammals independently evolved endothermic traits, and alongside this change, alterations in neural cognitive structures occurred: the number of neurons surged to twenty times the original number, and new brain structures simultaneously evolved. Through substantial evidence, the research team demonstrated a profound connection between whole-body endothermy and the development of cognitive abilities, highlighting the unique perspectives and significance of studying the brain from the viewpoints of evolutionary biology and comparative biology.
1. How Did Whole-Body Endothermy Evolve?
What exactly is “whole-body endothermy”? It is a physiological trait that enables endothermic animals to maintain a stable body temperature through internal metabolism, without relying on external environmental temperatures. This characteristic requires extremely high metabolic support, with birds and mammals being the most typical representatives among extant species. In terms of neural cognitive functions, birds exhibit astonishing similarities to mammals. This resemblance does not exist in reptiles, despite birds being more closely related to reptiles from an evolutionary perspective.
It can be said that the transition from ectothermy to endothermy is one of the most dramatic changes in the evolutionary history of the brain. As the only two groups of tachymetabolic endotherms on Earth, birds and mammals likely share these cognitive traits because they are essential for supporting their high-energy metabolic lifestyles. This commonality provides important clues for our understanding of the development of cognitive abilities.
▷ Figure 2. Simplified Evolutionary Diagram and Research Methodology. Source: Trends in Cognitive Sciences
2. What is the Relationship Between Cognition and Whole-Body Endothermy?
“Whole-body endothermy” pertains to energy consumption and biological metabolism, while “cognition” is a core function of the brain. Why are these two seemingly unrelated traits closely linked in evolution? The mere co-occurrence of “whole-body endothermy” and “cognition” may not be sufficiently convincing.
To sustain themselves, organisms must convert external energy into energy that supports their cellular activities and structures; this network of chemical activities is collectively known as “metabolism.” In addition, to prevent collapse, organisms seek out the substances required for metabolism while avoiding potential harm. This balance is primarily maintained through allostasis, which refers to the regulation of stability through change. Allostasis involves not only passively responding to changes but also actively predicting and coping with disturbances. Cognition is a homeostatic regulation system closely connected to an organism’s environment. Therefore, it is believed to be closely related to metabolism [1]:
“Cognition consists of sensory and other information processing mechanisms through which organisms become familiar with, evaluate, and interact effectively with environmental features to meet their survival needs. The most fundamental needs are survival/maintenance, growth/prosperity, and reproduction.”
From this perspective, the nervous system can be viewed as an evolutionary extension for maintaining the organism’s basic allostasis. The emergence of the brain enables organisms to move more quickly and over longer distances, thereby accessing more information and better storing this information in memory.
In this light, the brain is extremely important in energy intake—does this mean that a more developed brain is always better? Such a notion may overlook the “investment” cost associated with brain development. The “expensive brain hypothesis” posits that having a larger brain requires either an increase in total energy intake or a reduction in other high-energy-consuming bodily activities to compensate. In fact, in environments with lower energy availability or highly fluctuating conditions, animals typically have smaller brains, meaning these animals “choose” smaller brains to reduce the “expense” of brain maintenance.
▷ Figure 3. Evolution of Brain Size in Carnivores. Source: Nature
However, when comparing ectothermic and endothermic animals, this linear relationship in brain size does not hold. In endothermic animals, the number of high-energy neurons increases exponentially—representing a change in the cost/benefit ratio of neural tissue. An interesting observation is that in endothermic animals, having a larger brain often correlates with longer lifespans, whereas in ectothermic amniotes, the opposite is true. This implies that in reptiles, investment in neural tissue diverts energy from other bodily activities, accelerating the aging process.
Endothermy fundamentally changes this relationship: the abundance of high-energy neurons not only does not compete with other bodily activities but also supports the maintenance of these activities. For endothermic animals, brain development is a high-risk but higher-reward “investment”: possessing a more neuron-rich and functionally flexible brain incurs energy costs that are negligible compared to the enhancements in brain functionality.
Moreover, the brains of endothermic animals play a crucial role in parental care. Compared to ectothermic animals, endothermic animals exhibit more common and advanced parental care behaviors. The essence of this phenomenon lies in the fact that the energy acquired by endothermic animals not only supports their own bodily activities and brain functions but also meets the needs of other individuals.
3. The Evolutionary Battle Between Brain Development and Energy Consumption: Who Prevails?
By comparing endothermic birds with ectothermic amniotes, we have come to understand the close relationship between “whole-body endothermy” and “cognition.” But how does a qualitative change like brain function occur?
First, we need to identify the differences in cognitive functions supported by these two types of brains and understand the reasons behind these differences.
With the emergence of endothermy, life entered the “age of gluttony.” Prior to this, animals had never consumed as much food as endothermic species do. Existing endothermic animals have food intake (by weight) that is 8 times (mammals) to 11 times (birds) that of equally sized active reptiles [2].
While ectothermic animals also require food to sustain life, grow, develop, and reproduce, endothermic animals additionally depend on food to fuel their high metabolism and generate body heat. Ectothermic animals can survive for extended periods by relying on external heat sources without food, and their lower dependence on food allows them to further conserve energy by reducing foraging activities [3].
▷ Figure 4. Box Turtles in Hibernation. Source: Boxturtles.com
Based on these differences, it can be inferred that the evolution of cognitive functions in endothermic animals primarily serves the purpose of obtaining food. Their high metabolism enables them to forage over greater distances and cope with fluctuations in environmental temperatures. However, high metabolism necessitates such energy-intensive foraging behavior.
At the core of foraging lies an eternal dilemma faced by all mobile organisms: whether to fully exploit existing resources or to explore potentially better resources. This issue is so fundamental that the brain originally evolved to solve it and has since continuously developed in complexity around this theme [4].
In model-free exploration, animals guide their behavior through random searching (assigning equal value to any exploratory action) and stimulus-response associations that rely on habits, occasionally stumbling upon valuable information [5].
In contrast, model-based exploration relies on the animal’s cognitive model of the environment and its goal-directedness. It involves simulating various actions within an environment-based psychological model derived from past experiences, thereby evaluating options and predicting outcomes.
Thus, exploratory foraging transforms from a blind, unknown adventure into a continually optimized search process for energy intake.
4. How Does Endothermy Shape Neural Cognition?
The observed neurocognitive adaptations in birds and mammals support the hypothesis that endothermic brains have evolved primarily to achieve efficient model-based exploration, addressing the energy demands of endothermic animals.
