The following table compares major categories of Machine Learning, Deep Learning, Generative AI, and Reinforcement Learning models.
| Category | Model Type | Core Purpose / Characteristic | Ideal Input Data Type | Training Paradigm | Popular Examples |
|---|---|---|---|---|---|
| Traditional ML | Linear Models | Assumes linear relationships between features. | Structured/Tabular (Numbers) | Supervised | Linear Regression, Logistic Regression |
| Tree-Based Models | Splits data like flowchart branches based on values. | Structured/Tabular (Mixed) | Supervised | Decision Trees, Random Forest, XGBoost | |
| Distance-Based | Classifies data points based on geometric proximity. | Structured/Tabular (Normalized) | Supervised | K-Nearest Neighbors, SVM | |
| Probabilistic | Uses probability theory and Bayes' Theorem. | Structured, Text (Word counts) | Supervised | Naive Bayes, Hidden Markov Models | |
| Clustering | Unsupervised grouping of similar unlabeled points. | Structured/Tabular | Unsupervised | K-Means, DBSCAN | |
| Dimensionality | Compresses datasets by reducing redundant features. | High-Dimensional Tabular | Unsupervised | PCA, t-SNE | |
| RNNs & Sequence | Vanilla RNN | Processes sequences step-by-step with memory. | Sequential (Text, Time-Series) | Supervised/Self-Sup. | Standard Elman RNN |
| LSTM | Retains long-term context using gating mechanisms. | Sequential (Text, Audio, Sensors) | Supervised/Self-Sup. | Standard LSTM, BiLSTM | |
| GRU | Streamlined, faster version of LSTM with fewer gates. | Sequential (Text, Audio, Sensors) | Supervised/Self-Sup. | Standard GRU | |
| CNNs (Spatial) | Image Class. | Identifies the main subject within a static frame. | Spatial Grids (Images, Videos) | Supervised | ResNet, VGG16, MobileNet |
| Object Detection | Locates and labels multiple distinct items in space. | Spatial Grids (Images, Videos) | Supervised | YOLO, Faster R-CNN | |
| Segmentation | Classifies every single individual pixel. | Spatial Grids (Medical scans) | Supervised | U-Net, Mask R-CNN | |
| Transformers | Encoder-Only | Extracts context and meaning from sequences. | Sequential (Text, Code) | Self-Supervised | BERT, RoBERTa |
| Decoder-Only | Predicts the next sequence element autoregressively. | Sequential (Text, Code) | Self-Supervised | GPT-4, Llama 3, Claude 3.5 | |
| Encoder-Decoder | Translates/maps one sequence onto another. | Sequential (Source Text) | Self-Supervised | T5, BART | |
| Generative AI | Multimodal | Processes and outputs multiple mediums natively. | Mixed (Text, Image, Video, Audio) | Self-Supervised | Google Gemini, GPT-4o |
| Diffusion Models | Generates media by removing noise iteratively. | Text prompts, Random noise | Supervised (Latent) | Stable Diffusion, Midjourney, Sora | |
| GANs | Two networks compete to create realistic data. | Random noise vectors, Images | Unsupervised/Adverserial | StyleGAN, CycleGAN | |
| VAEs | Compresses data down and decodes new variants. | Images, Structured vectors | Unsupervised | Beta-VAE | |
| Reinforcement | Value-Based RL | Finds actions by calculating future rewards. | Environment States, Screen pixels | Trial-and-error Reward | Deep Q-Networks (DQN) |
| Policy-Based RL | Directly learns behaviors for a given environment. | Environment States, Screen pixels | Trial-and-error Reward |
No comments:
Post a Comment