AI 900 dump

Vision

Language detection frame returns

• ISO6391 code
• Language name
• Score

Azure AI Document

• Identifying common data fields
  • Business card model
  • invoice model

Azure face

Face identification

• One to Many

Face verification

• One to One

Spatial analysis

• Analyze cctv video

• EX. Person count

• Azure AI Vision does not need to

  • choosing a model
  • evaluating model
  • training a model

Semantic segmentation

• Each pixel in image is placed on what kind of object

Custom vision

• Image recognition service with custom model
• NOT FOR VIDEO
• Object detection != multilabel/multiclass
• Is not a computer vision service

Face

• Tagging friends automatically in image
• Face attributes
  • Wearing a glasses or not

Cognitive service

• Image processing
• Content extraction
• NLP to knowledge mining
• IN SINGLE ENDPOINT

Specialized domain of categorizing

• Celebrities
• Landmarks

Confidence level

• The calculated probability of correct image classification
• Also included with each phrase returned by an image description
• NaN for unknown language name

Types

• Verification : Do two images of a face belong to the same person
• Similarity : Does this person look like other people
• Grouping : Do all the faces belong together
• Identification : Who is this person in this group of people

Machine learning

Feature : 재료, Label: 결과

• Model evaluation : Examining(조사하다) the values of a confusion matrix
  • Split data beforehand for this
• Feature engineering : Preprocess dataes
• Feature selection : Extract features from data

Automated machine learning

• Can not include custom python script in a training pipeline
• Can not visually connect dataset and modules
• First! Create a dataset on Azure machine learning studio

Azure machine learning designer

• Can drag module, dataset on canvas
• Can use linear regression
• First Create a pipeline

If features are independence

• Do multiple linear regression

Responsible AI

Transparency principle

• Provide documentation to help developers debug code
• Much more understandable AI
• Model that explains well

Inclusiveness(포괄성) | make AI that everyone can use

• Ensure that all visuals have an associated text that can be read by a screen reader

Privacy and security | Securely

• Enable autoscaling to ensure that a service scales based on demand

Fairness | Not biased

• Ensure that a training dataset is representative of the population(모집단)

Reliability and safety | Without any harmness

• Handling of unusual and missing values provided to an AI system

Speech

• Can be used to transcribe phone call to text

Conversational

Ex. QnA Maker, Azure bot service, Chatbot

• Pre-recorded one is not a conversational AI workload

Azure bot service / Conversational AI

• Cannot import FAQ to QnA set
• Createds knowledge base

QnA Maker

• Can use SQL Database as knowledge base
• Can add chit-chat content
• CAN NOT create a qna made by machine learning
• Can manually enter question ans answer
• Can retreieve data from FAQ
• To populate knowledge base, upload PDF
• Image/Audio can't be knowledge base

Copilot

• provide plugin

NLP

• Identify issues from support question data
  • Key phrase extraction
• Idenify any people and products mentioned
  • Named entity recognition

Stemming

• NLP technique normalize word before counting it

Removing stop words

• First step in the statistical analysis of terms for NLP
• remove words without meaning
• For example: a, an, the

Tokenization

• Speech synthesis NLP invoked
• Breaking text into individual worsd, can be assigned phonetic sound

Generative AI

Safety system

• Filter to suppress prompts and responses

System message

• Can used to identify constraint and style

GPT model

• Understand/Create Natural language

Embedding

• Search, Classify, Compare source of text for similiarity

Vectorization

Responsible AI

First step

• Identify potential harms

Image generation

• Can't describe image

Universal language model used by the speech to text model

• Optimized for conversational, Dictation

---

Real time inference pipeline

• Azure Kubernetes Service

Language Understanding

• LUIS
• Working with user's intent

Form regognizer

• automatically read information from images/PDF
• Identifies and extracts data from documents, organize it
• EX. extract text, table / key-value pair

Entity recognition

• Extract person, location, organization from the text
• Returns links to the external website to disambiguate terms

Classification Model

• use true positive rate to evaluate

Anomaly detection

• Identifying suspisous sign-in by looking for usual pattern

Azure language service

• Entity linking
• PII detection
• Sentiment analysis

---

• Speech service and translator service can be used both at the same time

---

여러개 있는데 그것들의 종류를 알아내는것

• Object detection