| name | cognition-inspired-dual-stream-emotion |
| description | Cognition-Inspired Dual-Stream Semantic Enhancement (DuSE) for Vision-Based Dynamic Emotion Modeling. Implements hierarchical temporal prompt clusters (HTPC) for cognitive priming and latent semantic emotion aggregators (LSEA) for knowledge integration. Models neuro-cognitive mechanisms from Conceptual Act Theory for dynamic facial expression recognition. Use for: emotion recognition, cognitive-inspired computer vision, neuro-cognitive modeling, dynamic facial expression analysis. |
| version | v1.0.0 |
| last_updated | 2026-04-15T00:00:00.000Z |
| paper | https://arxiv.org/abs/2604.12777v1 |
Cognition-Inspired Dual-Stream Emotion Modeling (DuSE)
Source Paper: Cognition-Inspired Dual-Stream Semantic Enhancement for Vision-Based Dynamic Emotion Modeling
Authors: Huanzhen Wang, Ziheng Zhou, Zeng Tao, Aoxing Li, Yingkai Zhao
Published: 2026-04-14
Overview
Models human brain emotion perception through dynamic, hierarchical integration of sensory input with semantic and contextual knowledge. Implements neuro-cognitive mechanisms from Conceptual Act Theory for dynamic facial expression recognition.
Biological Inspiration
The human brain constructs emotional percepts by:
- Cognitive Priming: Linguistic cues pre-sensitize neural pathways (HTPC stream)
- Knowledge Integration: Sensory inputs synthesized with learned concepts (LSEA stream)
- Hierarchical Processing: Reflecting hippocampus and default mode network roles
Dual-Stream Architecture
Stream 1: Hierarchical Temporal Prompt Cluster (HTPC)
Purpose: Operationalize cognitive priming effect
Text Semantics ──► Fine-grained Temporal Features ──► Modulated Visual Processing
Mechanism:
- Linguistic cues pre-sensitize neural pathways
- Modulate processing of incoming visual stimuli
- Align textual semantics with temporal facial dynamics
Stream 2: Latent Semantic Emotion Aggregator (LSEA)
Purpose: Model knowledge integration (Conceptual Act Theory)
Sensory Inputs ──► Learned Conceptual Knowledge ──► Coherent Emotional Experience
Mechanism:
- Aggregates sensory inputs across modalities
- Synthesizes with learned conceptual knowledge
- Reflects hippocampus and default mode network function
Architecture Diagram
┌─────────────────────────────────────────────────────────────────┐
│ Dual-Stream DuSE Model │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────────┐ ┌─────────────────┐ │
│ │ HTPC Stream │ │ LSEA Stream │ │
│ │ (Cognitive │ │ (Knowledge │ │
│ │ Priming) │ │ Integration) │ │
│ │ │ │ │ │
│ │ Text Semantics │ │ Sensory Inputs │ │
│ │ │ │ │ │ │ │
│ │ ▼ │ │ ▼ │ │
│ │ Fine-grained │ │ Learned │ │
│ │ Temporal Feats │ │ Conceptual │ │
│ │ │ │ │ Knowledge │ │
│ │ ▼ │ │ │ │ │
│ │ Modulate Visual │ │ ▼ │ │
│ │ Processing │ │ Coherent │ │
│ └────────┬────────┘ │ Emotion Exp. │ │
│ │ └────────┬────────┘ │
│ │ │ │
│ └───────────────┬──────────────────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────┐ │
│ │ Fusion & │ │
│ │ Decision │ │
│ └───────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘
Key Features
1. Neuro-Cognitive Plausibility
Explicitly models:
- Prefrontal-limbic interactions in emotion processing
- Hierarchical temporal dynamics across brain regions
- Conceptual Act Theory mechanisms
2. Semantic Context Integration
Unlike traditional approaches that process facial expressions in isolation:
- Integrates temporal dynamics with semantic knowledge
- Context-aware emotion modeling
- Cross-modal (text + vision) alignment
3. Improved Interpretability
Neuro-cognitive grounding provides:
- Clear mapping to known brain mechanisms
- Interpretable attention weights
- Modularity aligned with brain function
Workflow
Phase 1: HTPC Processing
class HTPC(nn.Module):
"""
Hierarchical Temporal Prompt Cluster
Models cognitive priming effect
"""
def __init__(self):
self.semantic_encoder = TextEncoder()
self.temporal_aligner = TemporalAlignmentModule()
def forward(self, text_input, temporal_features):
semantic_repr = self.semantic_encoder(text_input)
modulated_features = self.temporal_aligner(
semantic_repr, temporal_features
)
return modulated_features
Phase 2: LSEA Processing
class LSEA(nn.Module):
"""
Latent Semantic Emotion Aggregator
Models knowledge integration
"""
def __init__(self):
self.sensory_encoder = VisualEncoder()
self.conceptual_knowledge = ConceptMemory()
def forward(self, visual_input):
sensory_repr = self.sensory_encoder(visual_input)
conceptual_repr = self.conceptual_knowledge(sensory_repr)
emotional_experience = self.aggregate(
sensory_repr, conceptual_repr
)
return emotional_experience
Phase 3: Dual-Stream Fusion
class DuSE(nn.Module):
"""
Dual-Stream Semantic Enhancement
"""
def __init__(self):
self.htpc = HTPC()
self.lsea = LSEA()
self.fusion = AttentionFusion()
def forward(self, visual_input, text_input):
htpc_output = self.htpc(text_input, visual_input)
lsea_output = self.lsea(visual_input)
output = self.fusion(htpc_output, lsea_output)
return output
Results
- Benchmarks: State-of-the-art on challenging in-the-wild datasets
- Performance: Superior to vision-only approaches
- Interpretability: Enhanced through neuro-cognitive alignment
Activation Keywords
- dual stream emotion recognition
- cognition inspired vision model
- conceptual act theory implementation
- hierarchical temporal prompt cluster
- latent semantic emotion aggregator
- neuro cognitive emotion modeling
- dynamic facial expression recognition
References