This evolutionary development is manifested in several ways: both birds and mammals have independently evolved larger relative brain sizes [6] and higher neuron densities [7, 8], resulting in significantly enhanced information processing and memory capacities. Compared to reptiles of equivalent body weight, birds and mammals have, on average, 17 times and 9 times more neurons in the telencephalon, respectively, and 45 times and 69 times more neurons in the cerebellum [8].
The expansion of the mammalian telencephalon is closely associated with the emergence of the six-layer neocortex. For example, visual information is organized in a topological and hierarchical manner, forming retinotopic coding. Perceptual information from different sensory modalities is integrated in association areas to create functional modules, such as the limbic system, executive modules, and premotor modules. Similar features are also present in the telencephalon of birds, where the hyperpallium contains multiple topological maps of visual fields [9]. This structure enables rapid and detailed visual analysis but requires more neurons than the non-retinotopic coding found in the visual cortices of reptiles [10]. In both birds and mammals, modular brain regions have become increasingly complex.
Additionally, birds and mammals have convergently evolved complex behavior coordination areas in their telencephalon—the prefrontal cortex (PFC) in mammals and the nidopallium caudolaterale (NCL) in birds [11]. Although these regions are not homologous anatomically, they are highly similar functionally. Both the PFC and NCL receive input from all sensory modalities, spatial and memory information from the hippocampus, and internal state information from the limbic system. They relay this information to the basal ganglia and motor (premotor) structures, thereby mediating goal-directed actions [12], such as inhibiting behaviors and managing working memory as part of a range of executive functions.
Beyond cortical areas, the evolution of the hippocampus in birds and mammals has also become more convergent. While the hippocampus of most vertebrates can create allocentric spatial maps centered on external environments [13], the structure and connectivity of the hippocampus in birds and mammals are evidently more complex. Their enlarged hippocampi receive preprocessed information from the pallium/cortex and possess rich intrinsic axonal collaterals, allowing them to remember numerous locations without extensive training. Birds and mammals may possess integrated cognitive maps, including valence maps associated with specific locations, to enable faster and more efficient memory encoding [14].
In addition to similarities in information organization and encoding, the brains of birds and mammals have independently evolved more complex basal ganglia to achieve finer control of behavioral patterns [15]. The number of neurons in the basal ganglia of birds alone can reach or even exceed the total number of neurons in the entire telencephalon of reptiles of comparable size [7]. Furthermore, to exert more direct control over movement, both birds and mammals have evolved central and downstream excitatory neural pathways in the brain, which are crucial for dexterous motor skills such as food manipulation and extractive foraging [16].
The cerebellum is primarily associated with motor control and the predictive computations required for learning [17]. Amphibians and squamate reptiles have smaller cerebellums with fewer neurons, while crocodilians (turtles and archosaurs) and lepidosaurs have increased cerebellum volume and neuron counts, with birds and mammals exhibiting the most significant expansions [18]. Additionally, birds and mammals have evolved telencephalon-cerebellum pathways, whose development is related to the size of the telencephalon. Although the expansion of the cerebellum in endothermic animals was initially driven by complex terrestrial locomotion and upright posture to facilitate efficient respiration during sustained movement, increasing evidence suggests that the cerebellum plays an important role in various cognitive functions.
The demand for effective foraging has driven the evolution of these functionally similar yet independently evolved brain structures. Beyond more direct sensory-motor control, birds and mammals have also developed larger and more detailed cognitive maps. A key feature of the newly evolved neocortex/cerebellum system is “offline” sensory simulation, allowing animals to evaluate different options through hippocampal cognitive maps before taking action.
▷ Figure 5. Cells Associated with Hippocampal Cognitive Map Functions. Source: Neuron
Furthermore, the basal ganglia cannot distinguish between perceptual simulations and reality—this may prompt animals to make action decisions based on more energy-efficient simulated scenarios. The prefrontal cortex (PFC) and nidopallium caudolaterale (NCL) play crucial roles in balancing outcome-oriented memory with habitual behavior, deciding whether to inhibit instinctual tendencies or replace old experiences with new knowledge. These top-down processes help make more informed decisions and are believed to have partially evolved to increase caloric intake.
5. How to Deeply Understand Neural Cognition Through the Dimension of Time
The Deep Time Cognition team has combined differences in the survival strategies of ectothermic and endothermic animals with neuroanatomical evidence, uncovering the simultaneous evolutionary emergence of the cognitive revolution and endothermy. Through cognitive neuroscience research, they have elucidated the evolutionary processes of the brain and the mechanisms of information perception. For an in-depth study of the information processing methods of ectothermic and endothermic animals, they recommend conducting comparative neuroscience research. This research should focus on species located at critical nodes of system development.
▷ Figure 6. Workflow for Deep Temporal and Spatial Interpretation of Neural Cognition. Source: Trends in Cognitive Sciences
Among archosaurs, all species except endothermic birds have gone extinct. However, within reptiles, numerous extant species remain, including ectothermic animals (lizards, snakes, tuataras, turtles, and crocodilians) and endothermic animals (birds). In reptiles, endothermy likely appeared within the Archosauria clade, whose current representative species include crocodilians (ectothermic) and birds (endothermic). Crocodilians largely retain the neuroanatomical structures of ancient archosaurs. These ancestral features are shared by crocodilians, non-avian dinosaurs, and birds.
Palaeognaths represent the most primitive birds discovered in neuroanatomical studies and share many similarities with non-avian paravians (Paravians). They also have the lowest metabolic rates among birds, which may reflect their primitive energy regulation methods. By comparing the cognitive abilities of crocodilians with those of palaeognaths, researchers can approximate the evolutionary transition of ancient archosaurs from ectothermy to early, fully high-level metabolic endothermy. Although research in this area is still limited, existing findings suggest that palaeognaths already possess robust social cognitive abilities.
Studying other ectothermic reptiles beyond archosaurs is also important. Neurocognitive adaptive changes that appeared even earlier in some archosaurs may have facilitated the emergence of endothermy. For example, specialized structures in crocodilians, such as the rudimentary nidopallium caudolaterale (NCL) and projections between the forebrain and cerebellum, play significant roles in endothermic animals. It is necessary to determine whether these structures are unique to ancient archosaurs.
6. Conclusion
The progressively increasing energy demands throughout evolutionary history are reflected in cognitive and neuroanatomical structures. This shift aligns with the trend from ectothermy to endothermy and has occurred independently along the evolutionary paths of birds and mammals. Cognitive abilities support the survival of organisms, while elevated metabolic rates promote an increase in neuron numbers and the emergence of new brain structures. These new structures, in turn, enhance cognitive efficiency. In essence, how animals “eat” and how they “think” are intricately interconnected processes.
