Cuestionario vinculado a una guía · Medio Nivel

Prueba de entrenamiento en el momento del examen

Compare la adaptación en tiempo de prueba y las capas de secuencia TTT, incluidas las pérdidas autosupervisadas, los estados del modelo interno, el escalado lineal y el aprendizaje de bucle externo.

Rutas de guía relacionadasEntrenamiento de tiempo de prueba
Pregunta 1 de 8

What happens in the cited classic test-time training adaptation procedure before the main prediction?

Vista previa de cada pregunta8 preguntas

Las respuestas y explicaciones aparecen a medida que realiza el cuestionario.

  1. What happens in the cited classic test-time training adaptation procedure before the main prediction?
    • The unknown main-task label is supplied as a correct training target
    • The model must delete all parameters
    • Every earlier training example is reloaded without adaptation
    • Model parameters are updated using the incoming unlabeled sample
  2. ¿Por qué el TTT clásico depende de una tarea auxiliar autosupervisada?
    • Porque la autosupervisión es siempre más precisa que la tarea principal
    • Debido a que las etiquetas no están disponibles en el momento de la prueba, necesita una pérdida que pueda calcular sin ellas.
    • Porque elimina la necesidad de cualquier red neuronal.
    • Para cifrar los datos de prueba
  3. What forms the recurrent hidden state of a TTT sequence layer?
    • The parameters of an inner model updated through learning
    • Only an unchanging raw input filename
    • A mandatory list of all previous full attention matrices
    • The vocabulary sorted alphabetically
  4. Under fixed inner-model dimensions and update work, how does TTT sequence-layer complexity scale with sequence length?
    • Quadratically for every possible implementation
    • Independently of processing any tokens
    • Linearly in sequence length
    • Exponentially because every token retrains from all possible sequences
  5. What problem motivated the cited classic test-time adaptation study?
    • A requirement to increase the alphabet size
    • A difference between training and test input distributions
    • A guarantee that all deployment inputs match training
    • A need to sort source filenames faster
  6. Why is a decrease in an auxiliary adaptation loss insufficient by itself to establish a successful main-task update?
    • Every auxiliary loss is the unknown main-label accuracy
    • A loss value proves perfect safety on every future sample
    • The auxiliary objective and the intended prediction measure are different
    • An auxiliary task cannot be computed on unlabeled input
  7. What inner models distinguish the studied TTT-Linear and TTT-MLP layer variants?
    • A decision tree and a mandatory image tokenizer
    • Two copies of the same fixed dictionary with no parameters
    • A full historical attention matrix and an alphabetical index
    • A linear model and a two-layer multilayer perceptron
  8. In the sequence-layer framework, which loop learns the parameters defining the inner self-supervised task?
    • The outer training loop
    • Only the unknown future test label
    • A required manual dictionary edit at every token
    • Neither loop can learn any task parameters