Gated Chemical Units

We introduce Gated Chemical Units (GCUs), a new type of gated recurrent cells which provide fresh insights into the commonly-used gated recurrent units, and bridge their gap to biologically-plausible neural models. We systematically derive GCUs from Electrical Equivalent Circuits (EECs), a widely adopted ordinary-differential-equations model in neuroscience for biological neurons with both electrical and chemical synapses. We focus on saturated EECs, as they are more stable, and chemical synapses, as they are more expressive. To define GCUs, we introduce a new kind of gate, we call a time gate (TG), in the associated difference-equations model of the EECs. The TG learns for each neuron the optimal time step to be used in a simple Euler integration scheme, and leads to a very efficient gated unit. By observing that the TG corresponds to the forget gate (FG) in traditional gated recurrent units, we provide a new formulation of these units as neural ODEs. We also show that in GCUs, the FG is in fact its liquid time constant. Finally, we demonstrate that GCUs not only explain the elusive nature of gates in traditional recurrent units, but also represent a very competitive alternative to these units.