Synaplume

Glossary

Every important AI and ML term, explained in one breath. Search or browse.

53 terms

Activation Function
A nonlinear function (like ReLU or sigmoid) applied to a neuron's output. Without it, stacked layers would collapse into a single linear model.
Deep Learning
Agent
A system that perceives its environment and takes actions to achieve goals. The unifying abstraction of AI — from thermostats to LLMs with tools.
Artificial Intelligence
AGI (Artificial General Intelligence)
A hypothetical AI with human-level flexibility across arbitrary domains, as opposed to today's narrow systems that excel only at specific tasks.
Artificial Intelligence
Attention
A mechanism letting every element of a sequence directly look at every other and pull in relevant information via query-key-value lookups. The core of transformers.
Deep Learning
AUC (Area Under the ROC Curve)
A threshold-free score of a classifier's ranking quality: the probability that a random positive example scores higher than a random negative one. 0.5 = coin flip, 1.0 = perfect.
Machine Learning
Backpropagation
The algorithm that computes the gradient of the loss for every weight in a network by applying the chain rule backwards through the layers.
Deep Learning
Bias (model parameter)
The constant term b in a model like y = wx + b, shifting the output independently of inputs. Not to be confused with statistical or societal bias.
Mathematics
Bias–Variance Trade-off
The tension between models that are too simple (high bias: systematically wrong) and too flexible (high variance: memorize noise). Generalization lives in between.
Machine Learning
Classification
Supervised learning where the output is a category — spam/not-spam, cat/dog/horse — rather than a number.
Machine Learning
Clustering
Unsupervised learning that groups similar examples together without labels, e.g. discovering customer segments with k-means.
Machine Learning
CNN (Convolutional Neural Network)
A network that slides small learned filters over images to detect patterns, exploiting locality and translation invariance. The architecture that conquered vision.
Deep Learning
Cross-Entropy Loss
The standard loss for classification: it measures how surprised the model is by the true labels, punishing confident wrong predictions hardest.
Machine Learning
Cross-Validation
Evaluating a model by splitting data into k folds and training k times, each holding out a different fold — every point serves as validation once.
Machine Learning
Data Leakage
When information unavailable at prediction time sneaks into training (e.g. a feature written after the outcome). Produces inflated test scores and production failures.
Machine Learning
Diffusion Model
A generative model trained to remove noise step by step; generation starts from pure noise and denoises until an image crystallizes. Powers Stable Diffusion and DALL·E.
Deep Learning
Dot Product
Multiply matching components of two vectors and sum. Measures alignment/similarity — the single most executed operation in AI.
Mathematics
Dropout
A regularization technique that randomly silences neurons during training so the network learns redundant, robust features.
Deep Learning
Embedding
A learned dense vector representing a word, image, user or any object, positioned so that geometric closeness encodes semantic similarity.
Machine Learning
Epoch
One full pass of the training algorithm over the entire training dataset. Training typically runs for many epochs.
Deep Learning
Feature
One measurable input property of an example — a house's size, an email's word counts. Features form the input vector x.
Machine Learning
Fine-Tuning
Continuing to train a pretrained model on a smaller, task-specific dataset — adapting general knowledge to a particular job.
Deep Learning
GAN (Generative Adversarial Network)
Two networks trained against each other: a generator forging samples and a discriminator detecting fakes. Produced the first photorealistic synthetic faces.
Deep Learning
Gradient
The vector of all partial derivatives of a function — a compass pointing in the direction of steepest increase of the loss.
Mathematics
Gradient Descent
The optimization algorithm behind nearly all model training: repeatedly step parameters in the direction that decreases the loss.
Mathematics
Hallucination
When a generative model produces fluent but fabricated content — invented facts, citations or APIs — stated with unwarranted confidence.
Artificial Intelligence
Hyperparameter
