Technical Lexicon

A mechanism that allows neural networks to focus on specific parts of the input sequence when generating output, enabling long-range dependencies.

A mathematical gate that decides a neuron's output, introducing non-linearity to the network.

A neural network that learns to compress input data into a latent representation and then reconstruct it.

The algorithm used to train neural networks by calculating independent gradients for each weight via the chain rule.

A technique to standardize inputs to a layer for each mini-batch, stabilizing the learning process.

A transformer-based model focusing on understanding context from both left and right directions.

Assumptions built into a model to help it learn effectively (e.g., CNNs assume spatial locality).

A prompting technique enabling LLMs to decompose complex problems into intermediate reasoning steps.

A network architecture specialized for processing grid-like data such as images.

A loss function typically used in classification tasks, measuring the difference between true and predicted distributions.

Increasing the diversity of training data by modifying existing samples.

The component of a Transformer that generates the output sequence.

Generative models that create data by reversing a noise addition process.

Regularization technique dropping random neurons during training.

A continuous vector representation of discrete variables (words, images) where semantic similarity translates to geometric proximity.

The part of a model that processes input data into a context vector or embedding.

One full cycle through the entire training dataset.

Providing a model with a small number of examples (shots) to guide its performance on a new task.

Taking a pre-trained model and training it further on a specific dataset.

An optimization algorithm used to minimize the loss function by iteratively moving in the direction of steepest descent.

Framework where a Generator and Discriminator compete to create realistic data.

A series of decoder-only transformer models developed by OpenAI.

Confident but incorrect outputs generated by an AI.

Parameters set before training (e.g., learning rate) rather than learned.

Using a trained model to make predictions.

transferring knowledge from a large 'teacher' model to a smaller 'student' model.

A parameter-efficient fine-tuning technique that freezes pre-trained weights and injects trainable rank decomposition matrices.

A compressed, abstract representation of data within a model.

Step size for the optimization algorithm.

The raw, unnormalized predictions generated by the last layer of a neural network before Softmax.

A type of RNN capable of learning long-term dependencies, resolving the vanishing gradient problem.

Degradation of generative models when trained on AI-generated data recursively.

AI focused on interaction with human language.

The function the model aims to maximize or minimize (also Loss Function).

Representing categorical variables as binary vectors.

Memorizing training data instead of generalizing.

Internal variables (weights/biases) learned by the model.

The simplest type of feedforward neural network classifier.

The art of crafting inputs (prompts) to guide Large Language Models to produce desired outputs.

The process of reducing the precision of a model's weights (e.g., from 32-bit float to 8-bit integer) to reduce memory usage and increase speed.

A method to align language models with human values by fine-tuning them using a reward model trained on human preferences.

A technique that retrieves relevant external knowledge and feeds it to an LLM to generate more accurate and up-to-date responses.

A CNN architecture using 'skip connections' that allow gradients to flow easily during training, enabling extremely deep networks.

A set of techniques used to prevent overfitting by penalizing complex models.

Network for sequential data processing.

A function that converts a vector of raw scores (logits) into a probability distribution summing to 1.

Empirical observation that model performance improves predictably with model size, data size, and compute.

Mechanism relating different positions of a single sequence to compute a representation of the sequence.

Activation function mapping predictions to 0-1.

Training on labelled data.

A deep learning architecture based entirely on attention mechanisms, dispensing with recurrence and convolutions.

Hyperparameter controlling randomness in LLM generation (High = creative, Low = deterministic).

A multi-dimensional array, the fundamental data structure in ML frameworks like PyTorch/TensorFlow.

The basic unit of text for an LLM (roughly 0.75 words).

Splitting text into tokens.

Applying knowledge from one task to a related one.

Model is too simple to learn the data.

Finding patterns in unlabeled data.

Data used to tuning hyperparameters, separate from training and test sets.

Gradients becoming too small to train deep networks efficiently.

Storage for high-dimensional embeddings.

The strength of connections between neurons.

Performing tasks without specific training examples.