This glossary covers essential ML/AI terms relevant for PMs working with related products/technologies.
Accuracy
The percentage of correct predictions made by a machine learning model, often used to evaluate performance.
Why important for PMs: Higher accuracy means a model is making better decisions, but PMs should also watch for trade-offs like bias or overfitting.
Example: "Before deploying the fraud detection model, let's check its accuracy to ensure reliable predictions."
Related Terms: Precision, Recall, Model Training
Agent
An AI-powered system that can perceive its environment, make decisions, and take actions to achieve a goal.
Why important for PMs: Agents automate tasks, making products more intelligent and interactive—PMs need to define their behaviors and guardrails.
Example: "Our chatbot is an AI agent that helps users troubleshoot issues without human intervention."
Related Terms: Agentic, Autonomous, Reinforcement Learning
Agentic
The ability of an AI system to operate independently, make decisions, and take actions without constant human input.
Why important for PMs: PMs must carefully design agentic AI systems to ensure they align with business goals while maintaining user trust and safety.
Example: "Since our AI assistant is agentic, we need to define strict boundaries for automated decision-making."
Related Terms: Agent, Autonomous, Artificial General Intelligence (AGI)
Anomaly Detection
The process of identifying unusual patterns or outliers in data that don’t conform to expected behavior.
Why important for PMs: Helps detect fraud, security threats, and operational issues before they escalate.
Example: "We should integrate anomaly detection to flag suspicious transactions in real time."
Related Terms: AI Bias, Model Training, Inference
Artificial General Intelligence (AGI)
A theoretical form of AI that can perform any intellectual task a human can, unlike today’s specialized AI models.
Why important for PMs: While AGI doesn’t exist yet, PMs working in AI should understand its implications for future product innovation.
Example: "We focus on narrow AI, but understanding AGI helps us anticipate long-term technological shifts."
Related Terms: Autonomous, Machine Learning, Artificial Neural Network
Autonomous
An AI system that can operate independently, making its own decisions based on data inputs.
Why important for PMs: Autonomous AI can improve efficiency but requires careful design to ensure it aligns with user needs and ethical considerations.
Example: "Before rolling out autonomous customer support bots, we should test their decision-making accuracy."
Related Terms: Agent, Agentic, Reinforcement Learning
Backpropagation
A training process in neural networks where errors are sent backward through layers to adjust weights and improve accuracy.
Why important for PMs: It’s the backbone of deep learning—understanding it helps PMs communicate with engineers about model training and improvements.
Example: "The metric dashboard indicates that our image recognition model isn't learning well—let’s check if the backpropagation process is converging correctly."
Related Terms: Neural Networks, Gradient Descent, Model Training
Batch
A group of data samples processed together instead of one at a time, improving model training efficiency.
Why important for PMs: Batch processing speeds up AI training and inference, impacting system performance and costs.
Example: "To optimize training time, we should increase the batch size while ensuring the model still generalizes well."
Related Terms: Model Training, Mini-Batch, Stochastic Gradient Descent
Bias (AI Bias)
A systematic error in an AI model that causes it to favor certain outcomes unfairly, often due to imbalanced training data.
Why important for PMs: AI bias can lead to unfair or harmful decisions—PMs must ensure models are trained on diverse, representative data.
Example: "We need to audit our AI model for bias to ensure fair recommendations for all users."
Related Terms: Anomaly Detection, Fairness, Ethics in AI
Computer Vision
A field of AI that enables machines to interpret and process visual data like images and videos.
Why important for PMs: it powers applications like facial recognition, object detection, and autonomous driving—PMs should understand its limitations and ethical considerations.
Example: "Our app uses computer vision to automatically tag objects in photos—let’s refine it to improve accuracy."
Related Terms: Deep Learning, Neural Networks, AI Bias
Cross-Validation
A technique for evaluating how well a machine learning model will perform on unseen data by training and testing it on different subsets.
