Colin Bredenberg

We do address this in the discussion! In short, replay occurs in both REM and SW sleep, and is important for consolidation--REM replay just isn't as well-studied afaik. We think the evidence is a little stronger for this mechanism acting during REM sleep, but we cite papers supporting both phases.

There are closely related models that include interneuron subtypes--I'll email you some papers that discuss the role different interneuron subtypes might play in these kinds of learning algorithms. We'd have to adapt the model to include them, but it shouldn't be difficult.

You're right about top-down vs. bottom-up being poorly defined, esp. in association cortex. It is probably better to view this in terms of apical/basal inputs instead, and to choose a cortical area where apical inputs (e.g. higher-order thalamic inputs) have distinct (multimodal) response properties

Would love to chat more! Personally, I don't think that the plasticity gating mechanism in our model is very important (which is why we tested both with and without the gating). Related models (predictive coding and Boltzmann machines) don't have this gate, so it may not be necessary for learning.

Predictive coding of reward in the hippocampus - Nature Calcium imaging of mouse hippocampal neurons while mice learn a reward-based task over several weeks provides insight into the evolution of the hippocampal reward representation during extended periods of experience.

Nature research paper: Predictive coding of reward in the hippocampus

go.nature.com/49mB13V

Our paper on the "Oneirogen hypothesis" is now up in its revised form on eLife!

This is the hypothesis that psychedelics induce a dream-like state, which we show via modelling could explain a variety of perceptual and learning effects from such drugs.

elifesciences.org/reviewed-pre...

🧠📈 🧪

Diagram of a recurrent neural network: input goes into the network, output is compared to a target to produce an error, and dotted feedback arrows show updates to neural activity and to synaptic weights.

1/7 How should feedback signals influence a network during learning? Should they first adjust synaptic weights, which then indirectly change neural activity (as in backprop.)? Or should they first adjust neural activity to guide synaptic updates (e.g., target prop.)? openreview.net/forum?id=xVI...

A theory of multi-task computation and task selection Neural activity during the performance of a stereotyped behavioral task is often described as low-dimensional, occupying only a limited region in the space of all firing-rate patterns. This region has...

1/X Excited to present this preprint on multi-tasking, with
@david-g-clark.bsky.social and Ashok Litwin-Kumar! Timely too, as “low-D manifold” has been trending again. (If you read thru the end, we escape Flatland and return to the glorious high-D world we deserve.) www.biorxiv.org/content/10.6...

Excited to share that our work ‘Simultaneous detection and estimation in olfactory sensing’ with @mattyizhenghe.bsky.social, @neurovenki.bsky.social , @cpehlevan.bsky.social, @jzv.bsky.social and @paulmasset.bsky.social has been launched!

1/7

I have open positions for graduate students in my lab. If you’re interested in joining, please apply through the Mila form.

I’m particularly interested in (thread below): 1/3

🧠🤖 #MLSky

Representational drift without synaptic plasticity Neural computations support stable behavior despite relying on many dynamically changing biological processes. One such process is representational drift (RD), in which neurons' responses change over ...

When neurons change, but behavior doesn’t: Excitability changes driving representational drift

New preprint of work with Christian Machens: www.biorxiv.org/content/10.1...

Great review! Section 4.2 in particular looks very consistent with our 'oneirogen hypothesis' model (elifesciences.org/reviewed-pre...). You've pointed us to lots of great references for our revisions :)

Top-down feedback is ubiquitous in the brain and computationally distinct, but rarely modeled in deep neural networks. What happens when a DNN has biologically-inspired top-down feedback? 🧠📈

Our new paper explores this: elifesciences.org/reviewed-pre...

Just a couple days until Cosyne - stop by [3-083] this Saturday and say hi! @nandahkrishna.bsky.social

Hi BlueSky fam, for my first post and to celebrate our recent paper being physically published I thought I’d do a summary thread!

This has been my most favourite (and toughest) work to date.

Please help share around!!

www.cell.com/cell/abstrac...
(Reach out if you can’t access)

How do interneurons reshape neural responses? I'm excited to present work with @eerosim.bsky.social at #NeurIPS2024 that proposes a nonlinear recurrent circuit model motivated by efficient coding theory.

Poster: 4:30p on Fri, Dec 13
Paper: openreview.net/forum?id=ojL...

Why does #compneuro need new learning methods? ANN models are usually trained with Gradient Descent (GD), which violates biological realities like Dale’s law and log-normal weights. Here we describe a superior learning algorithm for comp neuro: Exponentiated Gradients (EG)! 1/12 #neuroscience 🧪

11. We’re looking forward to your feedback, and we hope that some of you will be interested in testing aspects of our model!

10. To summarize: we’ve proposed a model of the hallucinatory effects of classical psychedelics, wherein hallucinations are caused by an increase in top-down inputs ordinarily reserved for offline generative replay underlying a form of representation learning in the brain.

9. We conclude our study by providing a variety of testable predictions that experimental neuroscientists could use to validate or invalidate our model.

8. Furthermore, consistent with the Entropic Brain Theory (Carhart-Harris et al. 2014), we show that simulated psychedelic administration in our model produces increases in stimulus-conditioned neural variability.

7. Next, we show that intermediate simulated psychedelic doses in our model cause large increases in synaptic plasticity at both apical and basal synapses, as seen in (Shao et al. 2021).

6. We next explored how our model could explain existing neural data. First, we show that Wake-Sleep learning induces strongly correlated tuning between apical and basal dendritic compartments, as has been observed experimentally (Beaulieu-Laroche et al. 2019; O’Hare et al. 2024).

5. Next, we point to experimental evidence that classical psychedelics may increase the influence of apical synapses on neural activity, while decreasing the influence of basal synapses. Simulating psychedelic doses this way in our models produces highly structured hallucinations.

We map the Wake-Sleep algorithm onto cortical architecture by proposing top-down, predictive synapses onto apical dendrites of pyramidal neurons dominate activity during the generative Sleep phase, while bottom-up synapses onto basal dendrites dominate during waking activity.

3. An ‘oneirogen’ is a compound that makes neural activity more dream-like. We propose that psychedelics shift neural activity towards a regime normally reserved for replay during sleep, used for learning. To model this form of learning, we use the classic Wake-Sleep algorithm (Hinton et al. 1995).

2. We put forward the ‘Oneirogen Hypothesis,’ arguing that psychedelics cause hallucinations because they hijack a system ordinarily used for learning in the brain.

The oneirogen hypothesis: modeling the hallucinatory effects of classical psychedelics in terms of replay-dependent plasticity mechanisms Classical psychedelics induce complex visual hallucinations in humans, generating percepts that are coherent at a low level, but which have surreal, dream-like qualities at a high level. While there a...

1. Hi all: I’m here to advertise our new preprint: www.biorxiv.org/content/10.1..., with Fabrice Normandin, @tyrellturing.bsky.social, and @glajoie.bsky.social!