A knob set by the practitioner rather than learned — learning rate, tree depth, k in k-means — tuned using the validation set.
Machine Learning
Inference
Using a trained model to make predictions on new data, as opposed to training. Also: the reasoning step in logic-based AI.
Machine Learning
Label
The correct answer attached to a training example — the price of a house, the 'spam' tag on an email. What supervised models learn to predict.
Machine Learning
Learning Rate
The step size of gradient descent. Too small: training crawls. Too large: the loss oscillates or explodes. The most important hyperparameter.
Mathematics
LLM (Large Language Model)
A transformer with billions of parameters trained to predict the next token over internet-scale text, then aligned into an assistant (e.g. GPT, Claude).
Artificial Intelligence
Loss Function
A formula measuring how wrong a model's predictions are. Training = minimizing it. Examples: mean squared error, cross-entropy.
Mathematics
Matrix
A rectangular grid of numbers: simultaneously a table of data and a transformation of vectors. Neural network weights are matrices.
Mathematics
Neural Network
Layers of simple units (weighted sums + nonlinearities) composed into one large trainable function. The foundation of deep learning.
Deep Learning
Overfitting
When a model memorizes training data's noise instead of learning the pattern — great training scores, poor real-world performance.
Machine Learning
Parameter
A number inside a model that training adjusts — the weights and biases. GPT-class models have hundreds of billions.
Machine Learning
Precision & Recall
Precision: of everything flagged positive, how much truly was? Recall: of all true positives, how many were caught? They trade off via the decision threshold.
Machine Learning
Prompt Engineering
Crafting the input text to an LLM — instructions, examples, output format — to reliably elicit the desired behavior.
Artificial Intelligence
RAG (Retrieval-Augmented Generation)
Grounding an LLM by vector-searching relevant documents for each query and including them in the prompt, so answers cite real sources.
Artificial Intelligence
Regression
Supervised learning where the output is a continuous number — a price, a temperature — rather than a category.
Machine Learning
Regularization
Deliberately constraining a model (weight penalties, dropout, early stopping) to trade a little training fit for much better generalization.
Machine Learning
Reinforcement Learning
Learning by acting: an agent tries actions, receives rewards, and improves its policy to maximize long-term reward. Trained AlphaGo; aligns LLMs via RLHF.
Artificial Intelligence
ReLU
max(0, x) — the simple activation whose non-fading gradient helped unlock deep networks. The default hidden-layer nonlinearity.
Deep Learning
RLHF
Reinforcement learning from human feedback: humans rank model outputs, a reward model learns those preferences, and RL optimizes the LLM against it.
Artificial Intelligence
Softmax
A function turning a vector of raw scores into probabilities that sum to 1. The standard final layer of classifiers and language models.
Machine Learning
Supervised Learning
Learning from examples labeled with correct answers — the dominant ML paradigm, covering classification and regression.
Machine Learning
Token
The unit LLMs read and write: a learned subword chunk of text. Tokenization explains many LLM quirks, including weaker performance in underrepresented languages.
Artificial Intelligence
Training / Test / Validation Split
The discipline of separating data: fit on training, tune on validation, and measure once on a locked-away test set.
Machine Learning
Transfer Learning
Reusing a model pretrained on a large dataset for a new task — freeze the body, retrain the head. The blueprint of the foundation-model era.
Deep Learning
Transformer
The architecture of modern AI: stacked attention + MLP blocks with skip connections, processing all tokens in parallel. Powers GPT, BERT and vision transformers.
Deep Learning
Underfitting
When a model is too simple to capture the pattern at all — poor performance even on training data. The opposite of overfitting.
Machine Learning
Unsupervised Learning
Finding structure in unlabeled data: clustering, dimensionality reduction, anomaly detection.
Machine Learning
Vector
An ordered list of numbers — the universal container of ML. Every input, output and embedding is a vector.
Mathematics
Weight
A learnable multiplier on an input or connection — the synapse strength of artificial neurons. Training means adjusting weights.
Deep Learning