Glossario IA
Il dizionario completo dell'Intelligenza Artificiale
GRU (Gated Recurrent Unit)
Recurrent neural network architecture introduced in 2014 that simplifies the LSTM structure by combining the forget and input gates into a single update gate, thus reducing the number of parameters while preserving performance.
Update gate
Control mechanism of the GRU that determines how much information from the previous state should be retained and how much new information should be added to the current state.
Reset gate
Essential component of the GRU that controls how much of the previous state should be forgotten when calculating the new candidate state, allowing the model to reset its memory when necessary.
Hidden state
Memory vector in a GRU that encodes relevant information from previous time steps and is transmitted to each new step to maintain temporal context.
Candidate vector
Intermediate representation calculated in a GRU that contains potential new information to be added to the hidden state, weighted by the update gate.
Gating mechanism
Regulation system in GRUs using sigmoid gates to selectively control the flow of information through the network, simulating a form of selective memory.
Temporal propagation
Process by which GRUs process sequences by propagating information from one time step to another through their hidden states and gating mechanisms.
Gate parameters
Weight matrices and bias vectors specific to the update and reset gates in a GRU, learned during training to optimize the control of information flow.
Previous state
Value of the hidden state at time step t-1 used as input to calculate the new state at time step t in a GRU, essential for maintaining temporal continuity.
Information fusion
Operation in GRUs where the update gate linearly combines the previous state and the candidate vector to produce the new hidden state, weighting their relative importance.
Computational complexity
Advantage of GRUs over LSTMs with approximately 33% fewer parameters, resulting in faster training while maintaining comparable performance on most tasks.
Vanishing gradient
Learning problem in traditional RNNs that GRUs mitigate through their gating mechanisms that preserve the gradient over long temporal sequences.
Bidirectional sequence
Architecture using GRUs in the forward and backward directions of a sequence, capturing past and future context for better understanding of sequential data.
Dropout regularization
Technique applied to GRUs where certain connections are randomly deactivated during training to prevent overfitting and improve generalization.