The Deep Time Cognition team combined differences in the survival strategies of ectothermic and endothermic animals with neuroanatomical evidence, uncovering the simultaneous evolutionary emergence of the cognitive revolution and endothermy. Through cognitive neuroscience research, they revealed the evolutionary processes of the brain and the mechanisms of information perception.
Tracing the evolution of the brain over geological timescales not only helps us explore the nature of “intelligence” but also prompts us to reassess our relationships with other species in the natural world. To better understand this cognitive shift, comprehensive comparative studies of cognition and neuroanatomy are necessary for extant ectothermic and endothermic animals located at critical points in system development. Although current research tasks are challenging and many core questions remain unresolved, the time is ripe for conducting such studies.
Questions Yet to Be Addressed:
• What are the primary cognitive differences between existing ectothermic and endothermic amniotes?
• How are cognitive abilities and brain anatomical structures related within and between ectothermic and endothermic animals?
• Can we identify brain region correlations unique to endothermic animals in extinct species?
• How have cognitive abilities gradually evolved during the process of endothermy?
In 2009, Swiss neuroscientist Henry Markram introduced an idea that stunned the scientific community: fully simulating the human brain—with its 86 billion neurons and 100 trillion synapses—using a computer.
▷ Figure 1. The Virtual Brain supports computational modeling of the brain at a network scale. Source: Jan Paul Triebkorn
If this “crazy” idea were realized, it could not only open up new avenues for treating brain disorders such as Alzheimer’s disease but also provide invaluable biological insights for the development of intelligent robots.
In 2013, with the backing of the European Union and other supporters, this ambitious concept finally became a reality when the Human Brain Project (HBP) was officially launched.
▷ Figure 2. Human Brain Project (HBP). Source: HBP official website
This grand scientific endeavor—comparable in scope to last century’s moon landing project—shifted its focus to an even more enigmatic realm: the “inner universe” within our skulls. Aiming to understand the complexity of the human brain and advance the treatment of neurological diseases, it boldly confronted various challenges in neuroscience, vowing to unlock the ultimate mysteries of the human brain.
Now, a decade later, in September 2024, an independent report comprehensively summarized the progress of the HBP in advancing brain science research, simulating the brain, and translating its findings into practical applications [1].
What breakthroughs has this project, which began as a scientist’s “wild dream,” achieved? What limitations has it encountered along the way? And in today’s era of rapidly advancing artificial intelligence, what lessons can we learn from the HBP’s experiences? Let us look back at the ten-year journey of this “grand scientific project.”
HBP Data Overview
The Human Brain Project (HBP) is a large-scale research initiative launched by the European Union in 2013. Its goal is to advance neuroscience, computer science, and medicine by simulating and understanding the complex functions of the human brain.
At that time, the global scientific community showed immense interest in brain science, computing technology, artificial intelligence, and their potential medical applications. Although previous studies had achieved significant progress, there was still a lack of integrated models—spanning multiple scales and dimensions—that could comprehensively explain how the brain operates at the biological, chemical, and physical levels. In this context, the HBP aimed to address key challenges in brain research—particularly the etiology of brain diseases and the operational mechanisms of neural networks—by leveraging neuroscience, computational science, and interdisciplinary collaboration.
▷ Figure 3. HBP Data Overview. Source: HBP official website
Over the past decade, the HBP has accomplished a series of breakthroughs in advancing brain science research and technological applications. To date, more than 500 scientists from 155 research institutions across 19 countries have participated in the HBP. Supported by this project, participating scientists have published over 3,000 papers, secured 92 patents, and incubated 12 companies. The EBRAINS database developed by the HBP is currently used by tens of thousands of scientists from nearly 1,500 research institutions.
▷ Figure 4. Facts Related to the EBRAINS Database. Source: HBP official website
Main Research Areas of the HBP
The EU Human Brain Project (HBP) has garnered worldwide attention and is regarded as comparable to the U.S. BRAIN Initiative. However, the two initiatives differ fundamentally in their research directions: while the U.S. BRAIN Initiative focuses on the development and application of neurotechnologies, the EU’s HBP places greater emphasis on brain simulation and the development of the information and communication technologies essential for such research.
▷ Figure 5. Fenix builds various EBRAINS platform services, enabling both the neuroscience community and other research fields to advance their work. Source: HBP official website
The primary research areas of the HBP include the following:
(1) Neurodegenerative Disease Models: By employing advanced simulation techniques, researchers have developed models to study neurodegenerative conditions such as Alzheimer’s disease, Parkinson’s disease, and depression. These models offer new perspectives for diagnosing and treating these disorders.
(2) Brain Simulation and Computation: A core objective of the HBP is to create a comprehensive model of the brain. By leveraging large-scale computational platforms, researchers can simulate neuronal networks and investigate the brain’s functional dynamics and disease mechanisms. In constructing digital brain models, the HBP aids scientists in unraveling the complexities of brain function.
▷ Figure 6. One of the primary goals of the EU Human Brain Project is to design and implement six computational platforms.
(3) Advancements at the Intersection of Brain Science and AI: The HBP has not only propelled advancements in neuroscience but has also significantly influenced artificial intelligence research. By simulating biological neural networks, the project has introduced innovative insights into neural network technologies. The outcomes of brain science research under the HBP have enabled AI systems to handle perception, learning, and memory tasks more efficiently.
(4) Brain Data Sharing Platform: The HBP has developed an open-access platform named “EBRAINS,” which allows brain science researchers from around the globe to share data, tools, and computational resources. This collaborative infrastructure has fostered international scientific cooperation and accelerated progress in brain disease research.
Leading Scientists and Notable Achievements
To achieve these ambitious research objectives, the HBP has assembled leading experts from diverse fields, including neuroscience, computer science, and medicine. Over 500 scientists from various disciplines have participated in the project. Using functional simulations, computational methods, and advanced imaging technologies, these researchers have enhanced our systematic understanding of brain function, particularly in simulating and predicting neurodegenerative diseases. European and American scientists involved in the HBP have primarily focused on large-scale neural simulations, sharing and analyzing neural data, and digitally studying neurological diseases.
Below are some representative achievements:
Henry Markram
École Polytechnique Fédérale de Lausanne, Switzerland
As one of the founders of the Human Brain Project (HBP), Markram is devoted to creating a whole-brain simulation platform to investigate neural functions across different scales. The “Blue Brain Project,” which he initiated in 2005, provided valuable insights and experience that helped launch and advance the HBP.
Karl Friston
University College London, United Kingdom
Friston proposed the brain’s predictive coding theory, enabling researchers to better understand the mechanisms behind the brain’s signal processing.
