Verfasst von / Written by Sebastian F. Genter
E
eager execution
A programming mode in TensorFlow where operations are immediately executed as they are called. Unlike graph execution which builds a computational graph first, eager execution evaluates operations dynamically, making debugging more intuitive. This mode is particularly useful for rapid prototyping and interactive workflows.
early stopping
A regularization technique that halts the training process before the model completes all iterations. This occurs when validation loss starts increasing, indicating potential overfitting. By stopping early, the model retains better generalization capabilities instead of memorizing training data patterns.
earth mover's distance (EMD)
A statistical measure quantifying the dissimilarity between two probability distributions. Often used in image processing and pattern recognition, it calculates the minimum "work" required to transform one distribution into another, where work is defined as the amount of probability mass moved multiplied by distance.
edit distance
A string metric measuring the minimum number of single-character edits (insertions, deletions, or substitutions) required to change one word into another. Common applications include spell checkers, DNA sequence alignment, and natural language processing tasks like text similarity assessment.
Einsum notation
A compact mathematical notation for specifying tensor operations using Einstein summation rules. Widely used in numerical computing libraries like NumPy and TensorFlow, it enables efficient expression of complex operations such as matrix multiplication, transposition, and tensor contraction through concise string descriptors.
embedding layer
A specialized neural network component that converts high-dimensional categorical data into dense, lower-dimensional vector representations. These learned embeddings capture semantic relationships between discrete entities, making them fundamental in NLP tasks and recommendation systems.
embedding space
The multidimensional vector space where embedded representations reside. In this latent space, geometrically close vectors represent semantically similar entities. For example, in word embeddings, "king" and "queen" would be closer than "king" and "apple".
embedding vector
A fixed-length numerical representation of discrete entities (words, products, users) in continuous vector space. These vectors are learned during model training to preserve relational properties - similar items have vectors with higher cosine similarity.
empirical cumulative distribution function (eCDF or EDF)
A non-parametric estimator of the underlying cumulative distribution function (CDF) for observed data. Constructed by ordering samples and creating a step function that increases by 1/n at each data point, providing a visual representation of data distribution.
empirical risk minimization (ERM)
A fundamental machine learning paradigm where model parameters are optimized to minimize loss on training data. Contrasts with structural risk minimization which incorporates regularization terms to prevent overfitting by considering model complexity.
encoder
A neural network component that transforms input data into compressed, meaningful representations. In sequence-to-sequence models, encoders process source inputs (like sentences) into context vectors that decoders use for target sequence generation.
ensemble
A machine learning technique combining predictions from multiple models to improve accuracy and robustness. Common implementations include:
- Bagging (e.g., Random Forests)
- Boosting (e.g., Gradient Boosted Trees)
These reduce variance and bias respectively.
entropy
In information theory, a measure of uncertainty or randomness in a system. For classification tasks, entropy quantifies the impurity of label distribution:
- A pure node (all same class) has zero entropy.
- Maximum entropy occurs with uniform class distribution.
environment
In reinforcement learning, the world context where an agent operates. Provides state information and rewards in response to agent actions. Environments range from simple grid worlds to complex simulations like game engines or robotic control systems.
episode
A complete sequence of agent-environment interactions in reinforcement learning, from initial state to terminal state. Episodes are used to evaluate policy performance, particularly in episodic tasks with natural termination points (e.g., winning/losing a game).
epoch
A full pass through the entire training dataset during model training. The number of epochs determines how many times learning algorithms will see each sample. Proper epoch selection balances underfitting and overfitting risks.
epsilon greedy policy
A reinforcement learning exploration strategy balancing exploitation (choosing best-known action) and exploration (trying random actions).
- With probability ε, selects random action.
- Otherwise chooses optimal action.
ε typically decays over time to favor exploitation.
equality of opportunity
A fairness metric ensuring equal true positive rates across protected groups. Requires that qualified individuals from all demographics have equal chances of receiving positive predictions, regardless of base rate differences between groups.
equalized odds
A stricter fairness criterion than equality of opportunity, requiring both equal true positive rates and equal false positive rates across protected groups. Ensures error distributions are similar for all demographics.
Estimator
A deprecated TensorFlow API for simplified model development. Superseded by Keras API, it provided high-level abstractions for training, evaluation, and prediction. Modern implementations should use tf.keras for better flexibility and integration.
evals
(LLM Evaluations) Systematic assessment protocols for large language models measuring performance across metrics like factual accuracy, reasoning ability, and safety. Includes:
- Automated metrics (BLEU, ROUGE)
- Human evaluation frameworks
evaluation
The process of assessing model performance using predefined metrics.
- For supervised learning, typically involves validation/test sets.
- For generative models, may include human assessments, adversarial evaluation, or task-specific benchmarks.
example
A single data instance used in machine learning systems. Contains features (input variables) and optionally labels (target values). In supervised learning, labeled examples are used for training, while unlabeled examples are used for inference.
experience replay
A reinforcement learning technique storing past experiences (state, action, reward, next state) in a buffer. During training, random batches from this buffer are sampled to break temporal correlations and improve learning stability.
experimenter's bias
A form of confirmation bias where researchers unconsciously influence experiments to confirm preconceived hypotheses. In ML, this might manifest through selective data collection, metric choice, or repeated model tuning until desired results appear.
exploding gradient problem
A training instability where gradients grow exponentially during backpropagation, causing drastic parameter updates. Common in deep networks and recurrent architectures, often addressed through:
- Gradient clipping
- Architectural modifications like LSTMs