Why important for PMs: Helps prevent overfitting, ensuring the model generalizes well to real-world data.
Example: "We should use cross-validation to validate our model’s accuracy before deploying it."
Related Terms: Overfitting, Model Training, Confusion Matrix
Confusion Matrix
A table used to measure a classification model’s performance by showing true positives, false positives, true negatives, and false negatives.
Why important for PMs: Helps analyze how well an AI model distinguishes between different outcomes, guiding improvements.
Example: "The confusion matrix shows a high false positive rate—let’s tweak the model’s threshold to improve precision."
Related Terms: Accuracy, Precision, Recall, False Positive
Dataset
A structured collection of data used to train, validate, or test machine learning models.
Why important for PMs: The quality of a dataset directly impacts model performance—PMs need to ensure data is relevant, diverse, and unbiased.
Example: "Before training our model, we need to validate if our dataset is representative of all user demographics."
Related Terms: Data Preprocessing, Training Data, AI Bias
Data Drift
Data drift occurs when the statistical properties of incoming data change over time, causing a machine learning model to perform less accurately. This happens when the new data the model encounters differs significantly from the data it was originally trained on, and leads to inaccurate predictions and decreased model reliability.
Why important for PMs: PMs working on AI/ML products must monitor data drift to prevent model degradation, maintain accuracy, and ensure a reliable user experience. Ignoring data drift can lead to poor predictions, biased decisions, and a decline in product effectiveness, requiring continuous model retraining and adaptation.
Example: An ML model predicting consumer purchasing patterns may perform well throughout the year but struggle during Black Friday or holiday seasons. The sudden increase in demand, shift toward discount-based buying, and surge in gift purchases differ from regular shopping trends, leading to inaccurate recommendations.
Related Terms: Model Degradation, Concept Drift, Continuous Learning
Data Preprocessing
The process of cleaning, transforming, and structuring raw data to make it usable for machine learning models.
Why important for PMs: Poorly processed data can lead to inaccurate models—PMs should ensure proper handling of missing values, duplicates, and inconsistencies.
Example: "Our raw user data has missing values—let’s refine our data preprocessing steps to handle them."
Related Terms: Dataset, Feature Engineering, Data Cleaning
Decision Tree
A machine learning model that makes predictions by mapping choices and their possible outcomes in a tree-like structure.
Why important for PMs: Decision trees are interpretable, making them useful for explaining AI decisions to stakeholders.
Example: "Since we need explainability, let’s start with a decision tree before moving to more complex models."
Related Terms: Random Forest, Model Interpretability, Classification
Deep Learning
A subset of machine learning that uses multi-layered neural networks to analyze complex patterns in large datasets.
Why important for PMs: Deep learning powers AI breakthroughs like image recognition and NLP—PMs should understand its capabilities and limitations.
Example: "Our product relies on deep learning to automatically generate captions for videos."
Related Terms: Neural Networks, Model Training, Backpropagation
Dimensionality Reduction
The process of reducing the number of input variables in a dataset while preserving essential information.
Why important for PMs: Helps improve model performance and efficiency by removing redundant or irrelevant data.
Example: "Since our dataset has too many features, let’s apply dimensionality reduction to speed up training."
Related Terms: Feature Engineering, Principal Component Analysis (PCA), Overfitting
Embeddings
A way of representing words, images, or other data as numerical vectors so that similar items are positioned close together in a multi-dimensional space.
Why important for PMs: Embeddings make AI models more efficient at understanding relationships between data points, improving recommendations, search, and NLP applications.
Example: "We should refine our word embeddings to improve the accuracy of our AI-powered search engine."
Related Terms: Natural Language Processing (NLP), Feature Engineering, Dimensionality Reduction
Epoch
A complete pass through the entire training dataset during model training. More epochs generally improve learning but can lead to overfitting.
Why important for PMs: Understanding epochs helps PMs balance model performance and training time to avoid excessive computation costs.
