Playing with "Biologically plausible" neural network models. Such as Hebb/Anti-hebbian networks and reinforcement learning.
I'm not sure where this is heading yet, but I'm basing a lot of this on previous work by Hu, Pehlevan & Chklovskii (2015), Pehlevan & Chklovskii (2014), Pehlevan & Chklovskii (2014)
And Q-AGREL developed by Pozzi, Bohté & Roelfsema, 2018.