Quiz de réseaux neuronals récurrents

Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.

1. What recurrent information does a conventional unidirectional RNN combine with each new input?
   • Only the previous printed prediction with no internal state
   • A fresh unrelated dataset at every step
   • The previous hidden state
   • No parameters or intermediate information
2. How should an RNN hidden state be understood?
   • An evolving internal representation that may summarize earlier inputs
   • A guaranteed lossless archive of every earlier input
   • The network's entire training dataset on disk
   • Only a final class label with no intermediate role
3. Which training difficulty can arise from propagating RNN gradients through many time steps?
   • A guarantee that every hidden state grows without bound
   • Vanishing or exploding gradient magnitudes
   • An automatic change from speech to image inputs
   • Permanent deletion of the training objective
4. naka la LSTMs ak GRUs di yokku ci RNNs yu leer yi?
   • Dañuy dindi bépp mémoire ci reso bi
   • Li ñuy liggéey ci nimero kese lañuy liggéey, du ci mbind
   • Dañuy tàggat te duñu am benn done
   • Dañuy jëfandikoo buntu yi ngir saytu ban leeral lañu wara denc, yeesal wala sànni
5. Ban xeetu done lañu jagleel RNN yi?
   • Done yu toppalante lu melni mbind, wax, ak seeni seeri waxtu
   • Benn nataal bu amul komànd
   • Nimero yuñ jagleel
   • Tabax bu am ñatti dimension kese
6. What training-computation difference does the original transformer architecture offer compared with a conventional sequential RNN?
   • Guaranteed fully parallel autoregressive generation
   • A guarantee that no training examples are needed
   • More parallel computation across training positions and direct attention interactions
   • A model with no learned parameters
7. In the stated linear recurrent illustration h-new equals one-half h-previous plus x, what is h-new for h-previous two and x three?
   • Juroom
   • Benn
   • Juróom ñaar
   • ñeent
8. Why does a fixed-size RNN inference state not establish constant memory for full-sequence training?
   • Training never uses intermediate values
   • Backpropagation through time may store or recompute information across many steps
   • Every model deletes its parameters before learning
   • A fixed state guarantees every historical input is retained exactly