Kvíz spojený s průvodcem · Střední úroveň

PANNs: Předtrénovaný kvíz o audio neuronových sítích

Test PANNs AudioSet pretraining, frozen versus fine-tuned features, weak timing labels, rare-event transfer, and task-specific heads.

Otázka 1 z 8

What training source does the original PANNs work use for broad audio representations?

Náhled každé otázky8 otázky

Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.

  1. What training source does the original PANNs work use for broad audio representations?
    • Only transcript-aligned conversational speech
    • AudioSet sound-event examples
    • Only isolated source stems from studio songs
    • Only unlabelled sensor-calibration sweeps
  2. A team needs a new machine-alarm classifier. How can PANNs be used?
    • Treat the existing web tagger as automatically validated
    • Use only the model name as a label
    • Adapt pretrained features with a task-specific head or fine-tuning
    • Skip target-domain audio tests
  3. Why might a frozen encoder underperform full fine-tuning on a very different acoustic domain?
    • Freezing always deletes the weights
    • Fine-tuning cannot use labels
    • The sample rate automatically doubles
    • Its features cannot adjust to that domain
  4. What AudioSet limitation matters for event-timing claims?
    • Clip tags do not directly mark exact onset and offset
    • It records no audio at all
    • Every clip contains one event only
    • All labels are word transcripts
  5. Which task is not delivered by an audio tagger alone?
    • Predicting broad event presence from a clip
    • Isolating a clean dialogue waveform from a music mix
    • Producing scores for a scene classifier
    • Reusing an encoder for another tagger
  6. Why test a PANN-based detector on the actual factory microphone?
    • Microphones cannot change audio features
    • A trained head guarantees every domain
    • The checkpoint license determines accuracy
    • Web-audio pretraining may not cover that sensor and noise pattern
  7. Which comparison tests the adaptation choice fairly?
    • Frozen and fine-tuned models on the same held-out target recordings
    • Only training accuracy for the fine-tuned model
    • Only the pretrained model’s paper score
    • Only file size of the two checkpoints
  8. What should be reported for a rare warning sound?
    • Only total AudioSet classes
    • Only average training loss
    • Per-class misses and false alarms, not just overall mean accuracy
    • Only model parameter count