Example: "Our model's accuracy plateaued after 10 epochs—let’s stop training to save resources."
Related Terms: Model Training, Overfitting, Batch
Feature Engineering
The process of selecting, transforming, or creating new input variables (features) to improve model performance.
Why important for PMs: Good feature engineering leads to better AI performance—PMs should ensure relevant and high-quality data is used.
Example: "Let’s explore additional user behavior metrics as part of feature engineering to enhance our recommendation model."
Related Terms: Feature Selection, Data Preprocessing, Dimensionality Reduction
Feature Selection
The process of choosing the most relevant features from a dataset to improve model performance and efficiency.
Why important for PMs: Reducing unnecessary features speeds up training, lowers costs, and prevents overfitting, making models more efficient.
Example: "Instead of using all customer attributes, let’s apply feature selection to focus on the ones that impact churn the most."
Related Terms: Feature Engineering, Dimensionality Reduction, Overfitting
Few-shot Learning
A machine learning approach where a model learns to make predictions with only a small amount of training data.
Why important for PMs: Few-shot learning allows AI systems to generalize from limited examples, useful in applications like personalization and rare-case recognition.
Example: "Since we don’t have much labeled data, we should explore few-shot learning for our intent detection model."
Related Terms: Transfer Learning, Fine-Tuning, Generalization
Fine-Tuning
The process of taking a pre-trained machine learning model and adjusting it for a specific task using new data.
Why important for PMs: Fine-tuning speeds up development and improves model performance while reducing training costs.
Example: "Instead of training from scratch, let’s fine-tune an existing model to suit our chatbot’s needs."
Related Terms: Transfer Learning, Few-shot Learning, Model Training
Foundational Model
A large-scale AI model trained on massive datasets that serves as a base for various downstream tasks, such as text generation, image recognition, or code generation (e.g., GPT, BERT, DALL·E).
Why important for PMs: Foundational models provide a strong starting point for AI applications, reducing the need for extensive training. PMs should understand their capabilities, limitations, and customization options.
Example: "Instead of building an AI model from scratch, we can fine-tune a foundational model like GPT to power our chatbot."
Related Terms: Fine-Tuning, Transfer Learning, Large Language Model (LLM), Few-shot Learning
Generative AI
A type of AI that creates new content—such as text, images, music, or code—based on patterns learned from existing data (e.g., GPT for text, DALL·E for images).
Why important for PMs: Generative AI enables innovative product experiences, from AI-driven content creation to chatbots. PMs must define its use cases, guardrails, and ethical considerations.
Example: "We can use generative AI to create personalised marketing copy for different customer segments."
Related Terms: Foundational Model, Large Language Model (LLM), Fine-Tuning
Gradient Descent
An optimisation algorithm that helps AI models learn by gradually adjusting their parameters to minimize errors.
Why important for PMs: Understanding gradient descent helps PMs grasp how models improve over time and why training takes time and computing power.
Example: "If our model isn't converging, we may need to tweak the learning rate in gradient descent."
Related Terms: Backpropagation, Model Training, Hyperparameter
Hallucination
A phenomenon when an AI model generates false, misleading, or entirely fabricated information that appears plausible but is incorrect. This is common in large language models (LLMs) and generative AI.
Why important for PMs: Hallucinations can erode user trust and lead to misinformation. PMs must implement safeguards, validation mechanisms, and user disclaimers to manage this risk.
Example: "Before launching our AI assistant, we need a strategy to detect and mitigate model hallucination to ensure accurate responses."
Related Terms: Generative AI, Large Language Model (LLM), Inference, AI Bias
Hyperparameter
A setting that controls how a machine learning model learns, such as learning rate, batch size, or number of training epochs.
Why important for PMs: Hyperparameters impact model performance, training speed, and cost. PMs should understand their role when discussing optimizations with data scientists.
Example: "Before retraining the model, let’s tune the hyperparameters to find the best combination for accuracy and efficiency."
