NCA-GENM Concept Glossary — 51 Terms Defined

Learning paradigms and cross-validationNCAM-C1

The four core ML paradigms — supervised, unsupervised, self-supervised, reinforcement — plus feature engineering and k-fold cross-validation for a reliable performance estimate.

Introduced in Machine learning fundamentals: learning paradigms, feature engineering, and cross-validation

Overfitting, underfitting, and the bias-variance tradeoffNCAM-C2

The tradeoff between a model that memorizes training noise (overfitting) and one too simple to capture the pattern (underfitting), judged only on held-out data.

Introduced in Overfitting, underfitting, and the bias-variance tradeoff

Model comparison metricsNCAM-C3

Accuracy, precision, recall, F1, and ROC-AUC for classification; MAE, MSE, and R² for regression — chosen because accuracy alone misleads on imbalanced data.

Introduced in Model comparison metrics: accuracy, precision, recall, F1, ROC-AUC, MAE, MSE, R²

Deep learning frameworksNCAM-C4

TensorFlow, PyTorch, and Keras, the dominant frameworks for building and training neural networks — PyTorch favored in research for its dynamic graphs, Keras as a high-level API on TensorFlow.

Introduced in Deep learning frameworks: TensorFlow, PyTorch, and Keras

Neural network basics and the training loopNCAM-C5

A neuron computing activation(Σwx+b), and the training loop of loss function, gradient descent, learning rate, backpropagation, and optimizer that updates its weights.

Introduced in Neural network basics: neurons, activation functions, and the training loop

ConvolutionsNCAM-C6

Learnable filters slid over grid-like data — images, spectrograms — to extract local, translation-invariant features; the workhorse of vision models and a building block inside U-Nets.

Introduced in Convolutions and the building blocks of vision models

Residual connections and nonsequential networksNCAM-C7

A skip connection that adds a layer's input to its output (F(x)+x), easing gradient flow in deep networks; nonsequential (functional) architectures allow the branches multimodal fusion needs.

Introduced in Nonsequential networks and residual connections

Multimodal loss functionsNCAM-C8

The loss families multimodal training combines — cross-entropy, contrastive, reconstruction, adversarial, and composite/weighted-sum — including CLIP's contrastive objective.

Introduced in Multimodal loss functions: cross-entropy, contrastive, reconstruction, adversarial, and composite

Training stability in multimodal settingsNCAM-C9

Normalization, learning-rate warmup and scheduling, loss weighting, and gradient clipping — the techniques that keep a composite multimodal objective from letting one modality dominate.

Introduced in Training stability in multimodal settings: normalization, LR warmup, loss weighting, gradient clipping

Multimodal transfer learningNCAM-C10

Adapting a pretrained encoder (vision backbone, text/LLM backbone, or CLIP encoder) to a new task via full fine-tuning or parameter-efficient methods, rather than training from scratch.

Introduced in Multimodal transfer learning: pretrained encoders, full fine-tuning vs. parameter-efficient adaptation

Model fusion and orchestrationNCAM-C11

Combining modalities early, intermediate, or late (fusion) versus coordinating multiple agents or tools (orchestration) — two distinct problems the exam tests as a pair.

Introduced in Model fusion and orchestration: early, intermediate, and late fusion; modality vs. agent orchestration

Multimodal prompt engineeringNCAM-C12

Steering a multimodal model's output with text-plus-image/audio prompts and context embeddings, with no training involved.

Introduced in Prompt engineering fundamentals and emerging multimodal trends

Data cleaningNCAM-C13

Handling missing values, outliers, scaling, and categorical encoding — with scalers fit on the training split only, never the full dataset before splitting.

Introduced in Data cleaning: missing values, outliers, scaling, and categorical encoding

Exploratory data analysis: descriptive statistics and correlationNCAM-C14

Profiling a dataset with descriptive statistics and correlation, where Pearson r measures only linear association and is never proof of causation.

Introduced in Exploratory data analysis: descriptive statistics and correlation

Chart selectionNCAM-C15

Matching a chart type to the question — histogram or box plot for distribution, scatter for relationship, bar for categories, line for time, heatmap for correlation across variables.

Introduced in Choosing the right chart and avoiding misleading visuals

Attention mapsNCAM-C16

A visualization of where a model's attention weight went, useful for explainability and debugging in multimodal settings but only suggestive, not a guaranteed causal explanation.

Introduced in Attention maps as an explainability and debugging tool in multimodal settings

Multimodal data preparation: augmentation and OCRNCAM-C17

Augmentation that expands and diversifies training data, and OCR as the ingestion step that extracts text from scanned or image PDFs feeding a multimodal or RAG pipeline.

