AI Exam - 3A Model + Inference

AI Practitioner Exam Prep - Model + Inference

Abbreviations
ATLAS = Adversarial Threat Landscape for AI Systems,
BERT metric = Bidirectional Encoder Representations from Transformers,
BLEU metric = Bi-Lingual Evaluation Understudy, CNN = Convolutional Neural Networks, 

GLUE benchmark = General Language Understanding Evaluation,
GPT = Generative Pre-trained Transformers, FM = Foundational Models, 

L = Labeled Data, LLM = Large Language Models, 

METEOR metric = Metric for Evaluation of Translation with Explicit ORdering,
NL or NLP = Natural Language Processing, NTM = Neural Topic Modeling, 

PDP = Partial dependence plots, PRA = Privacy Reference Architecture,
RAG = Retrieval-Augmented Generation, RL = Reinforcement Learning,
RLHF = Reinforcement Learning from Human Feedback,
ROUGE metric = Recall-Oriented Understudy for Gisting Evaluation,
RWK metric = Real World Knowledge score, SL = Supervised Learning,
SLM = Small Language Model, SQuAD benchmark = Stanford Question Answering Dataset,
U = Unlabeled Data, UL = Unsupervised Learning, 

WMT benchmark= Workshop on Machine Translation

Model Terms

Data efficiency = training with small, high-quality data rather than lots.

Diffusion models = GenAI that creates high-quality video/images by reversing noise. types: forward and reverse diffusion. 

Discriminative models = SL. discriminating doing classification or regression.

Forward Diffusion Model = hiding by adding noise till unrecognizable. Why? A: Maybe showing transform frame by frame. Similar to encryption.

GANs = generator and discriminator compete to create synthetic data.

Generative models = learns to create new content similar to training data.

Latent Space = hidden, vector cloud map of data with model's internal grasp of data relationships and semantic similarities.

LLM = a subset of FM. ex: Claude, Chat GPT, etc.

"Model" = software trained to recognize patterns and make predictions. Humans provide data and rules; computers find patterns and optimize weights.  

Model complexity = simpler models such as linear regression or decision trees are more explainable.

Neural networks = like brain. takes input then runs through hidden layer and outputs the answer.

Residual Neural Network = uses images and tries to "skip connections" method instead of CNN.

Pre-trained models = they don't need training data to get started using.

Reverse Diffusion Model = revealing. remove noise until clear image. Similar to decryption.

SLM = Rare. Used for edge devices.

Support Vector Machine = classifies tabular data; excels at high-dimensional, small datasets.

Transformer-based LLMs = Neural networks with parallel self-attention so processing multiple tokens together by positional encoding. Best for context in NLP, translation, generation of unique product descriptions, and summarization (e.g., BERT, GPT).

Inference (Doing) Parameters Terms

“Generation step” = controls how detailed (higher) or abstract (lower) image is.

Prompt Context Window parameter = affects number of tokens processed. small is fast. large is for more complex. 

Temperature of AI Output = randomness or creativity of the AI output. High = more chaos or creative. Set on model.

Top K Sampling = limits considered tokens to top K. Purpose: controlling diversity of output.

Top P/Nucleus Sampling = limits cumulative probability to top % of considered tokens. Purpose: controlling diversity of output.

Inference (Doing) Terms

Inference = when trained model views new data to calc the new output. affects latency speed.

Asynchronous Inferencing = ideal for workloads with large payloads (up to 1 GB), long processing times (up to one hour), or near real-time requirements.

Batch Inferencing/Batch Transform = start with lots of data and time.  Focuses on understanding all data to get answer(s). For slow decisions with historical understanding. Processes multiple inputs at once. Good for data analysis reports.

PDP = visualize the plot of one thing like age affects a second thing like income. good for transparency.

Real-Time Inferencing = Focus on newly arrived data. Good for real-time interactive data. EX: self-driving cars or missile defense systems). Old data just for context.
Serverless Inferencing = cheap option for intermittent or unpredictable traffic.

Shapley Values = each feature's value is its avg contribution to the prediction change relative to a baseline, uniquely satisfying 4 fairness properties: 1) efficiency, 2) symmetry, 3) dummy (null player), and 4) additivity.

Specialized App. Terms

Amplified Decision Making = helps humans in decisions in stressful times

CNN = DL for images; uses neural network filters to grid pixels (3x3 block) for pattern recognition. Filters scale from local to abstract. Ideal for computer vision, image classification, OCR, and medical imaging.

Generative Pre-trained Transformers = start with NLP text; translates to SQL or other.

Image Processing = processes vision (image/video) and time series (satellite/frames). Key areas: classification, object detection, and semantic segmentation.

RAG = pairs a dynamic retrieval system with Gen AI to search specific docs (company data/inventory) for answers. Used in chatbots, CRM agents, and legal/health analysis. Has offline content embedding and search indexing. Not for model training.

RAG Chunking = breaks large documents into small chunks. ensures only the most relevant snippets for a query, boosting the accuracy and precision of the response.

Sentiment analysis = NLP subset that identifies emotion/sentiment in text.

Text Analysis = processes text and speech.  Areas: classification, Word2Vec, and translation. Algorithms: BlazingText, Seq2Seq, LDA, and NTM.