IIoT PDF 15: Machine Learning (Slides: 2 - 19)

Definitions and short notes

Generative AI [Slide: 2]

• Create new content, e.g. text, code, images, music, videos, design prototypes

• Learns patterns from large amount of data and use those patterns to produce something original

• Works with large language models (LLM) and diffusion models trained on massive datasets (e.g. books, websites, articles) to resemble the data they learned from

LLM - trained on large datasets (e.g. books, websites, articles), learn grammar, language structure, facts about the world, patterns in how people communicate, to resemble the data they learnt from.

Machine Learning [Slide: 2,4]

• Teaches computers to do what comes naturally to humans and animals: learn from experience.

• ML algorithms::

  • Use computational methods to “learn” information directly from data without relying on a predetermined equation as a model.

  • Can adaptively improve their performance as the number of samples available for learning increases

  • Find natural patterns in data that generate insight and to make better decisions and predictions

Artificial Intelligence → Machine Learning → Deep Learning

CHATGPT [Slide: 3]

At high level → an AI system built by OpenAI that uses a type of ML model called a Large Language Model (LLM)

Very advanced autocomplete system that understands context really well and can write full conversations instead of just finishing sentences

CHATGPT

How CHATGPT works? [Slide: 3]

1. Trained on lots of text

   • Model is trained on a huge collection of books, websites, articles, and other text

   • During training, it learns patterns like: • How sentences are structured • How ideas connect • What words usually come next in a sentence

   • Learns probabilities and patterns (doesn’t memorize everything)
     

2. Predicts the next word

   • When you type something, ChatGPT:

     • Breaks your input into pieces (called tokens)

     • Predicts what word (or token) should come next

     • Repeats that process step-by-step to form a full answer

   • (Given this context, what’s most likely the next word?)
     

3. Uses neural networks

   • Core system → Based on a neural network architecture called a transformer, which is especially good at understanding context in language:

     • It can relate words far apart in a sentence

     • It keeps track of meaning across paragraphs

     • It adjusts responses based on your question
       

4. Fine-tuned for conversation

   • After initial training, models like ChatGPT are improved using:

     • Human feedback (people rating answers)

     • Safety rules (to avoid harmful responses)

     • Instruction tuning (to follow user requests better)
       

5. Doesn’t “think” like a human

   • Even though it sounds natural, ChatGPT:

     • Doesn’t have beliefs or feelings

     • Doesn’t truly understand like a person

     • Doesn’t know things in real time unless updated.
       
       It is generating responses based on patterns—not conscious thought.

MACHINE LEARNING

In what fields ML is used in? [Slide: 4]

• Natural language processing

• Image processing and computer vision

• Computational biology

• Computational finance

• Energy production

• Automotive, aerospace and manufacturing

• E-Commerce or media

• Retail

MACHINE LEARNING

General Manufacturing tasks where ML is helpful [Slide: 5]

1. Typically: Machine Learning

   1. Forecasting

   2. Planning

   3. Security

   4. Robot programming

   5. Process Optimization
      

2. Typically Gen AI

   1. PLC Programming

   2. UI development

   3. Code analyzing, testing, troubleshooting

   4. Specification

MACHINE LEARNING

When to use ML? [Slide: 6]

• For a complex task/problem involving a large amount of data and lots of variables

• When the task or problem cannot be solved easily with another algorithm e.g:

  • Hand-written rules and equations are too complex (e.g. face and speech recognition)

  • Rules of a task are constantly changing (e.g. Fraud detection from transaction records)

  • Nature of the data keeps changing, and the program needs to adapt (e.g: automated trading, energy demand forecasting, and predicting shopping trends.)

MACHINE LEARNING

Benefits of ML (Example) [Slide: 7]

Problems with Conventional Object Recognition:

Processing of data via programmed image capture system

• Each object to be recognized has to be precisely defined (deviations = rejection)

• Time-consuming programming for new objects

Benefits of ML using AI:

Processing of input data via neural networks

• Higher availability through detection of complex patterns

• Easier handling also of unknown objects

MACHINE LEARNING

Comparison b/w standard programs and ML [Slide: 8]

Traditional:
Data and Program (input) → Computer → Output

Machine Learning:
Data and Output (Input) → Computer → Program

MACHINE LEARNING

Types of ML [Slide: 8]

• Supervised learning

• Unsupervised learning

• Reinforcement learning

• Semi-supervised learning

Enhancement

• Transfer Learning

Supervised Learning [Slide: 9]

• Aim of S ML: To build a model that makes predictions based on evidence in the presence of uncertainty.

