spatiotemporal-tdann-mt-direction-maps

name: spatiotemporal-tdann-mt-direction-maps description: Spatiotemporal TDANN framework for modeling the emergence of direction-selective maps in primate MT cortex via self-supervised contrastive optimization with spatial regularization. Unifies ventral and dorsal stream topographic self-organization. arXiv: 2605.11718 (May 2026).

Spatiotemporal TDANN: Self-organized MT Direction Maps

This skill captures the methodology for modeling direction-selective map emergence in primate Middle Temporal (MT) cortex using a spatiotemporal Topographic Deep Artificial Neural Network (TDANN) trained with self-supervised contrastive learning and biologically-inspired spatial regularization.

Paper: Zhaotian Gu, Molan Li, Jie Su, Chang Liu, Tianyi Qian, Dahui Wang, "Self-organized MT Direction Maps Emerge from Spatiotemporal Contrastive Optimization", arXiv:2605.11718 (May 2026)

Core Problem

The spatial and functional organization of the primate visual cortex is a fundamental problem in neuroscience. While TDANN has successfully modeled ventral stream organization (e.g., face patches, object-selective regions), the dorsal stream's topographic organization — specifically direction-selective maps in area MT — has remained unresolved.

Key questions:

• Do MT direction maps emerge from the same universal principles as ventral stream maps?
• What computational trade-offs shape MT tuning properties?
• Can a single framework unify both streams?

Spatiotemporal TDANN Architecture

Model Components

1. 3D ResNet Backbone: Processes naturalistic video input (spatial + temporal dimensions)
2. Momentum Contrast (MoCo): Self-supervised contrastive learning paradigm
3. Biological Spatial Loss: Enforces cortical-like topographic organization
4. Direction Selectivity Readout: Extracts motion tuning properties

Architecture Design

# Conceptual architecture
class SpatiotemporalTDANN:
    """3D ResNet with MoCo contrastive learning + spatial regularization."""
    
    def __init__(self, backbone="resnet3d-18", spatial_loss_weight=1.0):
        # 3D ResNet processes video clips (B, T, C, H, W)
        self.backbone = ResNet3D(backbone)
        
        # MoCo contrastive learning components
        self.encoder_q = self.backbone  # Query encoder
        self.encoder_k = copy(self.backbone)  # Key encoder (momentum updated)
        self.queue = FeatureQueue(size=65536, dim=feature_dim)
        
        # Spatial regularization loss
        self.spatial_loss_weight = spatial_loss_weight
    
    def forward(self, video_clips):
        # Extract spatiotemporal features
        features = self.encoder_q(video_clips)  # (B, T, H', W', D)
        
        # Apply spatial regularization
        if self.training:
            spatial_loss = self.compute_spatial_loss(features)
        
        return features, spatial_loss

MoCo Contrastive Learning

def moco_contrastive_loss(query_features, key_features, queue, temperature=0.07):
    """Momentum Contrast loss for self-supervised video representation learning."""
    # Positive pair: query and key from same video (different augmentations)
    pos_logits = torch.einsum('nc,nc->n', [query_features, key_features]) / temperature
    
    # Negative pairs: query vs. all features in queue
    neg_logits = torch.einsum('nc,ck->nk', [query_features, queue.features]) / temperature
    
    logits = torch.cat([pos_logits.unsqueeze(1), neg_logits], dim=1)
    labels = torch.zeros(logits.shape[0], dtype=torch.long)
    
    return cross_entropy(logits, labels)

Biological Spatial Loss

The spatial loss enforces cortical-like organization:

def spatial_regularization_loss(features, spatial_distance_matrix):
    """
    Encourages neurons with similar functional preferences to be spatially close,
    mimicking cortical topographic organization.
    
    features: (B, H', W', D) - feature maps from backbone
    spatial_distance_matrix: precomputed spatial distances between grid positions
    """
    # Compute functional similarity between all neuron pairs
    functional_similarity = cosine_similarity(features, features)
    
    # Penalize when functionally similar neurons are spatially far apart
    spatial_loss = torch.mean(
        functional_similarity * spatial_distance_matrix
    )
    
    return spatial_loss

Key Findings

1. Spontaneous Emergence of MT-like Maps

Training the spatiotemporal TDANN on naturalistic videos leads to:

• Direction-selective columns: Neurons preferring similar motion directions cluster together
• Pinwheel structures: Topological singularities where direction preference rotates 360° around a point
• Direction preference diversity: Full coverage of motion directions across the map

2. MT Tuning Properties Match In Vivo Data

The model's neurons exhibit physiological properties matching macaque MT recordings:

3. Optimization Trade-off Discovery

Crucial insight: MT tuning properties emerge from a strict trade-off between:

• Task-driven discriminative pressure: The contrastive loss pushes neurons to distinguish different motion patterns
• Spatial regularization: The biological loss encourages spatially organized topography

This trade-off explains why MT neurons show:

• Strong direction selectivity (from discriminative pressure)
• Residual axial component (spatial regularization constraint)

4. Unified Cortical Self-Organization

The same principles govern both:

• Ventral stream: Object category maps, face patches
• Dorsal stream: Direction-selective maps, motion columns

This establishes a general mechanism for cortical self-organization across the entire visual cortex.

