@benjocowley.bsky.social @david-klindt.bsky.social
@dinanthos.bsky.social
Would love to hear what you think!

any chance I can still get the call link?
registrations are closed...

Now!

Or more likely, people should change what they think place cells mean.

Will give it a try! Thanks!!

Can't such much smart on CA2/3, will have to read up!
My only thought was that if (CA1) place cells "respond to the sequence of sensory inputs leading to a location",
then similar snapshots of activity will most often encode different things.
PCA only sees the instantaneous patterns, no sequences.

The paper doesn't look at place cells explicitly, but assuming everything transfers: at what point do we decide not to call it a place cell anymore?

I was going to comment about a biologically plausible implementation, but it seems you've heard that one enough.
As someone that looks at recordings all day, I'm constantly thinking about what signatures these models would leave behind.
I wonder if there's anything particular for CSCG.

2. I have found it hard to learn embeddings of space from hippocampal/MEC recordings in a 2D task.
In 1D, it is simpler because the trajectory to reach any position is always the same.
PCA/UMAP/CEBRA/... do not account for history.
Does the sequence hypothesis suggest that we need history to do it?

1. People are training RNNs with the same inputs as CSCG, either to predict the next visual input, or using RL. Neurons in those RNNs show spatial tuning.
Do you think that the CSCG graph is encoded in those RNNs? Is there a way to verify if the RNN is doing what CSCG does?

Great visualisations!
They have influenced my thinking.

some (potentially naive) questions I have: