Lightnews — Scholar-powered news

neurosock @neurosock.bsky.social · 2d

Thanks for your words

neurosock @neurosock.bsky.social · 2d

Disclosure: this is a simplification from a very complex paper!

Sorry if I omitted some details, it is all for clarity 😸

2

neurosock @neurosock.bsky.social · 2d

Limitations:

The precise biological mechanism for how the error is computed and broadcasted back also remains an open question.

They used back propagation (BP) to train the model but whether BP happens in the brain seems implausible to me.

2

neurosock @neurosock.bsky.social · 2d

Implications for AI:

This work provides ideas for building more robust and data-efficient AI.

Using architectural delays and self-supervised prediction could help create AI that learns the world's structure more like a human child does, without massive labeled datasets.

1

neurosock @neurosock.bsky.social · 2d

They also showed that learning works even with random, unstructured feedback connections from L5 to L2/3 (Fig. 4a).

This is a crucial finding, as it makes the whole scheme far more biologically plausible and robust to messy wiring.

1

neurosock @neurosock.bsky.social · 2d

Most impressively, the model reproduced the exact, layer-specific "surprise" signals seen in the brains of mice when their expectations are violated (Fig. 7c, d).

It showed a positive error in L2/3 and a negative one in L5.

1

neurosock @neurosock.bsky.social · 2d

The model also replicated the known fact that superficial cortical layers are more sparsely active than deep layers (Fig. 6a).

They show this is a natural consequence of L2/3's predictive role, as it learns to only encode essential features.

1

neurosock @neurosock.bsky.social · 2d

The model spontaneously learned to "see" through noise and fill in missing parts of images (Fig. 5a, b).

This robustness is an emergent property of the predictive learning process, not something it was explicitly trained to do.

1

neurosock @neurosock.bsky.social · 2d

This is what we learned from their results:

The model's learning rule, derived from math, perfectly matched the rules of synaptic plasticity measured in real brain tissue (Fig. 2b).

This suggests the model is grounded in real biological mechanisms.

1

neurosock @neurosock.bsky.social · 2d

But why?

For decades, anatomical data didn't fit our functional theories of the cortex.

This work was needed to provide a new computational theory that could explain *why* the cortex is wired with these strange, parallel pathways.

1

neurosock @neurosock.bsky.social · 2d

Here is another way to put it:

L2/3 is a "prophet" that uses the past to predict the future.

L5 is a "judge" that sees present reality, and its verdict on the prophet's accuracy is what drives all learning.

The connection between L2/3 and L5 is altered when learning.

1 1

neurosock @neurosock.bsky.social · 2d

After training, the circuit sees the same sequence and now makes the right prediction.

The prediction from Layer 2/3 matches the reality arriving at Layer 5, the error signal is minimal, and the model has learned the rule.

1

neurosock @neurosock.bsky.social · 2d

This error is the engine of learning, acting as a teaching signal that is fed back to the circuit.

It guides plasticity, strengthening the connections that would have led to the correct prediction while weakening the wrong ones (teaching signal).

1

neurosock @neurosock.bsky.social · 2d

Simultaneously, the actual new input arrives at Layer 5 via the direct "ground truth" pathway.

L5 compares the incorrect prediction with this reality, generating a "surprise" or a prediction error signal when they do not match.

1

neurosock @neurosock.bsky.social · 2d

At time 't', the predictive Layer 2/3 combines the past memory from Layer 4 with a top-down context cue.

With untrained connections, it makes a poor guess (horizontal bar) about what comes next and sends this faulty prediction to Layer 5.

1

neurosock @neurosock.bsky.social · 2d

At time 't-1', the first sensory input (the horizontal bar) arrives through thalamus and is stored in Layer 4.

This layer acts as a short-term memory of what just happened a moment ago, setting the stage for a prediction.

1

neurosock @neurosock.bsky.social · 2d

Let's unpack this:

In this example an mouse is learning the meaning of a green up arrow symbol 🟩⬆️ and how it predicts the change in a red bar stimulus 🟥⛔.

The mouse should learn up arrow ⬆️ at time 't' indicates the bar will rotate counterclockwise at 't+1'.

1

neurosock @neurosock.bsky.social · 2d

Core result:

The authors propose this new lane and layered architecture is a machine for self-supervised learning.

It uses the slow pathway as a memory of the past to predict the present, and the fast pathway to check if the prediction was right.

1 1

neurosock @neurosock.bsky.social · 2d

The classic view was a simple sequential flow of information from layer 4, to 2/3, and then 5.

But new evidence shows a direct "express lane" of sensory input that goes straight to the output layer 5, what is it there for?

1

neurosock @neurosock.bsky.social · 2d

A core function of cortex is predicting what happens next given the world's state.

This recent paper from Oxford shows how cortical layers may use a delay trick to learn to predict.

A simple illustration can explain the idea.

A🧵with my toy model and notes:

#neuroskyence #compneuro #NeuroAI

1 5 17

neurosock @neurosock.bsky.social · 3d

If you like this, and think I should keep doing these posts every week, hit a like on the first post!

It motivates me to spend 5-6 hours a week preparing these 👇🏽😇

neurosock @neurosock.bsky.social · 3d

Disclosure: this is a simplification from a very complex paper!

Sorry if I omitted some details, it is all for clarity 😸

1

neurosock @neurosock.bsky.social · 3d

Limitations:

The precise biological mechanism for how the error is computed and broadcasted back also remains an open question.

They used back propagation (BP) to train the model but whether BP happens in the brain seems implausible to me.

1

neurosock @neurosock.bsky.social · 3d

Implications for AI:

This work provides ideas for building more robust and data-efficient AI.

Using architectural delays and self-supervised prediction could help create AI that learns the world's structure more like a human child does, without massive labeled datasets.

1

neurosock @neurosock.bsky.social · 3d

They also showed that learning works even with random, unstructured feedback connections from L5 to L2/3 (Fig. 4a).

This is a crucial finding, as it makes the whole scheme far more biologically plausible and robust to messy wiring.

1