iiinova — where your digital self keeps evolving

Viktor Studenyak

@vstudenyak.bsky.social

PhD student in computational neuroscience @ Max Planck Inst. for Human Brain and Cog.Sciences || interested in n multi-level modelling, brain encoding, hippocampal formation, dentate gyrus

Posts Replies Media

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

Together, the model reframes the DG as a dual system: DG_SUP separates individual experiences, while DG_INF extracts common structure across them. 9/9

030

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

Finally, blade-specific prediction errors provide a possible control signal. We suggest that depending on whether information is represented by DG_SUP, DG_INF, both or neither, the system can separate new inputs, integrate them with existing representations, or selectively forget them. 8/9

130

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

The model generates experimentally testable predictions: • DG_INF place fields should be larger than DG_SUP fields • DG_SUP should show stronger episode-specific remapping • DG_INF integration should depend on similarity between experiences 7/9

120

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

With spatial inputs, DG_SUP formed small, sparse place fields that remapped strongly, whereas DG_INF produced broader, more stable fields. The two regimes qualitatively reproduce a key observation from Hainmueller & Bartos(2018): stable and dynamic spatial representations can emerge in parallel. 6/9

120

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

We tested the model with two very different inputs: MNIST and simulated entorhinal-cortex activity, including grid, object-vector, identity and random spatial cells. The same separation–integration distinction emerged in both domains. 5/9

120

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

In the model, suprapyramidal DG (DG_SUP) uses competitive k-winner-take-all dynamics and rapid plasticity to form episode-specific representations. Infrapyramidal DG (DG_INF) learns more gradually and integrates related episodes. 4/9

120

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

Experimental work proposed that separation and integration may be biased toward different DG blades (Berdugo-Vega et al., 2023). We provide a computational account of how blade-specific architectures and learning rules could generate this division. 3/9

120

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

The dentate gyrus is usually framed as a pattern separator: similar experiences are transformed into distinct representations. But experiments also find stable, integrated representations across related episodes. How could both functions coexist in the same structure? 2/9

120

Viktor Studenyak @vstudenyak.bsky.social · 07/09/2026

New preprint from my PhD at MPI CBS 🧠, with @doellerlab.bsky.social and @andrejbicanski.bsky.social “A computational model of the two dentate gyrus blades” What mechanisms allow the two DG blades to support distinct coding regimes? (www.biorxiv.org/content/10.6...) 1/9

1167

Reposted by Viktor Studenyak

Andrej Bicanski @andrejbicanski.bsky.social · 22/05/2025

Congrats Volker (‪@reisnerv.bsky.social‬). Collab with Doellerlab, modeling by Viktor Studenyak via the Neural Computation Group

073