Related Terms: Model Training, Gradient Descent, Overfitting
Inference
The process of using a trained machine learning model to make predictions or generate outputs based on new data.
Why important for PMs: Inference affects user experience, as it determines how quickly and accurately AI-driven features respond. PMs need to consider latency and infrastructure costs.
Example: "To improve app performance, we should optimize our AI inference time for real-time responses."
Related Terms: Model Training, Latency, Scalability
Intent
The goal or purpose behind a user's input, often used in Natural Language Processing (NLP) for chatbot and voice assistant interactions.
Why important for PMs: Understanding intent helps build AI-powered assistants that provide relevant and accurate responses. PMs need to design clear intent recognition models.
Example: "Our chatbot struggles with intent detection—let’s refine the training data to improve accuracy."
Related Terms: NLP, Embeddings, Few-shot Learning
Label
The category assigned to a data point in supervised learning, used to train models.
Why important for PMs: Labels define the quality of training data—PMs should ensure data labeling is accurate and unbiased to improve model performance.
Example: "To improve our sentiment analysis model, we need more accurately labeled data for training."
Related Terms: Dataset, Supervised Learning, Annotation
Large Language Model (LLM)
A type of AI model trained on massive text datasets to understand and generate human-like language (e.g., GPT, BERT).
Why important for PMs: LLMs power AI-driven experiences like chatbots, summarization tools, and content generation—PMs need to define use cases, guardrails, and fine-tuning needs.
Example: "Instead of building our own NLP model, we can fine-tune an existing LLM for our customer support bot."
Related Terms: Foundational Model, Generative AI, Fine-Tuning
Latent Variable
A variable that influences data but is not directly observed, often inferred by the model (e.g., customer intent in recommendation systems).
Why important for PMs: Understanding latent variables helps PMs improve AI-driven personalization and interpretability of models.
Example: "Our recommendation engine uses latent variables to infer user preferences beyond explicit actions."
Related Terms: Embeddings, Dimensionality Reduction, Feature Engineering
Learning Rate
Measure that controls how much a model updates its weights during training; too high may lead to instability, too low may slow learning.
Why important for PMs: Learning rate affects model training efficiency and performance—PMs should be aware of its role in model optimisation.
Example: "Our model isn’t converging properly—let’s check if the learning rate needs adjustment."
Related Terms: Gradient Descent, Hyperparameter, Model Training
Logistic Regression
A simple machine learning algorithm used for binary classification problems (e.g., predicting whether an email is spam or not).
Why important for PMs: Though basic, logistic regression is widely used in AI-powered decision-making, and PMs should understand its strengths and limitations.
Example: "For a quick and interpretable model, we can start with logistic regression before moving to complex algorithms."
Related Terms: Classification, Decision Tree, Neural Networks
Loss Function
A mathematical formula that measures how far a model’s predictions are from the actual values, guiding its learning process.
Why important for PMs: Loss functions help determine how well an AI model is learning—PMs should understand them when evaluating model performance.
Example: "If our model isn’t improving, we might need to experiment with different loss functions to optimize performance."
Related Terms: Gradient Descent, Backpropagation, Model Training
Machine Learning (ML)
A branch of AI that enables systems to learn from data and make predictions or decisions without being explicitly programmed.
Why important for PMs: ML powers many AI-driven products—PMs need to understand its capabilities, limitations, and how to integrate it into product strategy.
Example: "Our fraud detection system uses machine learning to continuously improve its ability to identify suspicious transactions."
Related Terms: Deep Learning, Model Training, Supervised Learning
Machine Learning Operations (MLOps)
A set of practices that make the deployment easier via monitoring and management of ML models in production environments.
Why important for PMs: MLOps ensures AI models remain accurate, scalable, and reliable—PMs must collaborate with engineering teams to implement best practices.
Example: "We need a strong MLOps strategy to continuously monitor and retrain our recommendation engine as user behavior evolves."
