api: YAML JSON TOON Database
version: 1.0.0
format: json
dataset:
  id: 29
  slug: ml-model-architectures
  title: Machine Learning Model Architectures
  description: "Overview of major ML model families: feedforward neural networks, CNNs, RNNs/LSTMs, Transformers, GANs, and when to use each."
  summary: "Machine-learning model architectures with type, structure, layers, uses, pros, cons and notes: feedforward networks, CNNs, RNNs, LSTMs, transformers, GANs and autoencoders, with a selection guide by data type."
  use_when[2]: "You are choosing an architecture for a data type such as images, text or time series.",You need the strengths and weaknesses of each.
  caveats[2]: "Guidance is general; strong results often come from architecture, data and training choices together.","Architectures are combined in practice (convolutional layers feeding transformers, for example)."
  related[4]: neural-network-types,attention-mechanisms,ml-training-algorithms,transfer-learning
  category: AI & ML Terms
  category_slug: ai-ml-terms
  tags: "machine-learning,neural-networks,architectures,cnn,rnn,transformer,gan"
  view_count: 0
  created_at: 1777673262
  updated_at: 1777673262
  entry_count: 7
  fields[14]{section,field,type,example}:
    architectures,name,string,Feedforward Neural Network (FNN)
    architectures,type,string,Dense / MLP
    architectures,structure,string,Input → Hidden layers (dense) → Output. No cycl…
    architectures,layers,array,"[3 items]"
    architectures,use_cases,array,"[3 items]"
    architectures,pros,array,"[3 items]"
    architectures,cons,array,"[2 items]"
    architectures,notes,string,Also called Multi-Layer Perceptron (MLP); found…
    architectures,key_layers,object,"{convolution,pooling,fully_connected}"
    architectures,variants,array|object,"[3 items]"
    architectures,gates,object,"{forget_gate,input_gate,output_gate}"
    architectures,core_mechanism,string,"Self-attention: each token attends to all other…"
    architectures,training,string,"Alternating: train D on real/fake, then train G…"
    selection_guide,*,object,"{tabular_data,images,sequences_short,sequences_long}"
data:
  architectures[7]:
    - name: Feedforward Neural Network (FNN)
      type: Dense / MLP
      structure: Input → Hidden layers (dense) → Output. No cycles or feedback connections.
      layers[3]: Input layer,One or more hidden layers (ReLU/common),Output layer (linear/softmax/sigmoid)
      use_cases[3]: Tabular data classification/regression,Simple pattern recognition,Baseline model
      pros[3]: Simple to implement,Universal approximator (with enough width/depth),Fast inference
      cons[2]: Ignores spatial/temporal structure,Parameter count grows quickly with input size
      notes: Also called Multi-Layer Perceptron (MLP); foundation for all deep nets
    - name: Convolutional Neural Network (CNN)
      type: Spatial hierarchy
      structure: "Convolutional layers (filters) → pooling → fully connected. Local receptive fields, weight sharing."
      key_layers:
        convolution: "Applies filters/kernels to detect local patterns (edges, textures)"
        pooling: Downsampling (max/avg) for translation invariance
        fully_connected: Final classification/regression layers
      use_cases[4]: Image classification,Object detection,Medical imaging,Time series (1D conv)
      pros[3]: Parameter sharing (efficient),Spatial invariance,Detects hierarchical features
      cons[2]: Fixed receptive field (unless dilated),Not inherently good for sequential data
      notes: "Architecture depth (ResNet, Inception) matters more than width for vision"
    - name: Recurrent Neural Network (RNN)
      type: Sequential / temporal
      structure: Recurrent connections; hidden state carries information across time steps.