Introduced in Preparing multimodal data: augmentation and OCR for PDF extraction

Relationships, trends, and confounding factorsNCAM-C18

Identifying real relationships in data while checking for confounders, including segmenting by subgroup before trusting an aggregate trend.

Introduced in Identifying relationships, trends, and confounding factors in an analysis

Experiment designNCAM-C19

Changing one variable at a time, A/B testing, fixed evaluation sets, and reproducibility — the discipline that lets a result be attributed to a specific change.

Introduced in Experiment design: one variable at a time, A/B testing, fixed evaluation sets, reproducibility

Managing multimodal data from multiple sourcesNCAM-C20

Aligning modalities — pairing captions with images — before augmentation and training, when data arrives from separate sources.

Introduced in Managing and preprocessing multimodal data from multiple sources

Diffusion models: forward and reverse process (DDPM)NCAM-C21

A generative model with two phases: forward diffusion adds Gaussian noise to data, and a learned reverse process removes it step by step to generate a sample from pure noise.

Introduced in Diffusion models: forward noising and reverse denoising (DDPM)

Controlling diffusion output with context embeddingsNCAM-C22

Using CLIP text conditioning to steer what a diffusion model generates, testing and refining the embedding to reach a desired image.

Introduced in Controlling diffusion output with context embeddings

GANs and the generator-discriminator gameNCAM-C23

A generative adversarial network trained as a min-max game between a generator and a discriminator, with no denoising step — a distinct mechanism from diffusion.

Introduced in GANs and the generator-discriminator min-max game

Fréchet Inception Distance (FID)NCAM-C24

A metric comparing generated-versus-real image statistics in a feature space to score generative image quality, where lower is better.

Introduced in Evaluating generative image quality: Fréchet Inception Distance (FID)

Conversational AI pipelines on RivaNCAM-C25

The fixed ASR → NLP/LLM → TTS pipeline order: automatic speech recognition transcribes, an NLP/LLM component understands and responds, and text-to-speech speaks the result back.

Introduced in Conversational AI on Riva: ASR, NLP/LLM, and TTS pipelines

Evaluation metrics by taskNCAM-C26

Task-specific metrics — BLEU for translation, ROUGE for summarization, perplexity for language models, FID for images, and faithfulness plus retrieval quality for RAG.

Introduced in Evaluation metrics by task: classification, regression, text generation, image generation, and RAG

Explainability and data/model quality testingNCAM-C27

Testing data and model quality and consistency, including checking a RAG system's retrieval relevance rather than judging only the final generated answer.

Introduced in Explainability and testing data/model quality and consistency

Making data modalities neural-network readyNCAM-C28

The modality-specific preprocessing needed before fusion: tokens for text, normalized patches for images, spectrograms for audio, and analogous steps for time-series and geospatial data.

Introduced in Data modalities and making them neural-network ready: text, image, audio, time-series, geospatial

Fusion tradeoffs: early, intermediate, and lateNCAM-C29

The accuracy, latency, and missing-modality tradeoffs across fusion points — early fusion lowers inference latency but is alignment-sensitive; late fusion tolerates a missing modality and often maximizes accuracy.

Introduced in Model fusion in depth: early, intermediate, and late fusion tradeoffs

CLIP and shared multimodal embedding spacesNCAM-C30

A model jointly training an image encoder and a text encoder with a contrastive objective so matched pairs land close together, enabling zero-shot classification with no task-specific fine-tuning.

Introduced in CLIP and shared multimodal embedding spaces

Handling missing or incomplete modalitiesNCAM-C31

Strategies for a missing modality: late fusion's graceful degradation, imputation/masking, cross-modal generation, and modality dropout.

Introduced in Handling missing or incomplete modalities

Autoencoders and anomaly detectionNCAM-C32

An autoencoder trained to reconstruct "normal" patterns, where a high reconstruction error on new input flags it as anomalous.

Introduced in Autoencoders and anomaly detection across modalities

Multimodal application patternsNCAM-C33

RAG, chatbots, and summarizers as application patterns over multimodal data — RAG's chunk/embed/store/retrieve/generate pipeline, and a chatbot's resent-history simulation of memory.

Introduced in Application patterns: RAG, chatbots, and summarizers over multimodal data

Python multimodal toolingNCAM-C34

NumPy, spaCy, Keras, and vector databases (FAISS, Milvus, Pinecone, Chroma, pgvector) as the tooling that indexes embeddings for the nearest-neighbour retrieval powering RAG.