Related Terms: Model Training, Inference, Observability
Model Evaluation Metrics
A set of measurements used to assess how well a machine learning model performs, such as accuracy, precision, recall, and F1-score.
Why important for PMs: Choosing the right metric helps PMs ensure the AI model meets business objectives and user needs.
Example: "Since we care more about minimizing false negatives, recall is the most important model evaluation metric for us."
Related Terms: Confusion Matrix, Accuracy, Loss Function
[DL/ML] Model
A trained AI system that takes input data and makes predictions or decisions based on learned patterns.
Why important for PMs: ML and DL models form the core of AI-driven products—PMs should understand their purpose, limitations, and resource needs.
Example: "Our DL model powers the image recognition feature, but we need to optimize it for mobile devices."
Related Terms: Neural Networks, Model Training, Inference
Model Training
The process of teaching an AI model by exposing it to labeled data so it can learn patterns and make predictions.
Why important for PMs: Training impacts performance, cost, and scalability—PMs should ensure training data quality and alignment with business goals.
Example: "Before launching, let’s run another round of model training to improve accuracy with the latest data."
Related Terms: Dataset, Fine-Tuning, Hyperparameter
Natural Language Processing (NLP)
A branch of AI that enables machines to understand, interpret, and generate human language, powering applications like chatbots, translation, and sentiment analysis.
Why important for PMs: NLP is key to conversational AI, search, and automated content—PMs need to define its use cases and manage challenges like bias and context understanding.
Example: "Our search feature uses NLP to understand user queries and provide more relevant results."
Related Terms: Large Language Model (LLM), Generative AI, Intent
Narrow AI
AI designed to perform specific tasks, such as image recognition or language translation, as opposed to Artificial General Intelligence (AGI), which would have human-like reasoning across multiple domains.
Why important for PMs: Most AI products today use narrow AI—PMs should focus on solving well-defined problems rather than expecting broad intelligence.
Example: "Our recommendation engine is an example of narrow AI—it excels at suggesting content but can't perform general reasoning."
Related Terms: Artificial General Intelligence (AGI), Machine Learning, Deep Learning
Neural Network
A machine learning model inspired by the structure of the human brain, consisting of layers of interconnected nodes (neurons) that learn patterns in data.
Why important for PMs: Neural networks power modern AI breakthroughs like deep learning and NLP—PMs should understand their strengths (pattern recognition) and challenges (training complexity).
Example: "We use a neural network to analyze customer sentiment and detect trends in user feedback."
Related Terms: Deep Learning, Model Training, Backpropagation
Overfitting
A process when a machine learning model learns patterns too specifically from training data, performing well on training but poorly on new data.
Why important for PMs: Overfitting leads to AI models that fail in real-world scenarios—PMs should ensure models generalize well to unseen data.
Example: "Our fraud detection model performs great on past data but is overfitting—let’s test it with fresh transactions to validate its accuracy."
Related Terms: Cross-Validation, Generalisation, Model Training
Precision
A metric that measures how many of the positive predictions made by a model are actually correct, helping assess accuracy in classification tasks.
Why important for PMs: Precision is crucial when false positives are costly, such as in fraud detection or medical diagnoses—PMs need to balance precision with recall.
Example: "Our spam filter has high precision, meaning most flagged emails are truly spam."
Related Terms: Recall, Confusion Matrix, Model Evaluation Metrics
Prompt Engineering
The process of designing inputs (prompts) to guide AI models, like chatbots or LLMs, to generate useful and relevant responses.
Why important for PMs: Good prompt engineering improves AI responses without retraining, making it a powerful tool for optimizing user interactions.
Example: "By tweaking the prompt, we made our AI assistant’s responses more concise and helpful."
Related Terms: Generative AI, Large Language Model (LLM), Fine-Tuning
Quantisation
A technique used to reduce the size and computational requirements of AI models by lowering the precision of numerical values (such as weights and activations in neural networks), converting 32-bit floating points to 8-bit integers. This speeds up inference while using less memory.
