Skip to content

Lecture12.md

Latest commit

 

History

History
172 lines (106 loc) · 6.66 KB

Lecture12.md

File metadata and controls

172 lines (106 loc) · 6.66 KB

📓 Applied AI Notes & Reflections

📅 Date: 25. october

🧑‍🏫 Instructor/Source: Henrik Strøm

🔍 Session Objective

Peer review feed back Principal component analysis

RAW Notes : Should be transformed ->Z

[ Jump in : henrik explains to output nodes _> when discussion of Softmax always the sum of 1 of a probability distribution -> and get a confidence in the result number / result -> you get to paramtric probabilty distribution in your output the probability sum is always 100 and when you have two output nodes in a binary classification _> so you have node1: 0.8 being a cat and node2:0.2 not a dog

TWO OUPUT NODES WITH A SOFTMAX -> SOFTMAX MAKES PROBABILITY DISTRIBUTION and can be interpreted as confidenc->

More than two ouput nodes are not needed for binary classification

MUST BE PRINTED TO BRAIN -> GROUND KNOWLEDGE FOR EXAM

CNN's:

- convolutional neual networks

###Features extraction : CNN is ment for image data -> and are really good at that

consists of poolings and convolutions -> look at a little area in picture 
pooling layer : feauture extraction from convultional layer 
                  -> try to enforce strongest signals and turn down weekeast 



    The outlining layers : will take care of contrasts ->
           innner is more detailed the longer we approach the inner layers

Features will become input to classfication network

MNIST -> steril data that benefit smaller from features extraction

Feature extraction is a strong toolset in terms of picture data

import tensorflow as tf
import tensorflow_datasets as tfds

(ds_test, ds_val, ds_train), ds_info = tfds.load(
    'mnist',
    split = [ 'test', 'train[0%:17%]

input image

convelutional kernel -> scans each picture pixel a small piece of the picture at the time ->

MAX Pooling reduces the and filters square don and uses stride 2 reduces 4 x 4 to 2 x 2

shuffle to remove the human errors in the loop

split by 42 to maske re

  • input layer is 28,28 -> goes in a "track" 384 -> 255 -> 128

  • Relu

  • adam adaptive learning rate adapts and finetune at end better hits optimum

  • Regularization -> we penalize

  • for the mnist the regularization does improve accruacy

  • so minmal difference in a high accuracy -> so it is not oiverfitted when the accuracy doesn't change

divide by 255 to have numbers between 0 - 1 bytes

Convolutional neural network using NOW

  • very very high accuracy

📝 Topic Overview

  • Topic: Peer review Principal componant analysis
  • Subtopics Covered:
    Bibliography with latex
  • Context in Course: How does this topic relate to previous sessions or future topics?

📚 Key Concepts

Concept Description/Definition Example/Context
Bibliography Find google scholar to generate latex bibliography
weights Generated randomely connecting line between perceptrons : reson being random : floats can give os better information as small numbers around zero Example or context to illustrate
Fully conected network All nodes are connect from input to hidden to ouput layer input circles connected to hidden to output layer are all connected on each other by lines that canno extend layers

🧠 Techniques, Algorithms & Models

Technique/Algorithm/Model Purpose Key Details Questions & Considerations
Technique 1 What it achieves or solves Important parameters, steps, or equations Any lingering questions, confusions, or considerations
Algorithm 2 What it achieves or solves Important parameters, steps, or equations Any lingering questions, confusions, or considerations

⚙️ Practical Applications

  • Use Case 1: Brief description of real-world application or industry use
  • Use Case 2: Brief description of real-world application or industry use
  • Additional Notes: Any additional thoughts on applications discussed
  • Current Industry Relevance: How is this applied AI concept evolving today? (Optional)

🔄 Connections to Prior Knowledge

Link this topic to any previous knowledge or related topics.

  • Concepts from Previous Sessions:
  • Related Personal/Professional Experiences:
  • Areas of Improvement or Clarification: What past concepts could you revisit to enhance your understanding?

💡 Key Takeaways & Insights

  • Main Insights:
  • Challenges & Limitations Discussed:
  • Personal Observations: Any unique observations or perspectives that arose while learning this content?

🎯 Action Steps for Reinforcement

  • Practice Goals: (e.g., code exercises, building a model, implementing an algorithm)
  • Reading/Research: Additional topics to read or research based on this session
  • Implementation Ideas: Real or hypothetical projects to apply this concept
  • Reflection on Learning: What strategies worked well today? Any adjustments for future learning?

🧩 Resources & Further Reading

  • Link 1: Title - Brief description of content
  • Link 2: Title - Brief description of content
  • Instructor’s Recommendations: Include any references or readings suggested by the instructor

❓ Questions, Reflections & Next Steps

  • Unanswered Questions: Any questions that remain unclear after the session
  • Reflective Thoughts: How has this session influenced your perspective on AI?
  • Topics for Deeper Dive: List topics you’d like to explore in more depth

📈 Tracking Progress (Optional)

Use this space to track your progress and self-assess your understanding.

  • Self-Assessment: Rate your understanding of today’s topic (1-5)
  • Progress Notes: Briefly note any progress since last session
  • Challenges Faced: Any specific difficulties that slowed down learning today