| AI Alignment | Ensuring AI behavior aligns with human goals and values. |
| API Latency | Time delay between request and response in a deployed AI service. |
| AUC (Area Under Curve) | Measure of model’s ability to distinguish between classes. |
| Accuracy | Proportion of correct predictions made by a model. |
| Activation Function | Function that introduces non-linearity into a neural network (e.g., ReLU, Sigmoid). |
| Artificial Intelligence (AI) | Simulation of human intelligence in machines programmed to think and learn. |
| Attention Mechanism | Mechanism that enables models to focus on relevant parts of the input. |
| Backpropagation | Algorithm for updating neural network weights. |
| Batch Normalization | Normalizes activations to improve training speed and stability. |
| Bias | Systematic error introduced by assumptions in data or model. |
| Chain-of-Thought (CoT) | Technique where models generate intermediate reasoning steps. |
| Computer Vision | AI field that enables machines to interpret and make decisions based on visual data. |
| Confusion Matrix | Table showing true vs predicted classifications. |
| Data Privacy | Protection of sensitive user data during model training and usage. |
| Deep Learning | Subset of ML using neural networks with many layers to model complex patterns. |
| Dropout | Regularization technique that randomly drops units in a neural network during training. |
| Embedding | Numerical representation of data, often used for similarity or search. |
| Epoch | One complete pass through the training dataset. |
| Explainable AI (XAI) | Techniques to interpret and understand model predictions. |
| F1 Score | Harmonic mean of precision and recall. |
| Few-shot Learning | Model learns from a few labeled examples. |
| Fine-Tuning | Training a pre-trained model on a specific task or dataset. |
| Generative AI | AI models that can generate new content like text, images, audio, or code. |
| Gradient Descent | Optimization algorithm for minimizing loss function. |
| Hallucination | Generated output that is fluent but factually incorrect. |
| HuggingFace | Ecosystem for pretrained NLP models and transformers. |
| Inference | Running a trained model to make predictions. |
| LangChain | Framework for building LLM-powered applications using composable chains. |
| Large Language Model (LLM) | A transformer-based model trained on large corpora of text data. |
| LlamaIndex | Tool for indexing and querying external data with LLMs. |
| LoRA (Low-Rank Adaptation) | Efficient method for fine-tuning large models with fewer parameters. |
| Loss Function | Function that measures error between predicted and actual values. |
| Machine Learning (ML) | Subset of AI that allows systems to learn from data without explicit programming. |
| Mixture of Experts | Model architecture that routes input through subsets of expert networks. |
| Model Checkpointing | Saving model states during training to resume or analyze progress. |
| Model Fairness | Ensuring model performance does not discriminate against subgroups. |
| Model Serving | Hosting and providing access to ML models for inference. |
| Multimodal AI | Models that process and generate across multiple data types (e.g., text, image, audio). |
| Natural Language Processing (NLP) | AI branch that deals with understanding and generation of human language. |
| Neural Network | Computational model inspired by the human brain, used in deep learning. |
| ONNX | Open Neural Network Exchange; format for model interoperability. |
| Open Weight Model | AI model with publicly available weights for reuse and fine-tuning. |
| Overfitting | Model learns training data too well and fails to generalize. |
| Positional Encoding | Adds information about the position of tokens in sequences to transformer models. |
| Pre-training | Initial training of a model on a large generic dataset. |
| Precision | Proportion of true positive predictions among all positive predictions. |
| Prompt Engineering | Crafting effective prompts to guide LLM responses. |
| Proprietary Model | Model with restricted access, typically hosted and maintained by a company. |
| PyTorch | Popular deep learning library developed by Facebook. |
| ROC Curve | Graphical plot showing performance of classification model. |
| Recall | Proportion of true positives among all actual positives. |
| Reinforcement Learning | Learning method where agents learn by taking actions and receiving rewards. |
| Residual Connection | Skip connections in neural networks that help prevent vanishing gradients. |
| Responsible AI | Ethical and accountable development and deployment of AI. |
| Retrieval-Augmented Generation (RAG) | LLM approach that augments prompts with relevant context from a document store. |
| Scikit-learn | Python library for traditional machine learning. |
| Self-Attention | Mechanism allowing models to weigh importance of different parts of input. |
| Semi-Supervised Learning | Uses a small amount of labeled data with a large amount of unlabeled data. |
| Supervised Learning | Training a model on labeled data. |
| TensorFlow | Google’s open-source deep learning framework. |
| TensorRT | NVIDIA platform for optimizing and deploying deep learning models. |
| Tokenization | Breaking text into smaller units (tokens) for processing. |
| Transformer | Deep learning model architecture that uses self-attention for sequence tasks. |
| Underfitting | Model is too simple to capture underlying patterns in data. |
| Unsupervised Learning | Training a model on data without labels to find patterns. |
| Variance | Model's sensitivity to small fluctuations in training data. |
| Vector Database | Database optimized for storing and querying high-dimensional embeddings. |
| Zero-shot Learning | Model predicts on tasks without having seen labeled examples. |