Why important for PMs: Quantization helps make AI models run efficiently on mobile devices and edge computing by reducing their size and power consumption. PMs should consider it when optimizing models for speed, cost, and deployment constraints.
Example: "To deploy our AI model on smartphones, we need to apply quantization to shrink its size and improve inference speed."
Related Terms: Model Optimisation, Inference, Latency, Edge AI
Recall
A metric that measures how many of the actual positive cases a model correctly identifies, helping evaluate model sensitivity.
Why important for PMs: Recall is critical when missing a positive case is costly, such as in disease detection—PMs must balance recall with precision.
Example: "Since recall is low, our fraud detection model is missing too many real fraud cases—let’s adjust its sensitivity."
Related Terms: Precision, Confusion Matrix, Model Evaluation Metrics
Regularisation
A technique used in machine learning to prevent overfitting by adding constraints or penalties to a model’s complexity.
Why important for PMs: Regularization improves model generalization, ensuring AI works well on new data instead of just memorizing training examples.
Example: "Our model overfits the training data—let’s apply regularisation to make it more robust."
Related Terms: Overfitting, Generalisation, Regression
Regression
A type of machine learning algorithm used to predict continuous values, such as price, temperature, or sales numbers.
Why important for PMs: Regression helps in forecasting trends and making data-driven product decisions, like predicting churn or revenue growth.
Example: "Using regression, we can predict next quarter’s revenue based on historical trends."
Related Terms: Logistic Regression, Model Training, Supervised Learning
Reinforcement Learning (RL)
A machine learning approach where an AI agent learns by interacting with an environment and receiving rewards or penalties based on its actions.
Why important for PMs: RL is useful for optimizing complex decision-making processes, such as recommendation engines or autonomous systems.
Example: "We’re exploring reinforcement learning to improve our dynamic pricing model."
Related Terms: Agent, Agentic, Supervised Learning
Retrieval-Augmented Generation (RAG)
A technique that enhances generative AI by retrieving relevant external data before generating a response, improving accuracy and reducing hallucinations.
Why important for PMs: RAG improves the reliability of AI-generated content by incorporating real-time, up-to-date information—helpful for chatbots, search engines, and research tools.
Example: "To make our AI assistant more factual, we’re integrating retrieval-augmented generation (RAG) to pull in real-time news updates."
Related Terms: Generative AI, Large Language Model (LLM), Model Hallucination
Supervised vs. Unsupervised Learning
Two approaches to training ML models (depending on a task and availability of annotated data).
In Supervised Learning the model learns from labeled data, meaning each input has a correct answer (e.g., classifying emails as spam or not).
In Unsupervised Learning the finds patterns in unlabelled data without predefined categories (e.g., grouping customers by purchasing behaviour).
Why important for PMs: Supervised learning is great for well-defined problems, while unsupervised learning helps uncover hidden insights. PMs should choose the right approach based on data availability and business goals.
Example: "Since we have labeled customer feedback, we should use supervised learning for sentiment analysis, but for customer segmentation, unsupervised learning might be better."
Related Terms: Semi-Supervised Learning, Training Data, Clustering
Semi-Supervised Learning
A hybrid approach where a model learns from a small amount of labeled data and a large amount of unlabeled data.
Why important for PMs: Semi-supervised learning reduces the need for expensive labeled datasets while still achieving good performance. PMs should consider this approach when labeling data is costly or time-consuming.
Example: "Since manually labeling all medical images is expensive, we can use semi-supervised learning to train our AI model efficiently."
Related Terms: Supervised Learning, Unsupervised Learning, Training Data
Transformer Model
A deep learning architecture designed for processing sequential data, especially text, by using self-attention mechanisms to understand context efficiently (e.g., GPT, BERT).
Why important for PMs: Transformers power modern NLP applications like chatbots, summarization, and translation—PMs should know how they enable large-scale AI models.
