graft-neural-population-transformer-recalibration - SKILL.md Agent Skill

name: graft-neural-population-transformer-recalibration description: "GRAFT: Transformer-based neural population activity model with gain-recalibrated adapters for cross-day BCI recalibration. Separates reusable temporal dynamics from recalibratable neuron interface. Achieves state-of-the-art 0.3866 co-bps on NLB'21 MC Maze. Recalibrates to new datasets by updating only 9.21% of parameters. Supports data-efficient cross-day generalization in brain-computer interfaces." version: 1.0.0 author: Hermes Agent license: MIT metadata: hermes: tags: [neural-population, transformer, bci, cross-day-recalibration, gain-recalibration, neural-activity, mc-maze, nlb-benchmark, adapter, temporal-dynamics] related_skills: [neural-population-decoding, bci-recalibration, transformer-neural-modeling] arxiv_id: "2606.11066v1" paper_title: "GRAFT: Gain-Recalibrated Adapters for Transformer-Based Neural Population Activity Modeling" paper_authors: "Xiangsheng Ge, Yang Xie" paper_date: "2026-06-09"

GRAFT: Transformer-Based Neural Population Activity Modeling with Cross-Day Recalibration

Overview

Neural population activity models recover rich temporal structure from binned spikes, but their read-in and readout layers are traditionally tied to fixed recorded neurons. This coupling limits reuse in long-term brain-computer interfaces where recorded neuron identities, counts, and response statistics change across days.

GRAFT introduces a Transformer-based neural population activity model that separates reusable temporal dynamics from a recalibratable neuron interface, enabling data-efficient cross-day generalization.

Core Innovation: Interface-Backbone Separation

Traditional Approach Limitation

Coupled architecture: Read-in and readout layers fixed to specific neurons
Cross-day problem: Neurons change (identities, counts, statistics)
Solution needed: Separate reusable dynamics from neuron-specific interface

GRAFT Architecture

Shared Backbone: Reusable temporal dynamics (Transformer)
Neuron Interface: Recalibratable read-in and readout
Auxiliary Mechanisms: Gain and positional mechanisms for neural activity

Methodology

Neuron Interface Design

Purpose: Controls how recorded neurons enter and leave the shared backbone

Key Mechanisms:

Gain Mechanism: Neuron-specific gain parameters for scaling
Positional Mechanism: Neuron position encoding in Transformer
Adapter Architecture: Lightweight recalibration layers

Transformer Backbone

Temporal Dynamics Modeling:

Self-attention for long-range dependencies
Position encoding for spike timing
Multi-head attention for population patterns

Gain-Recalibrated Adapter

Cross-Day Recalibration Strategy:

Freeze backbone (temporal dynamics)
Update only adapter parameters (neuron interface)
Minimal parameter changes (9.21%) for new dataset

Results: NLB'21 Benchmark

MC Maze Dataset

Performance: 0.3866 co-bps (ensemble)

New State of the Art on primary co-bps metric
Among public and reported NLB'21 results

Cross-Day Protocol

Dataset Series: MC Maze → Large/Medium/Small variants

Recalibration Results:

MC Maze → Large: 0.3749 co-bps (9.21% parameters)
MC Maze → Medium: 0.3112 co-bps (9.21% parameters)
MC Maze → Small: 0.3152 co-bps (9.21% parameters)

Restricted Support: Target-day support sets used for recalibration

Key Technical Components

1. Gain Mechanism

Purpose: Scale neuron-specific signals

Implementation:

Per-neuron gain parameters
Learned during training
Recalibrated for new neurons
Normalizes across recording sessions

2. Positional Mechanism

Purpose: Encode neuron positions in Transformer

Implementation:

Neuron identity encoding
Population-level position embedding
Supports variable neuron counts

3. Adapter Architecture

Lightweight Design:

Small parameter footprint (9.21% of total)
Rapid recalibration
Preserves backbone dynamics
Enables efficient cross-day adaptation

Applications

1. Long-Term BCI Systems

Problem: Neurons change over days/weeks Solution: Recalibrate interface without retraining backbone Benefit: Stable long-term performance

2. Patient-Specific BCI

Challenge: Inter-subject variability Approach: Train backbone once, recalibrate interface per-patient Advantage: Rapid deployment with minimal data

3. Multi-Session Experiments

Use Case: Same animal, different recording sessions Implementation: Freeze dynamics, adapt to session-specific neurons Efficiency: Avoid full retraining