Michael Hausser
University College London, United Kingdom
By developing cutting-edge optogenetic and neuronal recording techniques, Hausser’s team has explored the relationship between synaptic connectivity and brain function, thereby laying the groundwork for mapping the brain’s structure and function.
Alain Destexhe
Saclay University, Paris, France
Specializing in the study of the brain’s dynamic functions, Destexhe’s research focuses on the mechanisms behind the generation of brain waves and neuronal synchronization. His work has provided crucial support for deciphering neural signals.
Terrence Sejnowski
University of California, San Diego, USA
As a member of the HBP Advisory Board, Sejnowski concentrates on the computational functions of the brain and machine learning technologies, with a particular emphasis on enhancing artificial intelligence systems through brain modeling.
Christof Koch
Allen Institute for Brain Science, USA
At the Allen Institute, Koch and his team have developed a high-resolution brain atlas, achieving a more precise structural model of the human brain.
Participating Research Institutions and Budget
The HBP is a European-led initiative, supported by funding from the European Union’s Horizon 2020 research program. It involves research institutions, universities, and companies from multiple European countries, with key participants including Germany, Switzerland, France, the United Kingdom, Spain, and Italy. Major scientific institutions include European neuroscience research centers (such as the Max Planck Institute in Germany) and leading European universities (such as ETH Zurich and Paris VI University). Multinational companies like Intel and IBM have provided technical support for the development of computational platforms. In addition, scientists and institutions from non-EU countries—such as those from the United States and China—have also participated in collaborative projects.
The initially planned budget for the HBP was 1 billion euros, with the majority of the funding provided by the EU and supplemented by contributions from partner countries. The budget was allocated in phases, with approximately 500 million euros dedicated to the first five years (2013–2018). Subsequently, the project entered its second phase, securing hundreds of millions of euros in additional support. Over a ten-year period, the total budget may eventually reach roughly 1.5 billion euros, covering personnel costs, the construction of computing platforms, infrastructure investments, and research expenditures.
Other Brain Science Projects Similar to the HBP
There are several other research projects worldwide dedicated to in-depth studies of the brain. These initiatives have achieved significant results in advancing neuroscience, brain disease research, and artificial intelligence. Although they differ in budget, objectives, and technological approaches, they collectively contribute to a deeper understanding of brain function and its complex networks.
Project Achievements and Technological Innovations of the HBP
As a milestone project in humanity’s exploration of brain science, the HBP has undoubtedly been successful—evidenced by its research outcomes, the achievement of its goals, and favorable assessments from scientists both within and outside the project. Over the past decade, the HBP has not only made outstanding contributions to neuroscience but also paved the way for technological innovations in fields such as artificial intelligence and medical diagnostics.
By championing an interdisciplinary approach, the HBP has effectively integrated brain science, computer science, and medicine, thereby establishing a new research paradigm and advancing the application of digital brain models across various domains, including medicine and technology. In particular, in the area of functional modeling, the project employs computer simulations to replicate brain function, thereby partially revealing how the brain processes information. Notably, the cortical column model has provided a fundamental basis for understanding neural networks.
▷ Figure 7. A biomorphic model of the cortical column for content-based image retrieval. Source: MDPI, Telnykh, Alexander, et al. “A Biomorphic Model of Cortical Column for Content—Based Image Retrieval.” Entropy 23.11 (2021): 1458.
Thanks to substantial financial investment and an international network of collaboration, the HBP has deepened our understanding of brain structure and function while driving the practical application of digital brain models across diverse fields. The open-access platform established by the project provides researchers worldwide with invaluable tools and data resources, significantly advancing progress in the diagnosis and treatment of brain disorders.
Reflections on Limitations
Despite its remarkable achievements, several aspects of the HBP’s ten-year journey merit reflection. At its inception, the HBP set goals that were overly complex and optimistic. For instance, the project once attempted to simulate the entire brain comprehensively, yet current technology remains incapable of accurately replicating such an intricate system. As Peter Dayan, Director of Computational Neuroscience at University College London, noted, “The existing technology for simulating the human brain is not yet fully mature.”
Furthermore, these simulations remain limited when applied to larger-scale and more complex networks. In 2024, Nobel Prize in Physics laureate and “father of artificial intelligence” Geoffrey Hinton emphasized, “The real problem with the HBP is that scientists do not yet know how to enable large systems to learn autonomously.” On the project management front, transparency issues that emerged in 2015 attracted significant attention and criticism from the scientific community, prompting some researchers to withdraw. Additionally, some critics have argued that the project’s funds were not used efficiently, with substantial portions of the budget allocated to relatively immature or ambiguous technological developments.
These issues serve as a reminder that while large-scale scientific projects aim for groundbreaking outcomes, they also require pragmatic planning and meticulous management.
Future Prospects
Reflecting on the development of the HBP, we must acknowledge that brain science research still faces a number of fundamental challenges. These challenges are not only hurdles for the HBP but also key areas that the entire field of brain science must continue to explore:
· Overall Working Principles of the Brain: How are complex neuronal networks integrated to execute cognitive, emotional, and motor tasks?
· Neural Basis of Consciousness: How does consciousness emerge from the brain’s electrophysiological activity?
· Mechanisms of Memory and Learning: How is information encoded and retrieved through synaptic plasticity?
· Pathology of Neurological Disorders: How does an imbalance in brain health lead to psychiatric and neurological diseases?
Building on the research foundation established by the HBP, more than a hundred scientists have authored articles envisioning the next decade of brain science research [2]. They have highlighted several innovative technological directions worthy of attention:
· High-Resolution Brain Imaging Technologies: Next-generation functional magnetic resonance imaging (fMRI) and single-molecule resolution optical imaging will help reveal the flow of information within the brain with greater clarity.
· Brain-Machine Interface Technology: This is expected to yield breakthroughs in understanding neural signal transmission and treating brain disorders.
· Artificial Intelligence and Machine Learning: Their widespread application will further enhance our ability to analyze complex neural data, simulate brain function, and predict the behavior of neural networks.
· Next-Generation Molecular Biology Tools: The development of new tools—including, but not limited to, gene editing techniques—will deepen our understanding of the functions of specific neurons.
With the continuous advancement of new technologies and the invaluable experience accumulated by the HBP, our understanding of brain science is poised to accelerate in the coming years.
Postscript
Throughout its ten-year journey, the HBP has faced negative rumors. In 2015, some scientists issued an open letter to boycott the project; in 2017, the project’s lead figure was replaced; and in 2023, some media outlets even bluntly labeled the HBP as a failed project [3]. In Chinese discourse, the HBP has received limited attention, and existing reports—predominantly negative and echoing some international evaluations—have focused on issues such as unclear objectives, fragmented research efforts, and a lack of funding transparency.