Implementation Patterns

Training Pipeline

def train_spatiotemporal_tdann(
    video_dataset,
    spatial_loss_weight=0.1,
    moco_temperature=0.07,
    queue_size=65536,
    num_epochs=100,
    lr=1e-4,
):
    model = SpatiotemporalTDANN(spatial_loss_weight=spatial_loss_weight)
    optimizer = torch.optim.AdamW(model.parameters(), lr=lr)
    queue = FeatureQueue(size=queue_size)
    
    for epoch in range(num_epochs):
        for video_clips in video_dataset:
            # Generate augmented views
            view_q = augment(video_clips, view="query")
            view_k = augment(video_clips, view="key")
            
            # Forward pass
            q_features, spatial_loss = model(view_q)
            with torch.no_grad():
                k_features = model.encoder_k(view_k)
            
            # Contrastive loss
            contrastive_loss = moco_contrastive_loss(
                q_features, k_features, queue, moco_temperature
            )
            
            # Combined loss
            total_loss = contrastive_loss + spatial_loss
            
            optimizer.zero_grad()
            total_loss.backward()
            optimizer.step()
            
            # Update momentum encoder
            momentum_update(model.encoder_k, model.encoder_q, m=0.999)
            queue.update(k_features)

Analyzing Direction Selectivity

def compute_direction_selectivity(neuron_responses, directions):
    """
    Compute direction selectivity index (DSI) and preferred direction.
    
    neuron_responses: (n_trials, n_neurons) - response to moving stimuli
    directions: (n_trials,) - stimulus direction in degrees
    """
    # Vector sum method
    cos_sum = torch.sum(neuron_responses * torch.cos(torch.deg2rad(directions)), dim=0)
    sin_sum = torch.sum(neuron_responses * torch.sin(torch.deg2rad(directions)), dim=0)
    
    # Preferred direction
    pref_direction = torch.atan2(sin_sum, cos_sum) * 180 / torch.pi
    
    # Direction Selectivity Index
    response_vector_length = torch.sqrt(cos_sum**2 + sin_sum**2)
    total_response = torch.sum(neuron_responses, dim=0)
    dsi = response_vector_length / (total_response + 1e-8)
    
    return pref_direction, dsi

def compute_circular_variance(neuron_responses, directions):
    """Compute circular variance - lower means more selective."""
    pref_dir, dsi = compute_direction_selectivity(neuron_responses, directions)
    return 1 - dsi  # Circular variance = 1 - DSI

def detect_pinwheels(direction_map):
    """
    Detect pinwheel singularities in direction preference map.
    
    direction_map: (H, W) - preferred direction at each spatial location
    """
    # Compute phase winding around each 2x2 block
    pinwheel_map = torch.zeros_like(direction_map, dtype=torch.bool)
    
    for i in range(direction_map.shape[0]-1):
        for j in range(direction_map.shape[1]-1):
            block = direction_map[i:i+2, j:j+2]
            # Unwrap phases and compute total change around loop
            phase_diff = unwrap_phase_differences(block)
            winding = torch.sum(phase_diff)
            
            # Pinwheel if winding number ≈ ±2π
            pinwheel_map[i, j] = abs(winding) > torch.pi
    
    return pinwheel_map

Experimental Setup

Dataset

• Naturalistic videos: Unlabeled video clips from natural scenes
• Augmentation: Temporal cropping, spatial jitter, color jitter, motion blur
• Evaluation stimuli: Drifting gratings, random dot kinematograms (RDKs)

Hyperparameters

Validation Against In Vivo Data

To validate the model against biological data:

1. Direction Selectivity Index (DSI): Compare distribution of DSI values
2. Circular Variance: Compare tuning sharpness distributions
3. Pinwheel Density: Count pinwheels per mm², compare to ~3.14/mm² (universal constant)
4. Axial Component: Measure residual sensitivity to orthogonal motion
5. Orientation-Direction Correlation: Analyze joint tuning properties

When to Use This Framework

• Studying cortical self-organization: Understanding how topographic maps emerge
• Modeling dorsal stream: Motion processing, direction selectivity
• Unifying ventral and dorsal streams: Single framework for both pathways
• Testing developmental hypotheses: What constraints shape cortical maps
• Neuromorphic vision systems: Bio-inspired motion detection architectures

Related Skills

• spatiotemporal-tdann — broader spatiotemporal TDANN methodology
• kuramoto-phase-encoding — neuro-inspired phase encoding for vision transformers
• neural-code-language-characterization — automated neuron characterization
• eeg-visual-attention-decoding — visual attention decoding from neural signals

Key Insights

1. Universal self-organization principle: Same computational principles govern both ventral and dorsal stream topography
2. Trade-off shapes tuning: MT properties emerge from balance between discriminative pressure and spatial regularization
3. Self-supervised sufficiency: No supervised labels needed — naturalistic videos + contrastive learning produce brain-like maps
4. Pinwheel density as invariant: Model reproduces the ~3.14/mm² pinwheel density found in biological cortex
5. Direction + axial residual: Strong direction selectivity with residual axial component matches in vivo recordings
6. 3D ResNet + spatial loss: The combination of temporal processing and topographic regularization is key
7. Predictive power: Framework can generate testable predictions about MT organization under different constraints