      variants[3]: Vanilla RNN (simple recurrence),LSTM (Long Short-Term Memory),GRU (Gated Recurrent Unit)
      use_cases[4]: Language modeling,Time series prediction,Speech recognition,Sequence tagging
      pros[2]: Handles variable-length sequences,Theoretically infinite memory (with LSTM/GRU)
      cons[3]: Sequential processing (hard to parallelize),Vanishing/exploding gradients (vanilla RNN),LSTM/GRU complex
      notes: LSTM/GRU gates mitigate vanishing gradients; still largely superseded by Transformers for NLP
    - name: Long Short-Term Memory (LSTM)
      type: Recurrent with gates
      structure: Cell state + forget/input/output gates. Controls information flow over long sequences.
      gates:
        forget_gate: Decide what to discard from cell state
        input_gate: Decide what new information to store
        output_gate: Decide what to output based on cell state
      use_cases[3]: Machine translation (legacy),Speech synthesis,Text generation (pre-Transformer)
      pros[2]: Can learn long-range dependencies (~100s steps),Robust to vanishing gradients
      cons[3]: Still sequential (slow training),Many parameters (4× vanilla RNN),Requires more data than simpler models
      notes: LSTM superior to vanilla RNN but slower; Transformer now standard for sequence
    - name: Transformer
      type: Attention-based
      structure: Encoder-decoder or encoder-only/decoder-only. Self-attention + feed-forward. Positional encoding.
      core_mechanism: "Self-attention: each token attends to all others; O(n²) complexity"
      variants:
        encoder_only: "BERT (masked attention, pre-training)"
        decoder_only: "GPT (causal attention, autoregressive)"
        "encoder-decoder": "T5, original Transformer (translation)"
      use_cases[4]: NLP (all major tasks),Computer vision (Vision Transformer),Audio (Whisper),Multimodal (CLIP)
      pros[3]: Parallelizable (no recurrence),Long-range dependencies (attention),Scalable with data/model size
      cons[3]: Quadratic memory in sequence length,Requires massive data for training from scratch,Positional encoding limits context
      notes: Dominant architecture since 2017 (Attention Is All You Need). Sparse attention variants handle long sequences.
    - name: Generative Adversarial Network (GAN)
      type: Generative
      structure: Generator (creates fake samples) + Discriminator (distinguishes real vs fake). Min-max game.
      training: "Alternating: train D on real/fake, then train G to fool D. No explicit likelihood."
      use_cases[4]: Image synthesis (StyleGAN),Image-to-image translation,Super-resolution,Data augmentation
      pros[2]: "High-quality, sharp samples",No explicit likelihood modeling (flexible)
      cons[3]: "Training unstable (mode collapse, vanishing gradients)",Hard to evaluate,No explicit density
      notes: Diffusion models now dominate image generation; GANs still used for specific tasks (style transfer)
    - name: Autoencoder
      type: Representation learning
      structure: Encoder (compress to latent) → bottleneck → Decoder (reconstruct). Reconstruction loss trains end-to-end.
      variants[3]: Vanilla AE (dense layers),Variational Autoencoder (VAE) — probabilistic latent,Denoising AE — robust features
      use_cases[4]: Dimensionality reduction,Anomaly detection,Denoising,Generative modeling (VAE)
      pros[2]: Unsupervised pre-training possible,Learns compact representations
      cons[3]: VAE samples blurry,Simple AE doesn't learn useful latent space,Reconstruction loss not always semantic
      notes: Contrastive learning (SimCLR) often superior for representation learning
  selection_guide:
    tabular_data: FNN or Gradient Boosting (XGBoost/LightGBM)
    images: "CNN (ResNet, EfficientNet) or Vision Transformer"
    sequences_short: LSTM/GRU or Transformer (if enough data)
    sequences_long: Transformer with sparse attention or linear attention
    text_nlp: "Transformer (BERT for understanding, GPT for generation)"
    time_series: "Temporal CNN, LSTM, or Transformer (Informer, PatchTST)"
    image_generation: Diffusion model (Stable Diffusion) or GAN (legacy)
    anomaly_detection: Autoencoder or isolation forest