Meta and TRIBE v2: the AI that predicts your brain reactions

Meta and Its Predictive AI Model TRIBE v2

Meta has recently unveiled TRIBE v2, an innovative artificial intelligence model capable of predicting how the human brain responds to various stimuli, whether visual, auditory, or linguistic. This model, which relies on functional magnetic resonance imaging (fMRI) data, has demonstrated the ability to surpass the accuracy of individual measurements in anticipating brain responses.

During rigorous testing, TRIBE v2 successfully reproduced established results in neuroscience, such as the precise identification of brain areas dedicated to processing faces, places, and language. This advancement could potentially reduce the time and resources needed for laboratory research on the brain.

However, TRIBE v2 is not without limitations. It perceives the brain as a mere passive receiver and only accounts for three types of sensory stimuli. Additionally, it is dependent on the relatively slow temporal resolution of fMRI. Nevertheless, Meta has decided to make the source code, model weights, and an interactive demonstration freely accessible to the public.

An AI Model Trained on Massive Data

The development of TRIBE v2 is based on a vast database, with over 1,000 hours of fMRI data from 720 participants. fMRI is a technique that indirectly measures brain activity by tracking changes in blood flow and oxygen levels. With this information, TRIBE v2 is designed to predict the brain's reaction to any type of visual, auditory, or linguistic stimulus.

Three Models for Data Preprocessing

TRIBE v2 uses three types of inputs: video, audio, and text. Each type of data is first processed by a pre-trained AI model from Meta: Llama 3.2 for text, Wav2Vec-Bert-2.0 for audio, and Video-JEPA-2 for video. These models transform raw data into numerical representations, or embeddings, that capture visible, audible, or readable elements.

A transformer then processes these representations together, identifying common patterns across different stimuli, tasks, and individuals. A final subject-specific layer translates this information into a brain map composed of 70,000 voxels, the three-dimensional measurement units used in fMRI.

Reducing Noise in Predictions

Individual fMRI images are often affected by various noise factors, such as heartbeat, head movements, and device artifacts. To obtain a typical brain response to a stimulus, researchers usually need to average many fMRIs. TRIBE v2 circumvents this challenge by directly predicting an adjusted average response.

In tests, TRIBE v2's predictions showed a stronger correlation with the actual group average than most individual fMRIs. This effect is particularly notable in the Human Connectome Project dataset, captured with a 7 Tesla scanner, which offers superior signal quality compared to standard 3 Tesla machines.

Advancements Over Previous Models

TRIBE v2 surpasses optimized linear models, which had been the standard for this type of prediction. The previous version, TRIBE v1, was trained on only four subjects and predicted just 1,000 voxels. Despite this, it won the Algonauts 2025 competition, outperforming 263 other teams.

The accuracy of TRIBE v2's predictions improves with the increase in training data and has not yet plateaued, suggesting that the model will continue to enhance as fMRI databases grow. This phenomenon reflects the scaling laws observed in large language models, where more data generally leads to better performance.

Reproducing Decades of Laboratory Research

Researchers tested TRIBE v2 with everyday stimuli, such as movies and podcasts, where multiple sensory inputs stimulate the brain simultaneously, as well as with isolated stimuli typical of classical neuroscience. In these controlled contexts, a single image could be presented on the screen for one second to measure the response of a specific brain region.

The team used testing protocols from the Individual Brain Charting dataset, a well-established collection of neuroscientific experiments, and asked the model to predict which brain areas should activate. In visual experiments, TRIBE v2 identified brain regions specialized for faces, places, bodies, and characters.

Mapping Sensory Channels

By selectively disabling individual input channels, TRIBE v2 reveals how much each sense stimulates activity in specific brain regions. The results are consistent with current neuroscientific knowledge: audio best predicts activity near the auditory cortex, video is associated with the visual cortex, and text activates language areas and certain parts of the frontal lobe.

In regions where the brain combines inputs from multiple senses, using all three channels offers the greatest gains. At the junction of the temporal, parietal, and occipital lobes, the accuracy of predictions increases by up to 50% compared to a single channel.

The Limitations of TRIBE v2

Despite its advancements, TRIBE v2 has notable limitations. fMRI measures brain activity only indirectly through blood flow, with a delay of several seconds. The rapid dynamics of neural signals, which are measured in milliseconds, remain inaccessible. Furthermore, the model only covers three sensory channels: smell, touch, and balance are not accounted for.

More fundamentally, TRIBE v2 considers the brain as a passive receiver of sensory inputs, without modeling how it actively makes decisions or drives actions. It also cannot capture developmental changes or clinical conditions, which researchers view as a priority for future versions.

Meta envisions three use cases for the model: planning neuroscience experiments, building AI architectures closer to brain function, and eventually diagnosing brain diseases. The code, model weights, and an interactive demo are all publicly available.