Amit Yaron

Neuroscientist
Project researcher @ IRCN, Tokyo university

[Amit Yaron @YaronAmit]

@WashingtonExp cC1)

[Amit Yaron @YaronAmit]

Huge thanks to mentors Zenas Chao & Hirokazu Takahashi, and collaborators Tomoyo Shiramatsu-Isoguchi, Felix Kern & Kenichi Ohki!
@UTokyo_News WPI-IRCN
📄 Paper: https://t.co/xmYZEgO7EL

[Amit Yaron @YaronAmit]

A key asymmetry reveals the mechanism:
📌 Receptive field: responds to both tones when heard
📌 Predictive field: fires only when one specific tone is omitted
Same neuron — opposite selectivity for presence vs absence.

[Amit Yaron @YaronAmit]

Our circuit model explains these results. Lateral interactions between prediction-error neurons cause unexpected tones to suppress neighboring predictions, sharpening each neuron's predictive field and making omission coding more precise and efficient.

[Amit Yaron @YaronAmit]

This confidence built up over time:
🔹 Early in the sequence, gaps triggered little or no response
🔹 After repeated exposure to the target tone, the same gap evoked strong firing
This suggests the brain gradually accumulates statistical regularities to form stable expectations.

[Amit Yaron @YaronAmit]

New @PLOSBiology paper:
https://t.co/xmYZEgO7EL
We found neurons in rat auditory cortex that signal narrow predictions for specific tones.
Using omissions as a probe, we show that some neurons show frequency-specific expectations despite responding broadly to actual sounds. 🧵

[Amit Yaron @YaronAmit]

Omission responses tracked the target tone’s probability:
🔹 Gap in a sequence where the target tone was 90 % of items → strong firing
🔹 Gap in a sequence where the target tone was 10 % → weak or no firing
This shows predictions are feature-specific and confidence-graded.

[Amit Yaron @YaronAmit]

These are “Probability-Encoding Omission Neurons” (PEONs)
Their omission responses act as prediction error signals:
The stronger the expectation for this specific tone, the stronger the firing when that tone fails to appear.

[Amit Yaron @YaronAmit]

Two-tone sequences were played with occasional silent gaps (5% omissions). Many of these neurons responded to both tones when present. ~13% also fired on gaps, but only when they occurred in a sequence dominated by one specific tone.

[Amit Yaron @YaronAmit]

Looking forward to connecting with others thinking about:
– minimal systems
– spontaneous computation
– embodied vs disembodied learning
– how prediction might emerge from nothing at all.

[Amit Yaron @YaronAmit]

Our new paper just out:
“Dissociated neuronal cultures as model systems for self-organized prediction”
With Zhang, Akita, Shiramatsu, Chao & Takahashi
Published in Frontiers in Neural Circuits
🔗 https://t.co/GAdz6lqsRb

[Amit Yaron @YaronAmit]

Sometimes, a Petri dish is enough to explore the foundations of intelligence.

The capacity for prediction may not be something added later.
It may be what the cortex already is.

[Amit Yaron @YaronAmit]

@techdevnotes To be fair, you have to have a very high IQ to understand Grok 4. The reasoning capabilities are extremely subtle, and without a solid grasp of first-principles thinking most of the outputs will go over a typical user's head.

[Amit Yaron @YaronAmit]

@LiamCatterson94

[Amit Yaron @YaronAmit]

@strowhiro

[Amit Yaron @YaronAmit]

@616LEBEAUS