Example: "Since our chatbot needs to handle long conversations, we should leverage a transformer model like GPT for better contextual understanding."
Related Terms: Large Language Model (LLM), Natural Language Processing (NLP), Generative AI
Train/Test Split
The practice of dividing a dataset into two parts: a training set to teach the model and a test set to evaluate its performance.
Why important for PMs: Ensures models generalize well to new data rather than just memorizing training examples. PMs should check that train/test splits are properly balanced.
Example: "Before deploying our model, we need to validate it using a proper train/test split to avoid overfitting."
Related Terms: Training Data, Cross-Validation, Overfitting
Training Data
The dataset used to teach a machine learning model by helping it recognise patterns and relationships.
Why important for PMs: The quality of training data directly impacts model performance—PMs should ensure it is diverse, clean, and representative.
Example: "Our AI model is biased because our training data doesn’t include enough diverse user inputs—let’s fix that before launch."
Related Terms: Dataset, Label, Train/Test Split
Transfer Learning
A machine learning technique where a pre-trained model is adapted to a new but related task, reducing training time and data requirements.
Why important for PMs: Transfer learning enables faster AI development with fewer resources—PMs can leverage existing models instead of training from scratch.
Example: "Instead of training our AI assistant from scratch, we can apply transfer learning using an existing LLM."
Related Terms: Fine-Tuning, Foundational Model, Few-Shot Learning
Tuning
The process of adjusting a machine learning model’s hyperparameters to improve its accuracy and performance.
Why important for PMs: Proper tuning helps models perform better without excessive computation costs—PMs should track its impact on deployment time and efficiency.
Example: "We need to spend time tuning our recommendation model’s parameters to balance accuracy and speed."
Related Terms: Hyperparameter, Fine-Tuning, Model Optimisation
Validation Data
A subset of data used to fine-tune a machine learning model during training, ensuring it generalises well before testing on unseen data.
Why important for PMs: Validation data helps prevent overfitting and ensures that models perform well beyond just the training set. PMs should monitor validation accuracy to gauge model readiness for deployment.
Example: "Our model performs well on training data, but the validation data suggests overfitting—we need to regularize it."
Related Terms: Train/Test Split, Overfitting, Cross-Validation
Variance
A measure of how much a model's predictions fluctuate based on different training datasets. High variance means the model is too sensitive to the training data and may not generalize well.
Why important for PMs: Models with high variance may perform well in testing but fail in real-world scenarios—PMs should work with data scientists to balance bias and variance.
Example: "Our model has high variance, meaning it’s too dependent on training data and might not generalise well to new users."
Related Terms: Bias-Variance Tradeoff, Overfitting, Regularisation
Vector Embeddings
A numerical representation of complex data (like words, images, or user behavior) in a multi-dimensional space, making it easier for AI models to understand relationships.
Why important for PMs: Embeddings power recommendations, NLP, and search—PMs should understand their role in improving personalisation and AI-driven experiences.
Example: "By improving our vector embeddings, we can make product recommendations more relevant to each user."
Related Terms: Embeddings, Feature Engineering, Dimensionality Reduction
Weights
Numerical values in a neural network that determine the importance of different inputs when making predictions. Adjusting weights during training helps the model learn patterns in data.
Why important for PMs: Weights control how AI models learn—PMs should understand their role when discussing model accuracy, performance, and interpretability with engineers.
Example: "The model is misclassifying images—let’s check if the weight updates are improving learning."
Related Terms: Neural Networks, Backpropagation, Gradient Descent
Zero-Shot Learning
A machine learning approach where a model can make predictions on new, unseen categories without being explicitly trained on them.
Why important for PMs: Zero-shot learning allows AI models to generalize better, reducing the need for expensive labeled datasets. PMs can leverage this for scalability and adaptability in AI applications.
Example: "Using zero-shot learning, our chatbot can handle new user queries without needing specific training data for each topic."
Related Terms: Few-Shot Learning, Transfer Learning, Foundational Model