←
Daniel Bendor201 posts
Daniel Bendor avatar

Daniel Bendor

@BendorD

Professor in Behavioural Neuroscience, UCL

◷ Joined December 2013▧ 12 Photos and videos
Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

Big week for the lab: three new preprints out, spanning prediction errors, replay, and sleep oscillations. All from work led by fantastic students and postdocs. Threads below 👇

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

Huge thanks to everyone whose work is in these three preprints ♥ — Prez Jarzebowski, Margot Tirole, Eleonore Duvelle, Masahiro Takigawa, Diao Tong, Edward Horrocks, and Aman Saleem. So proud of this work from the whole team.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

Surprising twist: cortical activity before the hippocampal ripple already predicts which memory reactivates. Cortex isn't just a passive receiver of the "teaching signal," it actively nominates which memory can get consolidated.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

We split two VR tracks across left/right mouse visual cortex, then recorded V1 + hippocampus overnight. Ripple power, local spindle power, and slow-oscillation phase each independently boost cortico-hippocampal reactivation coherence, and coherence peaks when all three align.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

We propose each local awake replay event leaves a memory tag that decays over time, but remote replay of that episode keeps the tag alive. The cumulative amount of local + remote replay, weighted by proximity to sleep onset, most accurately predicts replay priority during sleep.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

Once time spent immobile was equalised across conditions, reward's apparent effect on replay rate vanished. What mattered instead was recency: the most recently experienced episode dominated sleep replay, not whichever had accumulated the most local replay.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

next, led by @MargotTirole with Eleonore Duvelle: does reward really drive what hippocampus "chooses" to replay during rest and sleep? We gave rats high- vs low-value reward across 3 sequential novel tracks and tracked CA1 replay all day.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

Two different E/I imbalances in the same circuit produce two distinct failure modes, resembling autism-like sensory overload and schizophrenia-like sign errors respectively - and the model's predictions hold up against mouse V1 Neuropixels data.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

Acting as a "temporal clamp", chandelier cells gate L2/3 output onto its downstream target, L5. Underpredicted inputs escapes the clamp early, and overpredicted inputs are released with a delay. this timing affects plasticity outcomes downstream in L5 neurons.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

So how does a single circuit tell these two error types apart and route them to opposite plasticity outcomes? We propose that chandelier cells act as a timing gate, turning spike delay into the sign of the error.

Daniel Bendor avatar
Daniel Bendor@BendorD·Jul 3

Our brains constantly predict what should happen next. But an overprediction (expecting something that doesn't come) and an underprediction (being surprised by something unexpected) are opposite kinds of mistakes, and should drive learning in opposite directions.