피지컬 ai 1

Main ideas / concepts covered

• Teachable Machine (Google) is presented as an easy way to build simple AI classifiers without deep machine learning knowledge.
• The video demonstrates three project types provided by Teachable Machine:
  1. Image project (visual classification)
  2. Audio project (sound classification, including background noise handling)
  3. Pose/action project (movement recognition using body poses)
• A core lesson repeated throughout: don’t just copy-paste code or rely entirely on automation—try to understand what the system is doing and build something meaningfully yourself.
• The presenter encourages iteration:
  • Increase classes
  • Collect more/diversified data
  • Adjust backgrounds/noise
  • Experiment to improve results

---

Methodology / step-by-step instructions (as described)

1) Image project (Teachable Machine – “standard image model”)

• Open Google Teachable Machine (search it on Google and start it).
• Select project type:
  • Choose Image Project
  • For the current purpose, select Standard image model (the “lighter/other model option” is mentioned as useful later)
• Sign in / account:
  • Use a Google ID to enter and create a project.
• Set up capture permissions:
  • If prompted, allow camera access (camera on laptop).
  • The video also notes you can use existing photos, but demonstrates webcam practice.
• Create classes:
  • The UI shows Class 1 and Class 2.
  • Rename classes to your own labels (names can be anything; the model learns the categories).
• Collect training examples:
  • Press and hold the capture/record button to take multiple photos for a class.
  • The presenter demonstrates taking images (e.g., changing/moving an attribute) to gather diverse samples.
• Train the model:
  • Click Train.
• Optional advanced training settings:
  • Open Advanced to adjust:
    • Epochs: how many times the model iterates over the full dataset (rounds it reviews)
    • Batch size: how many samples are used at once during training (SGD mini-batches)
    • Learning rate: how much the model updates during learning
  • Default settings are suggested as a good starting point.
• Evaluate results:
  • Look for accuracy increasing and loss decreasing (graph shown).
  • Test the trained model in real time.
• Export / use the model elsewhere:
  • Use Export Model.
  • Export can be converted for use with Keras or OpenCV.
  • Copy and run provided code to see the same behavior in the train/export workflow.
• Environment / dependencies mentioned:
  • If using Python libraries like OpenCV and TensorFlow, the presenter mentions using uv for environment management.
  • If needed: install packages via pip (the text mentions “PP Install,” likely referring to pip install).

---

2) Audio project (including background noise)

• Start the Audio Project (the video says it looks similar to the other projects).
• Understand the key feature:
  • The system captures background noise that exists during recording.
  • It records sound continuously while the user holds the record button.
• Record and create classes:
  • Record for about 20 seconds for a class sample (the system auto-extracts the sample).
  • Example class mentioned: “when it is still.”
• Record in intervals:
  • Example described: recording in 2-second intervals, then using labels/classes for training.
• Train and test:
  • Test whether the model can distinguish background noise vs. a particular event/action/sound.
  • The presenter claims it can reach very high accuracy (around 100% in the demo).
• Use the model on a website / via live server:
  • Use model upload / copy provided code.
  • Replace the model link in HTML, then serve with Live Server to run it on a PC.
  • The model can be used as a link after setup.
• Encouraged audio experiments:
  • Add more classes (e.g., rock/paper/scissors sounds).
  • Collect longer recordings (e.g., 30 seconds, 3 minutes).
  • Change sound types and background conditions.
  • Examples from students:
    • bus arrival sound
    • car brakes vs driving
    • heartbeat recording
    • distinguishing between voices (speech-related classification)

---

3) Pose / action project (“sports project” in the video)

• Start the Pose/action project.
• Define the movements you want to recognize:
  • Example: standing vs “puja hands” (hands held together/prayer gesture).
• Collect examples:
  • Record multiple samples for each action class (hand gesture vs standing motion).
• View performance using a confusion matrix:
  • The video explains the confusion matrix:
    • Correct predictions are on the diagonal
    • Errors appear off-diagonal
• Train, validate, and iterate:
  • The demo shows strong discrimination (“perfectly” distinguishing once trained).
  • It also shows accuracy/loss improving.
• Export and run on a laptop using live server:
  • Upload/create a runnable format similar to audio.
  • Instructions mentioned: rename folder, move files, right-click and use Live Server.
  • Start the project and see the recognized hand/movement appear.
• Encouraged applications:
  • Students used it for games (rock/paper/scissors-like mechanics).
  • Examples include exercise tracking and even sign language recognition using poses.

---

Key lessons emphasized

• Try at least one project (pose, audio, or image) rather than only watching.
• Improving results comes from:
  • adding more classes
  • collecting more data
  • diversifying samples (different angles, conditions, backgrounds)
  • changing environments/noise
• Even in an era where AI writes code, users should aim to understand what the code lines mean and how the logic flows, rather than treating everything as a black box.

---

Speakers / sources featured

• Teachable Machine / Google (tool) — Google’s Teachable Machine platform is the main source/tool demonstrated.
• Claude / AI coding assistant — mentioned as an example (“if I ask Claude to code it for me…”), but not shown producing the content.
• The video presenter — primary speaker (no name provided in subtitles).

---