AI Model Training

Teaching an AI to defend itself by constantly attacking it with tricky or malicious inputs during training.

A training technique where models are exposed to adversarial examples, inputs deliberately crafted to fool the model, to improve robustness. Widely used to harden AI systems against malicious attacks.

Why it matters: Critical for security-sensitive AI (self-driving cars, facial recognition) to prevent hacks via manipulated inputs.

Full definition: Adversarial Training →

A neural network that learns to compress data into a small code and then unzip it back to the original.

A type of neural network trained to compress input data into a compact latent representation (encoder) and then reconstruct the original input (decoder). Used for dimensionality reduction, anomaly detection, and generative modeling.

Why it matters: Excellent for unsupervised learning tasks like cleaning noisy images or finding anomalies.

Full definition: Autoencoders →

The algorithm that teaches neural networks by calculating how wrong each neuron was and adjusting it backward through the layers.

An optimization algorithm that computes the gradient of the loss function with respect to each weight by propagating errors backward through the network. The foundation of neural network training.

Why it matters: Without backpropagation, deep learning wouldn't exist, it's the mathematical engine behind all neural network learning.

Full definition: Backpropagation →

An AI system that keeps learning and improving from new data after deployment, instead of being frozen at launch.

Also called online learning or lifelong learning. A training paradigm where models are updated incrementally with new data in production, rather than requiring full retraining. Includes safeguards against catastrophic forgetting.

Why it matters: Critical for domains where data distributions shift rapidly, like fraud detection, recommendation systems, and market analysis.

Full definition: Continuous Learning →

Teaching a small, fast AI model to mimic a large, expensive one, so you get similar results at a fraction of the cost.

A technique where a smaller "student" model is trained to replicate the behavior of a larger "teacher" model. The student learns from the teacher's soft probability outputs, not just hard labels.

Why it matters: Enables deploying AI on edge devices and reducing inference costs while maintaining quality.

Full definition: Distillation (Model Distillation) →

Randomly turning off some neurons during training so the AI doesn't over-memorize and can generalize better.

A regularization technique for neural networks that randomly deactivates a fraction of neurons on each training step, so the network cannot lean too heavily on any single neuron or path. This forces it to learn redundant, more general representations. Dropout is applied only during training; at inference all neurons are active, with their outputs scaled to compensate.

Why it matters: Dropout is one of the simplest and most effective defenses against overfitting, where a model memorizes its training data but fails on new inputs. It costs almost nothing to add and routinely improves how well a model generalizes to the real world.

Full definition: Dropout →

One complete pass through the entire training dataset, the AI sees every example once per epoch.

A single iteration over the entire training dataset during model training. Multiple epochs are typically needed for the model to converge. Too many epochs can lead to overfitting.

Why it matters: A fundamental unit of training progress, monitoring loss across epochs tells you if the model is learning.

Full definition: Epoch →

Taking a pre-trained AI model and teaching it your specific domain knowledge, like hiring a generalist and training them on your business.

The process of further training a pre-trained model on a smaller, domain-specific dataset to specialize it for a particular task. Adjusts the model's weights to perform better in a specific context.

Why it matters: The primary mechanism for making general-purpose AI models useful for specific business applications.

Full definition: Fine-Tuning →

The AI learning process, adjusting its dials a tiny bit at a time, always moving toward less error, like rolling a ball downhill.

An optimization algorithm that iteratively adjusts model parameters in the direction that reduces the loss function. Variants include SGD, Adam, and AdaGrad. The fundamental mechanism by which neural networks learn.

Why it matters: The core algorithm that makes all neural network training possible, the engine under every deep learning model.

Full definition: Gradient Descent →

Adjusting the "settings" of an AI model (like learning speed or network size) to find the best performance.

The process of optimizing the configuration parameters that control the training process itself (learning rate, batch size, architecture choices). These are set before training begins, unlike model weights.

Why it matters: Can make the difference between a mediocre model and a world-class one, often overlooked in favor of data or architecture.

