Prueba de transformadores de visión

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

1. How does the original plain ViT convert an image into input tokens for its encoder?
   • It must read one raw pixel per recurrent step
   • It linearly embeds flattened non-overlapping patches
   • It converts the whole image into a mandatory single scalar
   • It removes all learned input projections
2. Which mechanism supports direct interactions between allowed patch tokens in the original dense ViT encoder?
   • Autoatención
   • Only a fixed local max-pooling operation
   • Only alphabetical ordering of patch labels
   • Only a random image crop
3. Which architectural difference did the original ViT study discuss relative to convolutional networks?
   • A guarantee that it contains fewer parameters than every CNN
   • An inability to process color images
   • A guarantee that every vision task needs the same huge dataset
   • Fewer convolution-style locality and translation-equivariance biases
4. ¿Cuál es el propósito de las codificaciones posicionales en un Vision Transformer?
   • Para comprimir el tamaño del archivo de imagen
   • Para mezclar aleatoriamente los parches para aumentarlos
   • Para decirle al modelo dónde estaba ubicado cada parche en la imagen original.
   • Para convertir imágenes en color a escala de grises
5. Un inconveniente práctico de la autoatención en muchos parches de imágenes es que su costo aumenta:
   • Cuadráticamente con el número de parches.
   • Linealmente con el número de parches.
   • Para nada con el tamaño de la imagen.
   • Sólo con la cantidad de canales de color.
6. What conclusion is supported by the original ViT classification experiments under their stated pretraining and transfer conditions?
   • Every transformer must outperform every CNN on every dataset
   • A pure patch transformer could perform competitively with compared CNNs
   • All convolutional operations are invalid for image recognition
   • No training data is needed for image classification
7. For a 224-by-224 image divided into non-overlapping 16-by-16 patches, how many patch tokens are there before adding a class token?
   • 197
   • 224
   • 196
   • 256
8. In the original ViT classification design, which encoder state supplies the classification head?
   • Only an unembedded raw input pixel
   • A required random token not processed by the encoder
   • Only the first patch’s original filename
   • The final state of the learned class token