Introduced in Python multimodal tooling: NumPy, spaCy, Keras, and vector databases

Mixed-precision trainingNCAM-C35

Training in FP16 with an FP32 master weight copy, loss scaling against underflow, and FP32 accumulation, using Tensor Cores to keep FP16 accuracy on par with FP32.

Introduced in Mixed-precision training: FP16, FP32, loss scaling, and Tensor Cores

Quantization: PTQ vs. QATNCAM-C36

Reducing model precision to shrink memory and speed inference — post-training quantization (PTQ) after training, or quantization-aware training (QAT), which recovers more accuracy by simulating quantization during training.

Introduced in Quantization: PTQ vs. QAT

Neural network pruningNCAM-C37

Removing weights to shrink a model — structured pruning removes whole channels/filters and is hardware-friendly; unstructured pruning zeroes individual weights and is sparser but harder to accelerate.

Introduced in Neural network pruning: structured vs. unstructured

Hyperparameter tuningNCAM-C38

Grid, random, and Bayesian search over hyperparameters, with learning rate the most sensitive one across every strategy.

Introduced in Hyperparameter tuning: grid, random, and Bayesian search

Transfer learning for efficiencyNCAM-C39

Reusing a pretrained encoder as a direct efficiency win — less data, less compute, lower energy cost — versus training from scratch.

Introduced in Transfer learning for efficiency

TensorRT and Triton for energy-efficient inferenceNCAM-C40

TensorRT optimizes a model (fusion, precision calibration, kernel tuning) while Triton serves it — two distinct jobs the exam treats as a standing trap.

Introduced in Energy efficiency and inference optimization with TensorRT and Triton

U-Net architecture and skip connectionsNCAM-C41

An encoder-decoder architecture whose skip connections carry fine spatial detail from encoder to decoder; removing them degrades output quality even though the network still runs.

Introduced in U-Net architecture: encoder-decoder structure and skip connections

The U-Net as diffusion backbone and autoencoderNCAM-C42

The same U-Net that reconstructs or denoises an image is the network run repeatedly inside a diffusion model's reverse process, generating images from pure noise.

Introduced in The U-Net as diffusion denoising backbone and as an autoencoder

CLIP plus diffusion: the text-to-image pipelineNCAM-C43

Encoding a prompt with CLIP's text encoder into a context embedding, conditioning a diffusion U-Net on it, then iteratively denoising to a final image.

Introduced in CLIP plus diffusion: building a text-to-image pipeline

The NVIDIA generative-AI SDK stackNCAM-C44

NeMo builds and customizes, Triton serves, TensorRT optimizes, Riva handles speech, ACE builds avatars, and cuDNN is a low-level primitives library — not a model or a server.

Introduced in NVIDIA SDKs: NeMo, Riva, Triton, ACE, cuDNN, and AI Blueprints/VIA

Prompt engineering and software quality practices for generative systemsNCAM-C45

Iterating prompts and context embeddings for a desired image alongside version control, reproducibility, and validating inputs/outputs before deployment.

Introduced in Prompt engineering for generative systems, and software quality practices

A text-to-image service end to endNCAM-C46

The full pipeline of a production text-to-image service: CLIP context embedding, diffusion U-Net denoising, TensorRT optimization, Triton serving, with monitoring and versioning around it.

Introduced in Putting it together: a text-to-image service end to end

Ethical principles of trustworthy AINCAM-C47

NVIDIA's trustworthy-AI framing around privacy, safety and security, transparency, and nondiscrimination, delivered through fairness, explainability, reliability, and accountability.

Introduced in Ethical principles of trustworthy AI: privacy, safety, transparency, nondiscrimination

Minimizing bias via disaggregated evaluationNCAM-C48

Per-subgroup evaluation that finds bias a high overall accuracy number masks, since bias usually originates in the data rather than the algorithm.

Introduced in Minimizing bias: disaggregated evaluation and mitigation

Data privacy vs. data consentNCAM-C49

Privacy protecting the data itself, and consent as informed, withdrawable permission for a stated use — sharpened for generative models, which can memorize and leak training data.

Introduced in Data privacy vs. data consent

Content authenticity for multimodal generative AINCAM-C50

Provenance, watermarking, disclosure, and detection for deepfakes and synthetic images, audio, and video — the multimodal-specific trust concern beyond text hallucination controls.

Introduced in Content authenticity for multimodal generative AI: provenance, watermarking, disclosure, detection

Hallucination, grounding, and guardrails checklistNCAM-C51

Grounding with RAG and citations to reduce hallucination, NeMo Guardrails to constrain behavior, and a trustworthy-AI checklist mapping each principle to a control.

Introduced in Hallucination, grounding, guardrails, and a trustworthy-AI checklist