Verfasst von / Written by Sebastian F. Genter
I
i.i.d.
Abbreviation for independently and identically distributed.
image recognition
A process within machine learning focused on classifying the objects, patterns, or concepts present within an image. It is often used synonymously with image classification.
imbalanced dataset
Synonym for #class-imbalanced dataset.
implicit bias
Refers to the act of automatically forming associations or assumptions based on an individual's mental models, experiences, and memories. Implicit bias can subtly influence various stages of machine learning development, including data collection, data classification, and the design choices made when building ML systems. For example, using the presence of a white dress as a feature for identifying wedding photos introduces implicit bias, as this custom is specific to certain cultures and historical periods. See also #confirmation bias.
imputation
A shortened form of #value imputation.
incompatibility of fairness metrics
The concept that various mathematical definitions of fairness in machine learning can be mutually exclusive, meaning it's impossible to satisfy all of them simultaneously for a given model. Consequently, no single, universal fairness metric exists that applies to every ML problem. This highlights the need to define and prioritize fairness contextually, focusing on mitigating specific harms relevant to the model's application and use case.
in-context learning
Synonym for #few-shot prompting.
independently and identically distributed (i.i.d)
Describes data points drawn from a probability distribution that remains constant over time and where each draw is statistically independent of all previous draws. While a fundamental assumption in many theoretical analyses (akin to an "ideal gas" in physics), truly i.i.d. data is rare in real-world applications. For instance, website visitor data might approximate i.i.d. conditions over a very short period but will likely exhibit dependencies and changing patterns (like seasonality) over longer durations. See also #nonstationarity.
individual fairness
A #fairness metric focused on the principle that individuals who are similar with respect to relevant attributes should receive similar predictions or outcomes from the model. For example, an admissions model satisfies individual fairness if two applicants with identical grades and test scores have an equal likelihood of admission. A key challenge lies in defining "similarity" appropriately; if the chosen similarity metric omits crucial information, enforcing individual fairness might inadvertently introduce other fairness issues.
inference
In the context of machine learning, inference is the stage where a trained #model is used to make predictions on new, previously unseen unlabeled examples. This contrasts with the training phase where the model learns from labeled data. (Note: Inference carries a different meaning in traditional statistics).
inference path
Within a #decision tree, this is the specific sequence of conditions (nodes) traversed by a particular #example during #inference, starting from the #root node and ending at a #leaf node, which provides the final prediction. The path taken depends on how the example's feature values satisfy the conditions at each node.
information gain
In the context of building decision forests, information gain quantifies the reduction in #entropy achieved by splitting a dataset based on a particular #condition (feature test). It is calculated as the difference between the entropy of the parent node (before the split) and the weighted average entropy of the child nodes (after the split), where the weighting is based on the number of examples in each child node. Decision tree splitters typically aim to select the condition at each node that maximizes information gain, leading to more homogeneous child nodes.
in-group bias
A tendency to favor or show partiality towards members of one's own group or those sharing similar characteristics. In machine learning development, if model testers or data raters predominantly consist of the developer's close associates (friends, family, colleagues), this bias can compromise the validity of testing and dataset representativeness. In-group bias is considered a form of #group attribution bias. See also #out-group homogeneity bias.
input generator
A mechanism or component responsible for loading and processing raw data into the format required by a #neural network, typically Tensors. It handles the transformation of data and iterates over it to produce batches suitable for model training, #evaluation, and #inference.
input layer
The initial layer in a #neural network that receives the input data, specifically the #feature vector representing an #example. This layer doesn't perform computations but simply passes the feature values to the first #hidden layer.
in-set condition
In a #decision tree, a type of #condition that checks if the value of a #feature belongs to a specified set of items. For example, a condition `house_style in [tudor, colonial, cape]` evaluates to 'Yes' if the `house_style` feature value is 'tudor', 'colonial', or 'cape', and 'No' otherwise. Using in-set conditions can often result in more compact and efficient decision trees compared to testing multiple one-hot encoded features individually.
instance
Synonym for #example.
instruction tuning
A refinement technique, considered a form of #fine-tuning, aimed at enhancing a #generative AI model's capability to understand and follow instructions given in a #prompt. This involves training the model on a dataset composed of varied instruction-based prompts and their desired outputs. An instruction-tuned model typically performs better at responding effectively to zero-shot prompts across a diverse range of tasks. Compare with #parameter-efficient tuning and #prompt tuning.
interpretability
The extent to which a human can understand the reasoning behind a machine learning model's predictions or decisions. Some models, like #linear regression (where weights directly indicate feature influence) and decision trees, are inherently more interpretable than others. Complex models like deep neural networks often require specialized tools and techniques, such as the Learning Interpretability Tool (LIT), to gain insights into their internal workings.
inter-rater agreement
A measure of consistency or consensus among multiple human raters (also called annotators) when performing a labeling or annotation task. High inter-rater agreement suggests that the task instructions are clear and the raters interpret them similarly. Low agreement might indicate ambiguous instructions needing refinement. Metrics like Cohen's kappa are commonly used to quantify inter-rater agreement. Also known as inter-annotator agreement or inter-rater reliability.
intersection over union (IoU)
A metric used frequently in object detection tasks within computer vision. It quantifies the degree of overlap between two bounding boxes: typically the predicted bounding box generated by the model and the #ground truth bounding box. IoU is calculated as the ratio of the area of overlap between the two boxes to the area of their combined union. The value ranges from 0 (indicating no overlap) to 1 (indicating perfect overlap where both boxes have identical coordinates).
IoU
Abbreviation for intersection over union.
item matrix
In the context of recommendation systems using #matrix factorization, this is one of the resulting matrices. It is composed of embedding vectors that represent latent (hidden) features or characteristics of each item being recommended (e.g., movies, products). Each row typically corresponds to a single latent feature, and each column corresponds to an item. The values capture the extent to which each item possesses each latent feature. This matrix usually has dimensions (number of latent features) x (number of items).
items
Within a #recommendation system, these are the entities that the system suggests or recommends to users. Examples include movies in a movie recommendation service, books in a bookstore application, or products on an e-commerce site.
iteration
A single update step for the model's parameters (weights and biases) during the training process. The number of examples processed within one iteration is determined by the #batch size. For neural networks trained with #gradient descent, one iteration typically involves processing one #batch through both a #forward pass (to calculate #loss) and a #backward pass (#backpropagation, to calculate and apply gradients).