4. Neural Population Analysis

Temporal Patterns: Rich structure from binned spikes Cross-Session Analysis: Compare dynamics across sessions Behavioral Correlation: Link population patterns to behavior

Comparison with Prior Methods

Traditional Neural Decoders

Aspect	Traditional	GRAFT
Architecture	Coupled	Separated
Cross-Day	Full retrain	Adapter recalibration
Parameters	100% update	9.21% update
Performance	Lower co-bps	0.3866 SOTA

Adapter-Based Methods

Existing adapters: Often for domain adaptation
GRAFT adapter: Specifically for neuron interface
Gain mechanism: Novel for neural population modeling
Positional mechanism: Tailored to spike timing

Technical Implementation

Model Architecture

Input: Binned spikes [T, N]
  ↓
Gain Layer: Neuron-specific scaling [N parameters]
  ↓
Position Encoding: Neuron identity + temporal position
  ↓
Transformer Backbone: Shared temporal dynamics
  ↓
Adapter Layer: Lightweight recalibration
  ↓
Output: Decoded behavior [T, B]

Training Pipeline

Stage 1: Train backbone on source dataset (MC Maze)
Stage 2: Freeze backbone
Stage 3: Train adapters on target dataset
Stage 4: Fine-tune gain parameters

Recalibration Protocol

Requirements:

Small target-day support set
Access to gain parameters
Frozen backbone weights

Steps:

Load pre-trained backbone
Initialize new adapter neurons
Train adapter on support set
Evaluate on target dataset

Extensions and Variations

1. Multi-Modal GRAFT

Integration: Add eye tracking, muscle signals Adapter Design: Multi-modal neuron interface Cross-Modal: Transfer between modalities

2. Real-Time GRAFT

Deployment: Online recalibration Latency: Adapter-only inference Hardware: Neuromorphic implementation

3. Hierarchical GRAFT

Architecture: Multi-scale temporal dynamics Adapters: Hierarchical neuron interfaces Applications: Multi-region recordings

4. Uncertainty Quantification

Bayesian Adapters: Probabilistic neuron interface Ensemble Methods: Multiple adapter samples Confidence: Uncertainty in decoded behavior

Experimental Validation

NLB'21 Benchmark

Task: Neural Latents Benchmark 2021 Dataset: MC Maze (monkey reaching task) Metric: co-bps (bits per second) Performance: 0.3866 co-bps (ensemble)

Cross-Day Datasets

Protocol: MC Maze → Large/Medium/Small Constraint: Restricted support sets Recalibration: 9.21% parameters Performance: Maintained accuracy

Behavioral Correlation

Behavior: Reaching trajectories Neural Activity: Motor cortex spikes Temporal Structure: Movement phases Decoding Quality: co-bps metric

Key Insights

1. Dynamics Preservation

Finding: Temporal dynamics transfer across sessions Implication: Learn once, apply multiple times Validation: Cross-day performance maintained

2. Interface Efficiency

Discovery: Small adapter (9.21%) sufficient Mechanism: Neuron-specific gains capture variability Application: Rapid cross-day deployment

3. Transformer Benefits

Advantage: Self-attention captures long-range dependencies Implementation: Positional encoding for spike timing Result: Rich temporal structure recovery

Future Directions

Research Extensions

Transfer Learning: GRAFT across species
Zero-Shot Recalibration: No target-day data
Online Learning: Continuous adapter updates
Multi-Region: Simultaneous cortical/subcortical

Methodological Advances

Transformer Variants: Sparse attention, efficient transformers
Adapter Types: LoRA, prompt tuning
Gain Mechanisms: Adaptive vs fixed gains
Positional Encoding: Relative vs absolute positions

Clinical Applications

Long-Term BCI: Years of stable performance
Patient-Specific: Rapid calibration for new patients
Rehabilitation: Adaptive decoding during recovery
Prosthetics: Real-time recalibration

References

Original Paper: arXiv:2606.11066v1 (2026-06-09)
NLB'21 Benchmark: Pei et al., 2021
Transformer Architecture: Vaswani et al., 2017
Adapter Methods: Houlsby et al., 2019
Neural Population Decoding: Georgopoulos et al., 1986

Activation Keywords

graft, neural population, transformer, bci recalibration, cross-day, gain adaptation, adapter, temporal dynamics, mc maze, nlb benchmark, neural interface, spike decoding, co-bps, brain-computer interface