Hidden units and hidden states AI
A hidden unit is one coordinate of an internal representation. A hidden state is the whole vector at one time step. In the activity report, that vector has 128 values; it is neither a set of 128 activities nor a set of physical sensors.
At one checkpoint, a window produces an array with 128 times × 128 units. The inspector retains every fifth time sample plus the final one: 27 states. A heatmap column is one saved state; a row follows one unit through these times. The tile grid displays the selected column, while the line plot follows the selected row. The original full history has equal-length time and unit dimensions, but they mean different things. The exported display is 27 times by 128 units.
A unit can combine information from all nine sensor inputs and the preceding state. Following the same unit number across checkpoints follows a coordinate in this particular model, not a guaranteed stable semantic feature. A different training run can organize its representation differently. Large response magnitude alone does not establish importance to a prediction.
The PyTorch GRU reference defines sequence outputs and final states. For a different setting, see language models and latent spaces.