Full definition: Hyperparameter Tuning →

Transferring the intelligence of a large, expensive AI model into a smaller, cheaper one that can run anywhere.

A compression technique where a compact "student" model learns to reproduce the behavior of a larger "teacher" model. The student learns from soft probability distributions rather than hard labels.

Why it matters: Enables running enterprise-grade AI on edge devices, reducing cloud costs by 10-100x.

Full definition: Knowledge Distillation →

A clever shortcut for fine-tuning AI models, adjusting only a tiny fraction of the weights to save time and money.

A parameter-efficient fine-tuning technique that adds small, trainable low-rank matrices to existing model layers instead of updating all weights. Reduces training compute and memory by 90%+.

Why it matters: Made fine-tuning foundation models accessible to companies without massive GPU budgets.

Full definition: LoRA (Low-Rank Adaptation) →

The AI's scorecard, a formula that measures how wrong the model's predictions are, guiding it to improve.

A mathematical function that quantifies the difference between a model's predictions and the actual target values. The model's training objective is to minimize this function. Common examples include MSE and cross-entropy.

Why it matters: The choice of loss function directly shapes what the model optimizes for, choose wrong and it learns the wrong thing.

Full definition: Loss Function →

When an AI memorizes the training data too well and fails on new data, like a student who memorizes answers but can't solve new problems.

A modeling error where a model learns the training data too precisely, including its noise and outliers, resulting in poor generalization to unseen data. Addressed through regularization, dropout, and cross-validation.

Why it matters: The most common failure mode in machine learning, a model that performs great in testing but fails in production.

Full definition: Overfitting →

Fine-tuning a foundation model by updating only a small fraction of its parameters, faster, cheaper, and nearly as good.

A family of techniques (LoRA, QLoRA, adapters) that enable fine-tuning large models by modifying only a small subset of parameters. Dramatically reduces compute and memory requirements.

Why it matters: Made it possible for companies to customize billion-parameter models on a single GPU.

Full definition: Parameter-Efficient Fine-Tuning (PEFT) →

The initial, massive training phase where an AI model learns general knowledge from enormous datasets before being specialized.

The first phase of training a foundation model on a large, diverse dataset to learn general patterns, language understanding, or visual features. Followed by fine-tuning for specific tasks.

Why it matters: The most expensive and resource-intensive phase of AI development, costing tens of millions of dollars for frontier models.

Full definition: Pre-training →

Training an AI to be more helpful and less harmful by having humans rate its outputs and feeding that feedback back into training.

A training technique where human preferences are used to fine-tune language models. Human evaluators rank model outputs, and a reward model is trained on these preferences to guide further training.

Why it matters: The technique that made ChatGPT conversational and helpful, the key innovation in aligning LLMs to human intent.

Full definition: Reinforcement Learning from Human Feedback (RLHF) →

Training an AI on unlabeled data by having it predict missing parts of the data, like a fill-in-the-blank quiz at scale.

A training paradigm where the model creates its own labels from the structure of the data. Examples include masked language modeling (BERT) and next-token prediction (GPT). Eliminates the need for manual labeling.

Why it matters: The reason foundation models are possible, no human could label the trillions of examples needed to train GPT-4.

Full definition: Self-Supervised Learning →

Taking an AI trained on one task and reusing its knowledge for a different but related task, like a doctor learning veterinary medicine.

A technique where a model trained on one task is repurposed for a different but related task. The model transfers learned representations, reducing the need for task-specific training data.

Why it matters: The reason we don't train models from scratch for every task, dramatically reduces data requirements and training time.

Full definition: Transfer Learning →

When an AI model is too simple to capture the patterns in the data, like trying to draw a curve with a straight line.

A modeling error where a model is too simple to capture the underlying patterns in the training data, resulting in poor performance on both training and test data.

Why it matters: The opposite of overfitting, often solved by using more complex models, more features, or more training time.

Full definition: Underfitting →