Lightnews — Scholar-powered news

Hugo Ninou @hugoninou.bsky.social · 1d

Huge thanks to my amazing supervisors @kadmonj.bsky.social and Alex Cayco-Gajic !

Read the full story : www.arxiv.org/abs/2510.02765

Curl Descent: Non-Gradient Learning Dynamics with Sign-Diverse Plasticity

Gradient-based algorithms are a cornerstone of artificial neural network training, yet it remains unclear whether biological neural networks use similar gradient-based strategies during learning. Expe...

www.arxiv.org

Hugo Ninou @hugoninou.bsky.social · 1d

9/10
Our work challenges the dominant view that learning must strictly follow a gradient. The diversity of plasticity rules in biology might not be a bug, but a feature—an evolutionary strategy leveraging non-gradient dynamics for more efficient and robust learning. 🌀

1 3

Hugo Ninou @hugoninou.bsky.social · 1d

8/10
In this latter regime where only one rule-flipped synapse was introduced in the readout layer, curl descent can speed up learning in nonlinear networks. This result holds over a wide range of hyper-parameters.

1 1

Hugo Ninou @hugoninou.bsky.social · 1d

7/10
The story is completely different for the readout layer. Surprisingly, even when curl terms make the solution manifold unstable, the network is still able to find other low-error regions!

1

Hugo Ninou @hugoninou.bsky.social · 1d

6/10
But beware! If you add too many rule-flipped neurons in the hidden layer of a compressive network, the learning dynamics spiral into chaos, thereby destroying performance. The weights never settle, and the error stays high.

1

Hugo Ninou @hugoninou.bsky.social · 1d

5/10
How does this scale up? We used random matrix theory to find that stability depends critically on network architecture. Expansive networks (input layer > hidden layer) are much more robust to curl terms, maintaining stable solutions even with many rule-flipped neurons.

1

Hugo Ninou @hugoninou.bsky.social · 1d

4/10
Toy example : In a tiny 2-synapse network, curl descent can escape saddle points and converge faster than gradient descent by temporarily ascending the loss function. But it comes at a cost: half of the optimal solutions become unstable.

1 1

Hugo Ninou @hugoninou.bsky.social · 1d

3/10
But can networks with such non-gradient learning dynamics still support meaningful optimization? We answer this question by focusing on an analytically tractable teacher-student framework, with 2-layer feedforward linear networks.

1

Hugo Ninou @hugoninou.bsky.social · 1d

2/10
This is motivated by the diversity observed in the brain. A given weight update signal can have an opposite effects on a network's computation depending on the postsynaptic neuron (e.g. E/I), which is inconsistent with standard gradient descent.

1

Hugo Ninou @hugoninou.bsky.social · 1d

1/10
We define the curl descent learning rule by flipping the sign of the updates given by gradient descent for some weights. These weights are chosen at the start of learning depending on the nature (rule-flipped or not) of the presynaptic neuron.

1

Hugo Ninou @hugoninou.bsky.social · 1d

🚨New spotlight paper at Neurips 2025🚨

We show that in sign-diverse networks, inherent non-gradient “curl” terms arise, and can, depending on network architecture, destabilize gradient-descent solutions or paradoxically accelerate learning beyond pure gradient flow.

🧵⬇️

www.arxiv.org/abs/2510.02765

Curl Descent: Non-Gradient Learning Dynamics with Sign-Diverse Plasticity

Gradient-based algorithms are a cornerstone of artificial neural network training, yet it remains unclear whether biological neural networks use similar gradient-based strategies during learning. Expe...

www.arxiv.org

1 2

Hugo Ninou @hugoninou.bsky.social · Jun 13

Unable to access, but would love to read this ! Any full text link 👀?

1

Reposted by Hugo Ninou

Leo Kozachkov @leokoz8.bsky.social · May 28

Big week for astrocyte research: 3 new Science papers link astrocytes to behavior. We're excited to add to the momentum with our new PNAS paper: a theory, grounded in biology, proposing astrocytes as key players in memory storage and recall. w/ JJ Slotine and @krotov.bsky.social
(1/6)

1 13 34

Reposted by Hugo Ninou

Friedemann Zenke @fzenke.bsky.social · May 27

1/6 Why does the brain maintain such precise excitatory-inhibitory balance?
Our new preprint explores a provocative idea: Small, targeted deviations from this balance may serve a purpose: to encode local error signals for learning.
www.biorxiv.org/content/10.1...
led by @jrbch.bsky.social

4 56 170

Reposted by Hugo Ninou

Jonathan Kadmon @kadmonj.bsky.social · Mar 23

(1/6) Excited to share a new preprint from our lab! Can large, deep nonlinear neural networks trained with indirect, low-dimensional error signals compete with full-fledged backpropagation? Tl;dr: Yes! arxiv.org/abs/2502.20580.

Training Large Neural Networks With Low-Dimensional Error Feedback

Training deep neural networks typically relies on backpropagating high dimensional error signals a computationally intensive process with little evidence supporting its implementation in the brain. Ho...

arxiv.org

5 22 63

Hugo Ninou @hugoninou.bsky.social · Mar 22

9/9
A huge thank you to my co-first author @SharonIsraely and @OriRajchert, @LeeElmaleh, @RanHarel, @FirasMawase,
@kadmonj.bsky.social , and @yifatprut.bsky.social ut.bsky.social for their invaluable contributions and support throughout this journey. 🙏

Bluesky

ut.bsky.social

1 1 1

Hugo Ninou @hugoninou.bsky.social · Mar 22

8/n
1️⃣ Our study provides new insights into how cerebellar signals constrain cortical preparatory activity, promoting generalization and adaptation.
2️⃣ We demonstrate that in the absence of cerebellar signals, cortical mechanisms are harnessed to restore adaptation, albeit with reduced efficiency.

1 2

Hugo Ninou @hugoninou.bsky.social · Mar 22

7/n
⚫ The increased dimensionality under HFS was accompanied by a decrease in generalization performance, both at the neural and behavioral levels.

e Cross condition generalization performance relative to chance level (0.5) calculated for Control (blue) or HFS (red) conditions. Each dot represents a pair of dichotomies with shared symmetries (e.g., top-down dichotomy on left targets and top-down dichotomy on right targets). Diamonds represent the average CCGP over all dichotomies. f Quantification of generalization across all sessions, calculated for early (1–5) and late (>= 10) trials in the FF (blue) and HFS (red) conditions (n samples 4 and 35 for early and late FF trials respectively, and 4 and 30 samples for early and late FF- HFS trials. Data are presented as mean values ± SEM.

1 1

Hugo Ninou @hugoninou.bsky.social · Mar 22

6/n
⚫ HFS led to higher dimensionality in neural activity, indicating a loss of structure in the neural representations, which is crucial for efficient motor learning and adaptation.

a Dimensionality of neural activity estimated by the participation ratio (see Methods) at different epochs (blue bars: conrol, red: HFS). Asterisks denote signifcant differences in dimensionality (Dpca) during control vs. HFS conditions (resampling test with n = 1000, p < 0.01) d Illustration depicting the effects of the topology of the neural representation.

1 3

Hugo Ninou @hugoninou.bsky.social · Mar 22

5/n
⚫ This compensation involved an angular shift in neural activity, suggesting a "re-aiming" strategy to handle the force field in the absence of cerebellar control.

e Polar histograms of the neural angles calculated for one monkey (monkey S) by aggregating data from all cued targets and trials in the control FF conditions with the same forcefield direction (left: clockwise CW trials, right: counter-clockwise CCW trials). f Same as (e) but during FF combined with HFS (FF-HFS).

1

Hugo Ninou @hugoninou.bsky.social · Mar 22

4/n
⚫ Under high-frequency stimulation (HFS), we observed a bigger difference between FF and null field (NF) neural activity, indicating a compensatory mechanism in the motor cortex to adapt to the perturbation.

c Decoding accuracy of adaptation conditions (NF vs. FF) based on epoch-specifc data for the control (blue bars) or HFS (red bars). Dashed line denotes chance level (0.5). Decoding accuracy values were obtained for different training and testing sets.

1

Hugo Ninou @hugoninou.bsky.social · Mar 22

3/n
⚫ Under high-frequency stimulation (HFS), neural activity was altered in both a target-dependent and independent manner, showing that cerebellar signals contain task-related information.

f, g Effects of HFS on coordinated neural activity, calculated to a data set with reach to 8 targets for both control and HFS conditions (see Methods). PCA was performed in a similar manner as in (d, e), concatenating control (solid) and HFS (dashed) data (explained variance: PC1:0.2; PC2: 0.16; PC3: 0.1; PC4: 0.07; PC5: 0.07). f PC1 and PC2 g PC1 and PC3. h PC1 and PC5. PC4 did now show HFS dependent difference.

1 1

Hugo Ninou @hugoninou.bsky.social · Mar 22

2/n 🔍 Key Findings:

⚫ Cerebellar Block Impairs Adaptation: Blocking cerebellar outflow thanks to high-frequency stimulations (HFS) in the superior cerebellar peduncle significantly impairs force field (FF) adaptation, leading to increased motor noise and reduced error sensitivity.

c Single-session motor noise was estimated by calculating the mean absolute deviation (MAD) of maximal deviations during HFS trials and plotting against the motor noise calculated during the matching control sessions (n = 191). Darker dots (n = 91) indicate sessions where the motor noise under HFS was comparable to the motor noise level during the control trials (i.e., HFS/Control ratio >0.6 and <1.4). d Mean error sensitivity ±SEM calculated for a subset of adaptation sessions (n = 91), for which the pre-adaptation noise level was comparable during FF (blue) and FF-HFS (red) conditions (i.e., highlighted sessions in c).

1 1

Hugo Ninou @hugoninou.bsky.social · Mar 22

🚨 Paper Alert! 🚨
1/n Thrilled to share our latest research, now published in Nature Communications! 🎉 This study dives deep into how the cerebellum shapes cortical preparatory activity during motor adaptation.
www.nature.com/articles/s41...
#neuroskyence #motorcontrol #cerebellum #motoradaptation

Cerebellar output shapes cortical preparatory activity during motor adaptation - Nature Communications

Functional role of the cerebellum in motor adaptation is not fully understood. The authors show that cerebellar signals act as low-dimensional feedback which constrains the structure of the preparator...

www.nature.com

4 16 50

Reposted by Hugo Ninou

Joao Barbosa @jbarbosa.org · Dec 16

Created a starter pack of neuroscience in/from Paris.

Let me know if you want to be added (the 'from' can include those not in Paris anymore) or just tap in if you want to know what we're talking about!

Regardless, please re-tweet!

go.bsky.app/3Zs9w5w

28 39 62