More than 450 researchers have signed this declaration in the past 24 hours. While we continue to focus most of our efforts within Israel, there is, and always has been, significant resistance in the country. The academic community is one of many voices expressing this sentiment.

(6/6) Our findings challenge some prevailing conceptions about gradient-based learning, opening new avenues for understanding efficient neural learning in both artificial systems and the brain.
Read our full preprint here: arxiv.org/abs/2502.20580 #Neuroscience #MachineLearning #DeepLearning

(5/6) Applying our method to a simple ventral visual stream model replicates the results of Lindsey, @suryaganguli.bsky.social and @stphtphsn.bsky.social and shows that the bottleneck in the error signal—not the feedforward pass—shapes the receptive fields and representations of the lower layers.

(4/6) The key insight: The dimensionality of the error signal doesn’t need to scale with network size—only with task complexity. Low-dimensional feedback can effectively guide learning even in very large nonlinear networks. The trick is to align feedback weights to the error (the gradient loss).

(3/6) We developed a local learning rule leveraging low-dimensional error feedback, decoupling forward and backward passes. Surprisingly, performance matches traditional backpropagation across deep nonlinear networks—including convolutional nets and even transformers!

(2/6) Backpropagation, the gold standard for training neural networks, relies on precise but high-dimensional error signals. Other proposed alternatives, like Feedback alignment do not challenge this assumption. Could simpler, low-dimensional signals achieve similar results?

Just realized I've mentioned the wrong Jan above. Sorry @japhba.bsky.social!

Just as the meaning of words depends on context, the brain must infer context and meaning simultaneously to adapt in real-time. We believe these insights uncover core principles of #CognitiveFlexibility—check it out! www.biorxiv.org/content/10.1...
Congrats to John and Jan on their outstanding work!

Remarkably, DeepRL networks converged on near-optimal strategies and exhibited the same nontrivial Bayesian-like belief-updating dynamics—despite never being trained on these computations directly. This suggests that inference mechanisms can emerge naturally through reinforcement learning.

By deriving a Bayes-optimal policy, we show that rapid context shifts emerge from sequential belief-state updates—driven by flexible internal models. The resulting dynamics resemble an integrator with nontrivial context-dependent, adaptive bounds.

We tackled this challenge with behavioral experiments in mice, Bayesian theory, and #DeepRL. Using a novel change-detection task, we show how mice and networks adapt on the first trial from a context change by inferring both context and meaning—without trial and error.