For a scientific project of such enormous scale, negative voices are not uncommon; they are a natural manifestation of free expression and public oversight. From the Manhattan Project to the Apollo program and the Human Genome Project, every large-scale initiative has faced doubts and criticisms. Scientific exploration is inherently filled with uncertainty, particularly for massive projects. The key lies in discerning the facts through firsthand sources and a thorough understanding of the project’s actual progress, rather than being swayed by sensationalist public opinion.
Therefore, when evaluating the HBP, it is essential to return to the core questions: What has it truly accomplished over the past ten years? Has it achieved its intended objectives? Have its research outcomes been recognized within the field? The answers to these questions are the true criteria for judging the success or failure of a scientific project.
In reviewing the HBP’s ten-year journey, it is clear that the project has not only pushed the frontier of brain science research but also facilitated breakthroughs in numerous technologies and theories. In this light, the HBP is undoubtedly a successful project. As a milestone in humanity’s exploration of brain science, its significance will become increasingly apparent over time.
Reference:
]]>
On a city street, a self-driving car is moving. Suddenly, a child chasing a ball dashes into the road. The car’s multiple onboard sensors immediately capture the situation—cameras, LiDAR, and millimeter-wave radar all operate simultaneously, while dedicated neural network processors and GPUs begin high-speed computations. Under peak conditions, the entire system consumes hundreds of watts of power and, through parallel processing, completes the perception-to-decision process in approximately 100 milliseconds.
In the same scenario, however, a human driver can hit the brakes in an instant. In such complex situations, the human brain consumes only about 20 watts of power—comparable to a small light bulb. Even more astonishingly, the brain, through distributed parallel processing, simultaneously manages countless tasks, such as breathing and regulating heartbeats. It performs parallel computations with extremely low energy consumption, learns quickly from limited experience, and adaptively handles a wide range of unforeseen situations.
This enormous efficiency gap has spurred scientists to pioneer a brand-new field—NeuroAI (neuro-inspired artificial intelligence). This emerging discipline seeks to overcome the limitations of traditional AI by mimicking the brain’s operational methods to create smarter and more efficient systems. For example, researchers draw inspiration from how the brain processes visual information to improve computer vision, reference neuronal connectivity patterns to optimize deep learning networks, and study the brain’s attention mechanisms to reduce the energy consumption of AI systems.
This article will take you on an in-depth journey into NeuroAI, reviewing its definition, development history, current research status, and future trends. It will also examine how NeuroAI simulates the brain’s learning mechanisms, information processing methods, and energy utilization efficiency—and how it addresses the bottlenecks currently faced by AI—in light of the latest research presented at a recent NeuroAI symposium hosted by the National Institutes of Health (NIH).
What is NeuroAI?
NeuroAI exemplifies the two-way integration of artificial intelligence and neuroscience. On one hand, it leverages artificial neural networks as innovative tools to study the brain, thereby testing our understanding of neural systems through constructing verifiable computational models. On the other hand, it draws on the brain’s operational principles to enhance AI systems, transforming the advantages of biological intelligence into technological innovations.
For years, the fundamental goal of AI research has been to create artificial systems capable of performing every task at which humans excel. To achieve this, researchers have consistently turned to neuroscience for inspiration. Neuroscience has spurred advancements in artificial intelligence, while AI has provided robust testing platforms for neuroscience models—forming a positive feedback loop that accelerates progress in both fields.
The relationship between AI and neuroscience is genuinely symbiotic rather than parasitic. The benefits that AI brings to neuroscience are as significant as the contributions of neuroscience to AI. For example, artificial neural networks lie at the core of many state-of-the-art models of the visual cortex in neuroscience. The success of these models in solving complex perceptual tasks has led to new hypotheses about how the brain might perform similar computations. Similarly, deep reinforcement learning—a brain-inspired algorithm that combines deep neural networks with trial-and-error learning—serves as a compelling example of how AI and neuroscience mutually enhance one another. It has not only driven groundbreaking achievements in AI (including AlphaGo’s superhuman performance in the game of Go) but also deepened our understanding of the brain’s reward system.
The Development History of NeuroAI
The interplay between computer science and neuroscience can be traced back to the very inception of modern computers. In 1945, John von Neumann, the father of modern computing, dedicated an entire chapter in his landmark EDVAC architecture paper to discussing the similarities between the system and the brain. Notably, the paper’s sole citation was a brain study by Warren McCulloch and Walter Pitts (1943), which is widely regarded as the first work on neural networks and laid the foundation for decades of mutual inspiration between neuroscience and computer science.
▷ Figure 1. A depiction of the logical computing units in computers from von Neumann’s paper, inspired by excitatory and inhibitory neurons. Source: von Neumann, J. (1993). First draft of a report on the EDVAC. IEEE Annals of the History of Computing, 15(4), 27-75.
The concept of neural networks achieved a major breakthrough in 1958 when Frank Rosenblatt published a paper entitled “The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain.” In this paper, he first proposed the revolutionary idea that neural networks should learn from data rather than being statically programmed. This breakthrough was featured in The New York Times under the headline “Electronic ‘Brain’ Learns by Itself,” sparking an early surge of interest in artificial intelligence research. Although Marvin Minsky and Seymour Papert pointed out the limitations of single-layer perceptrons in 1969—triggering the first “neural network winter”—the core concept that synapses are the plastic elements, or free parameters, in neural networks has persisted to this day.
In recent developments in artificial intelligence, numerous innovations have been inspired by neuroscience. One of the most prominent applications of NeuroAI is the highly successful convolutional neural network in the field of image recognition, whose inspiration can be traced back to the modeling studies of the visual cortex conducted by David Hubel and Torsten Wiesel forty years ago. Another example is the dropout technique, in which individual neurons in an artificial network are randomly deactivated during training to prevent overfitting. By simulating the random misfiring of neurons in the brain, dropout helps artificial neural networks achieve greater robustness and generalization capabilities.
Three Insights from the Brain for NeuroAI
Over the past decade, AI has made significant strides in various fields—it can write articles, pass law exams, prove mathematical theorems, perform complex programming, and execute speech recognition. However, in areas such as real-world navigation, long-term planning, and inferential perception, AI’s performance remains, at best, mediocre.
As Richard Feynman once said, “Nature’s imagination far surpasses that of humans.” The brain—currently the only system that flawlessly performs these complex tasks—has been refined over 500 million years of evolution. This evolutionary process enables animals to accomplish intricate tasks, such as hunting, with ease—tasks that contemporary AI still struggles to master.
