Recurrent Networks

Check your understanding

4 questions · answer all to submit

1. 1.What fundamental limitation of standard feedforward neural networks is addressed by Recurrent Neural Networks (RNNs)?

   AFeedforward networks are unable to perform gradient-based optimization due to their fixed architectureBFeedforward networks lack an inherent mechanism to capture temporal dependencies or maintain state across sequential inputsCFeedforward networks struggle with processing non-numeric, categorical data inputs effectivelyDFeedforward networks cannot generalize to unseen input patterns without extensive retraining

2. 2.What conceptual role does the 'hidden state' play within a Recurrent Neural Network (RNN) architecture?

   AIt serves as a static, pre-defined activation function applied universally across timestepsBIt is a randomly initialized vector used solely to introduce stochasticity into the network's computationsCIt represents the final output prediction of the network at the current processing timestepDIt acts as an internal memory vector that aggregates and encodes information from all preceding sequence elements processed so far

3. 3.Regarding the weights in a recurrent neural network (RNN) across timesteps, which statement is accurate?

   AA distinct and entirely independent set of weights is allocated for each timestep.BA unified set of weights is consistently shared and applied across all computational timesteps.COnly the initial timestep employs learnable weight parameters, with all subsequent steps fixed.DWeights are dynamically reinitialized at every single timestep, based on immediate input data.

4. 4.Prior to the widespread adoption of transformer architectures, which neural network architecture predominantly excelled in tasks such as machine translation and speech recognition?

   ARecurrent Neural Networks (RNNs), specifically long short-term memory (LSTM) variants, excelled.BGenerative adversarial networks (GANs) combined with attention mechanisms were widely used.CDeep convolutional neural networks (CNNs) without any recurrent components were predominant.DShallow feedforward networks with extensive regularization techniques were the primary architecture.