deep-learning-expert

Deep Learning Modeling and Optimization

Description

Provide deep learning model design, training optimization, problem diagnosis, and engineering practice guidance to ensure models meet requirements for accuracy, efficiency, and deployability.

When to Use

• User requests "design a deep learning model" or "how to train XX task"
• User asks about training issues (overfitting, underfitting, gradient vanishing/explosion, NaN, slow convergence)
• User needs paper reproduction or cutting-edge technology analysis (Transformer, Diffusion, ViT, etc.)
• User seeks framework usage recommendations (PyTorch, TensorFlow, JAX) or code implementation
• User asks about best practices in CV, NLP, reinforcement learning domains
• User needs model compression, quantization, deployment, or inference optimization solutions
• User requests data processing, data augmentation, or sampling strategy recommendations
• User compares different model architectures or training strategies

When NOT to Use

• User only needs traditional machine learning methods (e.g., decision trees, SVM, linear regression) without deep learning
• User's problem is pure data analysis or statistics without model training
• User only needs data preprocessing or feature engineering without model design
• User's problem is algorithm competition or pure programming without deep learning modeling
• User needs MLOps platform setup or team management rather than specific technical solutions

Input

{
  task: {
    type: string                  // Task type (classification/detection/segmentation/generation/sequence-modeling/RL)
    domain: string                // Application domain (CV/NLP/audio/multimodal/time-series)
    specificTask?: string         // Specific task (e.g., image-classification → CIFAR-10)
  }
  data: {
    size?: string                 // Dataset size (e.g., "10k samples")
    quality?: string              // Data quality description (noise, annotation quality)
    distribution?: string         // Class distribution (balanced/long-tail)
    labelAvailability?: string    // Labeling status (fully-supervised/weakly-supervised/unsupervised)
    augmentationNeeded?: boolean  // Whether data augmentation is needed
  }
  constraints: {
    computeResources?: string     // Compute resources (single-GPU/multi-GPU/TPU/CPU)
    latencyRequirement?: string   // Latency requirement (e.g., "inference <50ms")
    accuracyTarget?: string       // Accuracy target (e.g., "top-1 acc >90%")
    modelSizeLimit?: string       // Model size limit (e.g., "<100MB")
    budget?: string               // Budget constraints (affects pretrained model selection/compute resources)
  }
  challenges?: string[]           // Known challenges (few-shot/domain-shift/real-time/long-tail)
  existingSetup?: {
    framework?: string            // Existing framework (PyTorch/TensorFlow/JAX)
    baseModel?: string            // Existing base model
    trainingIssues?: string[]     // Current training issues
  }
}