This is precisely the direction that NeuroAI seeks to emulate: overcoming current AI bottlenecks by studying the brain’s operational mechanisms. In practice, this inspiration is reflected in the following aspects:
(1) Genomic Bottleneck
Unlike AI, which requires massive amounts of data to train from scratch, biological intelligence inherits evolutionary solutions through a “genomic bottleneck.” This process allows animals to perform complex tasks instinctively. In this process, the genome plays a crucial role: it provides the fundamental blueprint for constructing the nervous system, specifying the connectivity patterns and strengths between neurons, and laying the foundation for lifelong learning.
▷ Figure 3. Species that employ both innate and learned strategies tend to have an evolutionary advantage over those relying solely on innate instincts. Source: Zador, A. M. (2019). A critique of pure learning and what artificial neural networks can learn from animal brains. Nature Communications, 10, Article 3770.
It is noteworthy that the genome does not directly encode specific behaviors or representations, nor does it explicitly encode optimization principles. Instead, it primarily encodes connectivity rules and patterns that, through subsequent learning, give rise to concrete behaviors and representations.
Evolution acts upon these connectivity rules, suggesting that when designing AI systems, greater emphasis should be placed on the network’s connectivity topology and overall architecture. For example, by mimicking the connectivity patterns found in biological systems—such as those in the visual and auditory cortices—we can design cross-modal AI systems. Furthermore, by compressing weight matrices through a “genomic bottleneck,” we can extract the most critical connection features as an “information bottleneck,” enabling more efficient learning and reducing dependence on large training datasets.
▷ Figure 4. Architecture and performance of an artificial neural network designed based on the genomic bottleneck principle for reinforcement learning tasks. Source: Shuvaev, S., Lachi, D., Koulakov, A., & Zador, A. (2024). Encoding innate ability through a genomic bottleneck. Proceedings of the National Academy of Sciences, 121(38), Article e2409160121.
(2) The Brain’s Energy-Efficient Approach
There is a vast discrepancy in energy consumption between artificial neural networks and the biological brain. Currently, models like ChatGPT require at least 100 times more energy for real-time conversation than the human brain consumes. Moreover, comparing the energy consumption of GPU arrays to that of the entire brain significantly underestimates the brain’s energy advantage—maintaining a conversation actually uses only a small fraction of the brain’s energy budget.
The brain’s remarkable energy efficiency can be attributed to two key factors: sparse neural firing and high noise tolerance.
First, most of a neuron’s energy is expended on generating action potentials, with energy consumption roughly proportional to the frequency of neural spikes. In the cerebral cortex, neurons operate sparsely, firing only about 0.1 spikes per second on average. In contrast, current artificial networks exhibit high spike rates and operate under conditions of elevated energy consumption. Although the energy efficiency of artificial networks has improved somewhat, we remain far from mastering the brain’s energy-saving computational mode based on sparse spiking.
Second, the brain is highly tolerant of noise. During synaptic transmission, even if up to 90% of spikes fail to trigger neurotransmitter release, the brain continues to function normally. This stands in stark contrast to modern computers, where digital computations rely on the precise representation of zeros and ones—where even a single bit error can lead to catastrophic failure, necessitating significant energy expenditure to ensure signal accuracy. Developing brain-inspired algorithms that can operate effectively in noisy environments could lead to substantial energy savings.
(3) Biological Systems Balancing Multiple Objectives
There is a marked difference between biological systems and current AI in terms of goal management. Modern AI systems typically pursue a single objective, whereas organisms must balance multiple objectives simultaneously over both short-term and long-term scales—such as reproduction, survival, predation, and mating (the so-called “4Fs”). Our understanding of how animals balance these diverse goals remains limited, largely because we have yet to fully decipher the brain’s computational mechanisms. As we strive to develop AI systems capable of managing multiple objectives, insights from neuroscience can provide essential guidance. Conversely, AI models can serve as platforms to test and validate theories of multi-objective management in the brain. This symbiotic interaction between neuroscience and AI research is propelling advancements in both fields.
Frontier Advances in NeuroAI Research
The intersection of neuroscience and artificial intelligence continues to deepen, with researchers drawing inspiration from biological intelligence systems and exploring scales ranging from microscopic molecules to macroscopic structures. Below is an overview of a series of innovative NeuroAI breakthroughs presented at a special symposium hosted by the NIH in November 2024, which enhance our understanding of the nature of intelligence.
1.Astrocytic Biocomputation
Living neural networks exhibit rapid adaptability to their environments and can learn from limited data. In this context, astrocytes—serving as carriers of information—play a critical role in the gradual integration of information within neural networks. This discovery offers a fresh perspective on the unique capabilities of biological neural networks and holds significant implications for both neuroscience and artificial intelligence.
2.The Feedback Loop Between In Vivo and Virtual Neuroscience
Recent advances in neurotechnology have enabled researchers to record neural activity under natural conditions with unprecedented coverage and biophysical precision. By developing digital twins and foundational models, researchers can conduct virtual experiments and generate hypotheses to simulate neural activity, thereby exploring brain function in ways that transcend the limitations of traditional experimental methods. This shift toward virtual neuroscience is crucial for accelerating progress and provides valuable insights for developing flexible, secure, and human-centric AI systems.
3.Advanced NeuroAI Systems with Dendrites
Dendrites, which serve as the receptive components of neurons, play a pivotal role in biological intelligence. Incorporating dendritic features into AI systems can improve energy efficiency, enhance noise robustness, and mitigate issues such as catastrophic forgetting. However, the core functional characteristics of dendrites and their applications within AI remain unclear, limiting the development of brain-inspired AI systems. Addressing this challenge will require interdisciplinary research to explore the anatomical and biophysical properties of dendrites across different species. With the aid of computational models and new mathematical tools, researchers can better understand dendritic functions, advance dendrite-based AI systems, and deepen our understanding of the design principles and evolutionary significance of biological systems.
4.The Future of NeuroAI: Drawing Inspiration from Insects and Mathematics
Although current AI systems are powerful, they rely on massive networks, enormous datasets, and tremendous energy consumption. Compared to natural intelligence, they lack key biological mechanisms such as neuromodulation, neural inhibition, physiological rhythms, and dendritic computation. Incorporating these mechanisms into AI to enhance performance remains an important challenge. Recent connectomics research on fruit flies has opened a new direction by drawing inspiration from the simple yet efficient biological brain. However, this approach faces significant mathematical challenges, particularly in managing high-dimensional, nonlinear dynamic systems like neural networks.
