Source-linked AI summary
Toward Next-Generation Artificial Intelligence: Catalyzing the NeuroAI Revolution
Anthony Zador, Sean Escola, Blake Richards, Bence Ölveczky, Yoshua Bengio, Kwabena Boahen, Matthew Botvinick, Dmitri Chklovskii, Anne Churchland, Claudia Clopath, James DiCarlo, Surya Ganguli, Jeff Hawkins, Konrad Koerding, Alexei Koulakov, Yann LeCun, Timothy Lillicrap, Adam Marblestone, Bruno Olshausen, Alexandre Pouget, Cristina Savin, Terrence Sejnowski, Eero Simoncelli, Sara Solla, David Sussillo, Andreas S. Tolias, Doris Tsao
TL;DR
AI has achieved strong performance in human-centered domains but still lacks robust physical interaction, causal common sense, and animal-level sensorimotor abilities. The paper proposes NeuroAI and an embodied Turing test that incrementally benchmarks artificial agents against living animals, providing a roadmap centered on evolved capabilities and supported by shared computational infrastructure.
Problem
Current AI systems struggle with novel physical situations and cannot match the sensorimotor abilities of young children or simple animals, while conventional language-based tests omit these capabilities.
Method
The paper proposes NeuroAI research organized around species-specific embodied Turing tests, evolutionary task progression, animal behavioral benchmarks, and shared platforms for agent development.
Results
The paper concludes that embodied Turing tests can provide a roadmap for developing AI with animal-level sensorimotor intelligence.
Takeaways & Limitations
Progress toward general AI should draw more heavily on neuroscience and on the evolved sensorimotor solutions of animals.
Abstract
from arXiv · showhide
Neuroscience has long been an essential driver of progress in artificial intelligence (AI). We propose that to accelerate progress in AI, we must invest in fundamental research in NeuroAI. A core component of this is the embodied Turing test, which challenges AI animal models to interact with the sensorimotor world at skill levels akin to their living counterparts. The embodied Turing test shifts the focus from those capabilities like game playing and language that are especially well-developed or uniquely human to those capabilities, inherited from over 500 million years of evolution, that are shared with all animals. Building models that can pass the embodied Turing test will provide a roadmap for the next generation of AI.
NeuroAI Grand Challenge: The Embodied Turing Test
The embodied Turing test extends evaluation beyond language to species-specific sensorimotor interaction, exposing how far AI remains from animal-level physical intelligence. NeuroAI research is proposed as a path toward capabilities grounded in perception, action, flexibility, and efficient computation.
- Limits of conventional AI evaluation: Large language models have advanced conversational performance, but language-based evaluation omits physical interaction and causal common sense.These abilities are shared with animals and shaped by evolution.
- The embodied Turing test: The embodied Turing test asks whether an artificial animal’s behavior is indistinguishable from that of its living counterpart.Each species defines its own test, such as dam construction for an artificial beaver.
- Limits of current AI: Current AI systems remain brittle in novel situations and cannot match the sensorimotor abilities of young children or simple animals.They struggle with tasks such as walking to a shelf, manipulating objects, building nests, foraging, and caring for young.
- Shared animal capabilities: Animals’ flexibility reflects the ability to use general knowledge to master novel situations from limited experience rather than large labeled datasets.Evolution and development provide animals with strong foundations for real-world interaction.
- Shared animal capabilities: Biological computation is substantially more energy-efficient than modern AI, with the human brain using about 20 watts and sparse neural spikes supporting efficiency.Spike-based computation has also shown orders-of-magnitude improvements in speed and energy efficiency in hardware implementations.
A roadmap for solving the embodied Turing test
The proposed roadmap decomposes the embodied Turing test into progressively harder challenges guided by evolutionary history and evaluated against animal behavior. Standardized, species-specific benchmarks can support comparison across research groups and capabilities.
- A roadmap for solving the embodied Turing test: An evolutionary strategy breaks the embodied Turing test into incrementally challenging tasks, beginning with foundational abilities such as goal-directed locomotion.More sophisticated skills are layered on top of these basic capacities.
- A roadmap for solving the embodied Turing test: The roadmap spans organisms studied in neuroscience, including worms, flies, fish, rodents, and primates.This scope connects artificial-agent development with accumulated biological knowledge.
- Evaluation against animals: Rich behavioral datasets and biomechanical measurements can benchmark artificial agents against species-specific ethograms and realistic animal body models.New 3D videography tools are expanding behavioral datasets.
- Evaluation against animals: Species-specific tests can assess sensorimotor control, learning, generalization, memory-guided behavior, and social interactions.The challenges can be standardized to quantify progress and compare research efforts.
What we need
The NeuroAI program requires coordinated investment in interdisciplinary training, shared computational infrastructure, and fundamental research on neural computation. These resources are intended to support iterative development and evaluation of embodied agents.
- What we need: A new generation of researchers must be trained across engineering, computational science, and neuroscience.Their role is to draw on neuroscience to chart new directions in AI research.
- What we need: A shared platform is needed to develop and test virtual agents across increasingly complex embodied Turing tests.Training one large neural network on a single embodied task can already take days on specialized distributed hardware.
- What we need: Large-scale collaboration will require substantial computational power to iteratively optimize and evaluate many agents over multiple generations.The paper compares this shared resource to a particle accelerator or large telescope.
- What we need: Fundamental theoretical and experimental research is needed to understand neural computation using knowledge accumulated about brain cells and circuits.The paper identifies existing neuroscience investments, including the BRAIN Initiative, as an important foundation.
Conclusions
The paper argues that neuroscience remains relevant to AI because achieving animal-level sensorimotor intelligence requires understanding the biological solutions animals evolved. It calls for renewed attention to neuroscience within AI research.
- Conclusions: Neuroscience has historically driven major AI advances, but its influence and opportunities are often underrecognized in current AI research and education.The paper notes that leading AI conferences now focus largely on machine learning rather than computational neuroscience.
- Conclusions: If AI seeks animal-level common-sense sensorimotor intelligence, studying animals is directly relevant to reproducing their abilities in unpredictable environments.The paper uses bird flight through dense forest as an analogy for this design goal.