As mentioned, the LSTM architecture relies on a series of gates that can independently influence the activation values (a(t-1)), as well as the memory (c(t-1)), from previous timesteps as information flows through an LSTM unit. These values are transformed as the unit spits out the activations (at) and memory (ct) vectors pertaining to the current timestep at each iteration. While their earlier counterparts enter the unit separately, they are allowed to interact with each other in two broad manners. In the following diagram, the gates (denoted with the capital Greek letter gama, or Γ) represent sigmoid activation functions applied to the dot product of their respectively initialized weight matrix, with previous activations and current input:
