Questionário de aprendizagem auto-supervisionada

As respostas e explicações aparecem conforme você responde ao teste.

1. Where does the supervision signal in the described self-supervised examples come from?
   • A requirement for no training objective
   • Only new human semantic labels for every pretraining example
   • A rule that examples contain no useful information
   • The data and defined transformations of it
2. Which input positions contribute prediction targets to the original BERT masked-language-model loss?
   • The positions selected for the MLM prediction task
   • Every position in the entire input regardless of selection
   • Only the first position of each new sentence
   • Only positions that appear after the final sentence boundary
3. Na aprendizagem contrastiva como o SimCLR, um 'par positivo' é:
   • Duas imagens completamente não relacionadas
   • Um par de rótulos escritos por humanos
   • Duas visualizações aumentadas da mesma imagem que devem estar próximas no espaço de incorporação
   • Duas cópias do conjunto de dados
4. What annotation requirement can self-supervised pretraining reduce in the described examples?
   • The need for any usable input data
   • Separate task-specific semantic labels for every pretraining example
   • All computational cost by definition
   • The need to assess downstream performance
5. How can a representation learned through self-supervision be used for a later task?
   • By evaluating a task head or adapting the representation through fine-tuning
   • By assuming every target task requires no labels or validation
   • By discarding all representations before making predictions
   • By treating the pretraining loss as proof of every application's accuracy
6. Como a tarefa autosupervisionada de um modelo estilo GPT difere da do BERT?
   • Não usa nenhum texto
   • Requer rótulos manuais para cada palavra
   • Ele prevê apenas a primeira palavra de um documento
   • Ele prevê o próximo token em vez de preencher tokens mascarados
7. In the original SimCLR framework, which component is discarded for the reported downstream representation use?
   • The encoder representation itself
   • The contrastive projection head
   • All training inputs before pretraining begins
   • The positive-pair source-image relationship
8. When original BERT pretraining replaces a selected input position with a random token, what remains its loss target?
   • The newly inserted random token
   • The next sentence's first token in every case
   • The original token before replacement
   • No target, because every random replacement is omitted from training