Menu
Multi-task GNN learning with Adaptive Checkpointing with Specialization (ACS) to mitigate negative transfer.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Mitigating the effect of activity cliffs using Triplet Soft Margin (TSM) loss on High-Value Activity Cliff Triplets (HV-ACTs).
Running molecular geometry optimization and molecular dynamics (MD) simulations using ASE and MLIPs.
Generating 3D molecular conformations and preparing input files for quantum chemistry DFT calculations.
Constructing Gaussian output neural networks and training deep ensembles to quantify aleatoric and epistemic uncertainty.
Implementing Delta-ML (residual learning between low and high levels of theory) and model transfer learning.
Optimizing molecular feature weights and performing feature selection via Differentiable Information Imbalance (DII).
| name | multitask_learning_with_acs |
| description | Multi-task GNN learning with Adaptive Checkpointing with Specialization (ACS) to mitigate negative transfer. |
Use this skill when training machine learning models on multiple property prediction tasks simultaneously, particularly when labels are sparse, imbalanced, or when tasks exhibit negative transfer.
In Multi-Task Learning (MTL), a shared encoder (e.g., GNN backbone) is optimized based on the average gradients of multiple tasks. When tasks are contradictory, updates for Task A can destroy performance on Task B, a phenomenon known as Negative Transfer. This is amplified in low-data regimes or with imbalanced labels.
ACS addresses Negative Transfer without changing model architecture:
graph TD
Epoch[Epoch Validation Run] --> Eval[Compute Validation Loss for Tasks 1..T]
Eval --> Check1{Task 1 Loss < Best Task 1?}
Eval --> Check2{Task 2 Loss < Best Task 2?}
Check1 -- Yes --> Save1[Save Checkpoint for Task 1]
Check2 -- Yes --> Save2[Save Checkpoint for Task 2]
Execute train_multitask_acs.py to simulate ACS training:
python scripts/train_multitask_acs.py --num_tasks 3 --epochs 20
This script builds a multi-task network and demonstrates how task-specific checkpoints are captured and saved.