5.Learning from Neural Manifolds: From Biological Efficiency to Engineered Intelligence
Recent breakthroughs in experimental neuroscience and machine learning have revealed striking similarities in how biological systems and AI process information across multiple scales. This convergence presents an opportunity for a deeper integration of neuroscience and AI over the next decade. Research suggests that the geometric principles underlying neural network representation and computation could fundamentally change the way we design AI systems while deepening our understanding of biological intelligence. To realize this potential, several key areas must be addressed: 1) Developing new technologies to capture the dynamic evolution of neural manifolds across various time scales during behavior; 2) Constructing theoretical frameworks that connect single-neuron activity with collective computations to reveal principles of efficient information processing; 3) Elucidating cross-modal representational theories that explain the robustness of neural manifolds and their transformations, along with their underlying mechanisms; 4) Leveraging the efficiency of biological neural systems to develop computational tools for analyzing large-scale neural data. Interdisciplinary research spanning statistical physics, machine learning, and geometry promises to develop AI systems that more closely emulate the traits of biological intelligence, offering superior efficiency, robustness, and adaptability.
6.Advancing Toward Insect-Level Intelligent Robotics
Thanks to modern advances in control theory and AI, contemporary robots are now capable of performing nearly all tasks that humans can—from climbing ladders to folding clothes. Looking ahead, key challenges remain, such as designing systems that can make autonomous decisions and execute diverse tasks using only onboard computing without relying on cloud services. The fruit fly provides an ideal research model. With the complete connectome of the fruit fly’s brain and ventral nerve cord already mapped, we know that these tiny organisms can autonomously switch between multiple behaviors based on both internal and external states. However, to apply this model to robotics, higher-resolution connectomics data must be obtained and expanded to include more specimens. In addition, the connectomes of insects with more complex behaviors, such as praying mantises, need to be studied to better understand the relationship between neural architecture and intelligence. Finally, in-depth research into dendritic and axonal structures—beyond simple point-to-point connectivity—must be undertaken, along with the development of neuromorphic hardware capable of simulating millions of neurons that is readily accessible.
7.Mixed-Signal Neuromorphic Systems for Next-Generation Brain–Machine Interfaces
While traditional AI algorithms and techniques are effective for analyzing large-scale digital datasets, they face limitations when applied to closed-loop systems for real-time processing of sensory data—especially in the context of bidirectional interfaces between the human brain and brain-inspired AI, as well as neurotechnologies that require real-time interaction with the nervous system. These challenges are primarily related to system requirements and energy consumption. On the system side, low-latency local processing is necessary to ensure data privacy and security; in terms of energy, wearable and implantable devices must operate continuously with power consumption strictly controlled at sub-milliwatt levels. To address these challenges, research is increasingly turning toward bottom-up physical approaches, such as neuromorphic circuits and mixed-signal processing systems. These neuromorphic systems, employing passive subthreshold analog circuits and data-driven encoding methods, can execute complex biomedical signal classification tasks—such as epilepsy detection—at micro-watt power levels.
Afterword
As the field of NeuroAI continues to evolve, we find ourselves at a unique juncture in history: the deep integration of neuroscience and artificial intelligence not only helps us gain a deeper understanding of human intelligence but also inspires entirely new approaches to designing next-generation AI systems. From genomic bottlenecks to dendritic computation, and from energy efficiency to multi-objective balancing, the diverse characteristics observed in biological intelligence systems are reshaping our understanding of what intelligence truly is. Looking ahead, as brain science research deepens and AI technology advances, NeuroAI may spark a revolution in computational paradigms—enabling us to build artificial systems that more closely resemble biological intelligence. This journey is not merely about technological innovation; it is also a profound exploration of the nature of intelligence. In the process, we are not only creating new technologies but also rediscovering ourselves.
Reference:
]]>
One day, a wormhole appears in your garden, and you retrieve a book from it. The text in the book is complex and difficult to understand, resembling an alien language. How would you go about deciphering it? Would you start by analyzing whether these characters form a fixed set of symbols, much like our alphabet, or by observing the patterns in which they combine? Or perhaps you would consider enlisting the help of a large model, hoping it could assist you in understanding the book’s contents? So, can large models really learn “alien languages”?
Before attempting to learn an alien language, consider that large models have already successfully deciphered the languages of animals such as whales. Moreover, large models are capable of rapidly mastering a myriad of programming languages. So, what characteristics of an alien language would make it easier for large models to decipher? Recently, a study published in Nature Communications pointed out that the compositionality of language structure not only enhances the efficiency with which large models learn but also facilitates language acquisition for humans.
▷ Figure 1. Source: Galke, L., Ram, Y. & Raviv, L. Deep neural networks and humans both benefit from compositional language structure. Nat Commun 15, 10816 (2024). https://doi.org/10.1038/s41467-024-55158-1
What Is the Compositionality of Language?
Compositionality refers to the idea that combining two words in a language can yield a new, more complex concept. Consider two languages: In Language A, describing a “black horse” requires only combining the word for “black” with the word for “horse” to form the appropriate term, whereas in Language B the concepts of “horse,” “black,” and “black horse” are represented by three entirely distinct words. If Language A exhibits a greater number of such compound words than Language B, then it is said to possess higher compositionality.
▷ Figure 2. For aliens, which term—“zebra” or “斑马”(the horse with dots)—is more likely to lead them to associate the picture with the text? Source: AlLes
For adults, learning a language with strong compositionality requires a higher level of logical reasoning. This quality enables learners to deduce a set of generative rules instead of having to memorize each word individually. We have all experienced this: when learning English, understanding word roots makes vocabulary acquisition much easier than memorizing words directly from a vocabulary list. A language with high compositionality allows learners, after being exposed to a limited set of elements, to use these rules to produce an infinite array of expressions. In fact, research indicates that modern languages typically feature a robust compositional structure—an evolutionary development that enhances the efficiency of both learning and using the language.
A similar concept exists in programming languages. In low-level assembly language, every operation on a variable requires its own corresponding instruction. In contrast, high-level languages such as Python allow multiple operations to be consolidated into a single function call, which can, for instance, perform thousands of calculations on a matrix in one go. Large-scale models have shown advantages in understanding and applying programming languages, especially high-level languages with strong compositionality. However, previous studies have suggested that these models did not necessarily benefit from languages with a strong organizational structure.
To address this, Galke et.al sought to answer the following question: When trained on more structured language inputs, do deep neural network models exhibit the same learning and generalization advantages as adult humans? The researchers employed GPT-3.5 as the pre-trained model and an RNN as the language model to be trained. They used artificially simulated languages with varying degrees of compositionality as training texts to assess the learning capabilities of both human subjects and large models when confronted with these laboratory-generated virtual languages. The results revealed that within the training texts, stronger structural compositionality corresponded with greater improvements in generalization performance. This pattern was observed in humans as well as in both pre-trained and untrained artificial neural networks (Figure 3).
