Recurrent Network
An artificial neural network structure with feedback connections, in which the current state influences future states, enabling internal memory and causal loops that may give rise to irreducible causal interactions.
Timeline (last 90 days)
Convolutional neural networks implicitly assume "locality" and "translation equivariance", decision trees prefer axis-aligned splits, and recurrent networks assume temporal dependencies in sequences.
In the random networks constructed in the paper, a recurrent network and a feedforward network can have exactly the same input-output response latency law, yet their φ values are 0.28 bits and 0 bits, respectively.
Behavior alone cannot distinguish systems with different internal causal structures, but internal recordings require only 4-53 trials to separate the above recurrent network from the feedforward network.
Recurrent networks have feedback connections, and their current state affects future states; feedforward networks transmit information in only one direction with no feedback loops, so a feedforward network can be fully decomposed into independent paths.
Recurrent networks possess hidden states, while reinforcement learning agents implicitly carry a world model through value functions and transition dynamics.
The state update of recurrent networks is recursive, with serial dependencies between time steps, making it hard for GPU parallel efficiency to compete with Transformer, and this bottleneck may become more pronounced as model scale grows.
All Facts (6)
Convolutional neural networks implicitly assume "locality" and "translation equivariance", decision trees prefer axis-aligned splits, and recurrent networks assume temporal dependencies in sequences.
50%UnverifiedIn the random networks constructed in the paper, a recurrent network and a feedforward network can have exactly the same input-output response latency law, yet their φ values are 0.28 bits and 0 bits, respectively.
50%UnverifiedBehavior alone cannot distinguish systems with different internal causal structures, but internal recordings require only 4-53 trials to separate the above recurrent network from the feedforward network.
50%UnverifiedRecurrent networks have feedback connections, and their current state affects future states; feedforward networks transmit information in only one direction with no feedback loops, so a feedforward network can be fully decomposed into independent paths.
50%UnverifiedRecurrent networks possess hidden states, while reinforcement learning agents implicitly carry a world model through value functions and transition dynamics.
50%UnverifiedThe state update of recurrent networks is recursive, with serial dependencies between time steps, making it hard for GPU parallel efficiency to compete with Transformer, and this bottleneck may become more pronounced as model scale grows.
50%