• Supervised Learning Algorithm → Takes a known set of input data and known responses to the data (output) and trains a model to generate reasonable predictions for the response to new data.

SUPERVISED LEARNING

Techniques used to develop predictive models [Slide: 9]

1. Classification Techniques

   • Predict discrete responses (e.g: whether an email is genuine or spam, or whether a tumor is cancerous or benign.)

   • Classify input data into categories

   • Typical Applications

     • Medical imaging

     • Speech recognition

     • Credit scoring
       

2. Regression Techniques

   • Predict continuous response (e.g. changes in temperature or fluctuations in power demand)

   • Typical Applications

     • Electricity load forecasting

     • Algorithmic trading

Binary Classification vs. Multiclass Classification [Slide: 10]

• Binary Classification
  Classifying elements of a set into two groups on the basis of a classification rule.
  
  E.g. Scrap detection during quality inspection of the manufacturing of the relay in the AutFab
  Categorize the product into one of two groups: “Scrap” or “Yield”
  
  

• Multiclass/Multinomial Classification
  Problem of classifying instances into one of three or more classes.
  
  E.g. Quality inspection of the manufacturing of the relay in the AutFab
  class 0: Pass,
  class 1: Screws of socket unscrewed,
  class 2: Relay pin bended,
  class 3: Printed circuit board of protection module missing, …
  

SUPERVISED LEARNING

Phases of Supervised Learning [Slide: 11]

1. Training Phase of a model

   • ML algorithm is trained
     

2. Online/inference Phase

   • Signals are made available online to the deployed trained model, which uses them to make predictions.

SUPERVISED LEARNING

Features [Slide: 12]

• Features → Extracted from the input data

  • Feature: Individual measurable property or characteristic of a phenomenon being observed

  • Usually numerical, but can also be strings, graphs, etc.

SUPERVISED LEARNING

Labelled Outputs [Slide: 12,13]

• Expands each piece of unlabeled data with meaningful tags (e.g. whether the signal indicates a specific production line failure, a photo contains a speed sign of 100 km/h or which words were uttered in an audio recording)

• Labels can be obtained by asking people to make judgements about a particular piece of unlabeled data.
  
  
  

• SL algorithm → Trained by pairs of Features and Labels

• Model can then make a prediction of the Label from the online data, because it knows the Feature-Label pairs from the training.
  
  

Unsupervised Learning [Slide: 14,15]

• Finds hidden patterns or intrinsic structures in data

• Used to draw inferences from datasets consisting of input data without labeled responses.

• ML without predetermined targets and without rewards during the training.

• The algorithm is trained only by Features.

• UL tries to detect previously unknown patterns in the input data.

• Goal: to find correlations between the Features and to group them in clusters.

UNSUPERVISED LEARNING

Uses [Slide: 14]

1. Clustering

   • Most common unsupervised learning technique.

   • Used to find groupings in data

   • Applications: Gene sequence analysis, Speech recognition.
     

2. Anomaly detection

   • Used for identification of rare items, events or observations differing significantly from the majority of the data

   • Applications: Bank fraud detection, Structural defect analysis, Medical diagnosis
     

3. Association discovery

   • Used for exploratory data analysis to find hidden patterns

   • Applications: Market research

Reinforcement Learning [Slide: 16]

• Algorithm tries different actions on a problem and checks the reward it receives against the status it reaches

• Tries to maximize the rewards as long as the status update is positive

Supervised Learning vs Semi-Supervised Learning vs Unsupervised Learning [Slide: 18]

• Supervised Learning
  All data is labelled (training data) → Model
  

• Semi-Supervised Learning
  Small portion of data is labelled & Lots of data is unlabelled → Model
  

• Unsupervised Learning
  All data is unlabelled → Model

What problems can be solved with ML? [Slide: 19]

Machine Learning

• Supervised Learning

  • Classification

  • Regression
    

• Unsupervised Learning

  • Clustering (find similar instances)

  • Anomaly detection (find unusual instances)

  • Association Discovery (find feature rules)