▷ Figure 3. Overview of the experimental design. The researchers devised artificial languages with different levels of structure, divided into two categories: low-structure and high-structure languages. Low-structure languages lack systematic organization and compositionality, while high-structure languages exhibit both systematic organization and compositionality with respect to their shape and angular properties. The experimental procedure consisted of multiple rounds of training, each comprising an exposure phase, a guessing phase, and a generation phase. At the end of each round, memory tests and generalization tests were conducted to evaluate the models’ ability to reproduce previously seen items and to generate new items.
Highly Structured Languages Are Easier to Learn
First, the researchers explained why large-scale models do not tend to favor highly compositional languages. In short, deep neural networks usually possess an enormous capacity, which means they can easily memorize every individual linguistic expression without needing to rely on identifying compositional patterns to reinforce memory. However, this does not imply that highly compositional languages are irrelevant to large-scale models. In languages with greater compositionality, individual semantic units are reused across various contexts, leading to a higher frequency of occurrence in the training data. Consequently, these recurring semantic units and their contextualized patterns are learned more effectively through repeated exposure during the training process.
Let us return to the alien example. Suppose the alien book contains a reference table indicating that the meaning of “next” is “to the right” and that of “question” is “upwards.” Then, how would you express “up and to the right”? In a highly compositional language, one can deduce a pattern whereby combining “next” and “question” conveys the meaning “up and to the right,” whereas in a language with low compositionality such a pattern might be absent. This ability to apply learned knowledge or skills to new, unseen situations or data is known as “generalization.” When comparing languages with high compositionality to those with low compositionality, both human learners and large-scale models achieve significantly higher generalization scores with highly structured languages (see Figure 4).
▷ Figure 4. The figure shows the final generalization scores achieved by humans (A), GPT-3.5 (B), and RNN (C) across different input languages. The horizontal axis represents the structural score of the input language, and the vertical axis represents the generalization score. Each point corresponds to the overall generalization score of an input language, reflecting the extent to which the learner systematically generalizes new labels based on the learned labeling system. For instance, if a learner successfully recombines previously used components—such as combining “muif,” which represents shape, with “i,” which denotes direction, to form “muif-i”—the generalization score will be high. The shaded area around the regression line indicates the 95% confidence interval estimated via bootstrapping.
Furthermore, when training on more structured languages—that is, languages with clear grammatical rules and syntactic hierarchies—GPT-3.5 begins to exhibit prediction patterns that closely resemble those of human subjects. Figure 5B illustrates the similarity between GPT-3.5’s predictions and those of human participants for the next word in the new language under the same conditions. Similarly, Figure 5A shows that as the structure of the training text improves, the similarity among human learners during the generalization process also increases.
▷ Figure 5. The figure displays the final similarity scores between the outputs produced by humans (A), GPT-3.5 (B), and RNN (C) during the generalization process compared to human production. The horizontal axis represents the structural score of the input language, while the vertical axis indicates the production similarity score (calculated as the length-normalized edit distance), which measures how closely the labels generated by the models match those produced by human participants.
In essence, when learning highly structured languages, both large-scale models and human learners tend to converge—each opting to exploit the inherent structure of the language—which in turn leads to more accurate predictions for subsequent word generation.
Moreover, when confronted with languages that exhibit a higher degree of structural organization, large-scale models not only predict subsequent words and phrases more accurately but also learn at a faster pace (see Figure 6C). Their performance in terms of memory retention and generalization becomes more similar to that of humans as well (see Figures 6A, 6B, and 6D).
▷ Figure 6. The figure demonstrates how more structured languages lead to improved and faster reproduction of the input language (A), better generalization to novel scenarios (C), higher consistency with human participants in memory (B) and generalization (D) processes, as well as greater convergence among networks (E).
Ultimately, the degree of language structure also affects the trajectory of generalization. In highly structured languages, rules are explicit and transparent, with each semantic unit consistently corresponding to its form in a regular pattern. When learning such a language, both humans and neural networks encounter little ambiguity, and all possible generalization paths converge to a consistent outcome. In contrast, less structured languages lack clear rules and compositionality, which results in multiple plausible generalizations. Different options may all seem reasonable, leading to linguistic diversity, such as the formation of dialects.
Therefore, highly structured languages facilitate better generalization and enhance consistency in language cognition both among different neural networks and between neural networks and humans. This finding supports the view that large language models are valuable for studying human cognitive mechanisms and provides additional evidence of the similarities between human and machine language learning.
Can Large Language Models Learn Alien Languages?
In terms of language learning, large language models have been shown to possess capabilities similar to those of humans. Considering that these models also boast superior memory, it is conceivable that one day, when faced with extraterrestrials, they might indeed help us learn an alien language. However, the real challenge lies in the fact that if an alien language lacks sufficient systematic structure, our understanding and usage of it could suffer from high error rates and uncertainty.
The alien language depicted in the science fiction film Arrival, with its highly non-linear and complex symbolic structure, appears to offer a mode of thinking that transcends our current cognitive abilities. Its uniqueness lies in its departure from traditional linear structures, allowing learners to grasp all the information in a sentence simultaneously and even predict future events. From the perspective of structured language learning, an alien language might exhibit even greater systematicity than terrestrial languages, providing learners with richer information and thereby enabling them to forecast what lies ahead.
▷ The alien script used by the extraterrestrials in Arrival. Source: Film Industry Network
From this viewpoint, input composed of highly structured language can enable large language models to generalize more effectively, thereby enhancing their ability to comprehend new contexts. Consequently, if an alien language were to possess a more precise and orderly structure, models trained on big data might gradually acquire and understand its grammatical rules—much like humans do—thereby eventually learning the alien language and even transforming their cognitive processes to comprehend the future, much like the protagonist in Arrival.
Returning from science fiction to reality, communication among agents powered by large language models has even given rise to new languages. However, these emergent languages often lack structure and are not easily understood by other agents [1]. This may be because agents that do not face “survival pressure” tend to generate disordered and hard-to-learn forms of communication when new languages emerge [2]. The evolutionary history of human language reflects this as well; in the absence of actual survival needs, languages often struggle to remain efficient and systematic [3].
Looking even further ahead, if one day humanity seeks to break down the language barriers between different countries and ethnic groups, we wilal also face the challenge of learning an entirely new language. At that point, if we intend to design a new language, its systematic structure must be taken fully into account. Only a language with clear, structured grammatical rules can be rapidly mastered by diverse groups worldwide and easily understood by various intelligent agents. Perhaps the book delivered to your garden by a wormhole is, in fact, an “Esperanto” dictionary sent by future humans across time.
Reference:
]]>