Reading a neuron's tuning from its correlations with other neurons
summary
The gist
What a neuron represents is usually established by relating its activity to a stimulus The decoder trained on the correlation matrix of 512 sampled neurons…
In short
A decoder trained only on a correlation matrix of 512 neurons can determine preferred orientation from in-vivo mouse visual cortex responses, achieving a mean error of 25 degrees. The system recovers absolute orientation by exploiting subtle deviations from circulant symmetry in the class structure, proving that tuning information is encoded in relational patterns rather than stimulus labels or specific neuron samples.
Key concepts
- Correlation Matrix Input
- The decoder's input is strictly the correlation matrix of a sampled population, where each row represents a neuron. This allows the system to infer neuronal properties without needing any external labels or reference populations, focusing solely on internal relationships between neurons.
- Preferred Orientation Decoding
- The primary task is to decode the preferred orientation of visual cortex neurons. The system achieves this by analyzing how tuning profiles relate to each other within the population's correlation structure, even when no specific stimulus information is provided during training.
- Circulant Symmetry and Anisotropy
- The relations between orientation classes are nearly circulant, which fixes the class structure up to rotation and reflection. Recovering the absolute frame requires finding a departure from this symmetry—an anisotropy in the signal correlations—which serves as the key to fixing the absolute orientation.
- Receptive Field Cross-Animal Transfer
- The decoder can read receptive field positions across different animals because it focuses on where an animal's imaged patch looks on a screen, rather than the specific layout of neurons within that image. This allows for transfer of receptive field information based on spatial relationships.
Terminology used across episodes
This episode discusses
- Reading a neuron's tuning from its correlations with other neurons · Paper Radio
- Unambiguous Representations in Neural Networks: An Information-Theoretic Approach to Intentionality
The paper
Reading a neuron's tuning from its correlations with other neurons · Read on arXiv
What a neuron represents is usually established by relating its activity to a stimulus. We ask whether the same facts are contained in the correlation matrix of a recorded population, which keeps the pairwise relations among the neurons and discards the stimulus. A transformer that sees only the correlation matrix of 512 sampled neurons, with no labels and no reference population in its input, decodes the preferred orientation of mouse visual cortex neurons from in-vivo MICrONS responses to a mean error of 25° (uninformed: 45°; best constant guess: 41°) and their receptive-field position at R squared = 0.23; on the digital-twin substrate, 20° and 0.48. Asked whether a neuron prefers horizontal or vertical, it is right for 79% of neurons in vivo and 87% on the twin, against 56% for the majority answer; it does not resolve the obliques. The ordering is recoverable because the relations between orientation classes are close to circulant, which fixes the class structure up to rotation and reflection; the rotation of the frame is recoverable only because of a departure from that symmetry. The departure survives equal representation of every orientation and disjoint training and test stimuli, and it is an anisotropy of the signal correlations: in V1 the neurons preferring the dominant orientation have the most similar responses to one another, and a model population with only that property recovers the frame. Across 33 mice of the Allen Brain Observatory, with relations from drifting gratings, orientation transfers to unseen animals only when a sampled population mixes animals; populations drawn from one animal are not decoded measurably below the best constant, a limit of the diversity of populations one animal provides. For receptive fields, what crosses animals is where each animal's imaged patch looks on the screen, not the layout of neurons within it.
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Reading a neuron's tuning from its correlations with other neurons".
Marcus: Here's the gist of the paper. What a neuron represents is usually established by relating its activity to a stimulus The decoder trained on the correlation matrix of 512 sampled neurons, with no labels and no reference population in its input, decodes the preferred orientation of mouse visual cortex neurons from in-vivo MICrONS responses to a mean error of 25◦ The ordering is recoverable because the relations between orientation classes are close to circulant, which fixes the class structure up to rotation and reflection The core question addressed by this work The paper investigates whether the same facts about a neuron's tuning can be contained in the correlation matrix of a recorded population, discarding any stimulus information This relational setting is common, as recordings under natural behavior lack controlled stimuli and neurons from different sessions or animals share no stimulus-aligned coordinates but do share the structure of their correlations The decoder's input is strictly the correlation matrix of a sampled population, one row per neuron, and its output is either a preferred orientation or a receptive-field position Decoding Orientation and Receptive Fields The decoder reads orientation and receptive-field position from the Gram of 512 neurons of the test half on MICrONS, with predictions in the true frame About half the neurons are decoded to within 15◦ (0.46 in vivo, 0.54 on the twin) The decoder resolves cardinal axes and little else Orientation is scored by the mean absolute angular error between the decoded angle, atan2 of the output, and the label, taken modulo 180◦ Receptive fields are scored by R2 against a baseline of 0 Recovering the Absolute Frame The class-Gram of orientation is close to circulant, meaning it is unchanged by every rotation and reflection of the orientation circle, so that part of the relations fixes the classes only up to the 16 elements of D8 Whatever fixes the frame must be a departure from that symmetry, which is an anisotropy of the signal correlations A model population with only that property recovers the frame The absolute frame is recovered by a decoder trained with each population’s targets rotated at random, and the loss taken after the best alignment of predictions to targets Limits and Mechanisms The frame is not the label density and not the particular sample of frames The anisotropy factor in V1, where 28% of neurons fall in the 22.5◦ class centred on 90◦, peaks at 90◦ on both substrates (1.28 on the twin, 1.77 in vivo) The anisotropy is one of tuning similarity, not of trial-to-trial co-fluctuation Orientation is carried by the relations between tuning profiles, which exist between any two neurons that saw the same conditions, in the same brain or not Receptive Field Cross-Animal Transfer The receptive-field position crosses animals where each animal’s imaged patch looks on the screen, not the layout of neurons within it The decoder recovers where an animal’s imaged patch looks, not which neuron within it looks where Scaling and Population Size Beyond 256 to 512 neurons the orientation error moves by at most 2◦, up or down The twin receptive fields keep gaining slowly Population size beyond 256 to 512 neurons changes orientation error by at most 2◦ in either direction Controls and Validation The standard decoder, with the three row statistics as the only input per neuron and no relational term, yields an error of 37.7◦ The control that subtracts the whole class structure, which should destroy decoding, improved it further to 17.2◦ raw and 15.6◦ aligned The twin substrate is a model fitted to the same recordings, so the in-vivo results carry the claim about cortex and the twin results show what cleaner correlations would give The paper adds the measurement in cortex, at the level of single neurons, and the finding that the absolute readout is possible and rests on a remainder that belongs to the correlations, not to the sample of neurons or stimuli Code and Data Availability Code, run configurations and per-neuron predictions are in the directory microns ambiguity (branch microns-ambiguity) of https://github.
Title and authors: Ines: So we're looking at this paper by Lässig and the team titled "Reading a neuron's tuning from its correlations with other neurons". It’s about seeing what a single neuron is tuned to just by looking at how it talks to all the other neurons in its population, even when you don't give the decoder any labels or any stimulus information.
Marcus: That sounds like a pretty wild claim, Ines. Basically, they take this massive correlation matrix from five hundred twelve sampled neurons and try to figure out two things about that neuron: what direction it prefers—like horizontal or vertical—and where its receptive field is located on the screen.
Yuki: From a population genetics standpoint, that’s fascinating because it suggests we might be able to infer internal neural structure from just observing the statistical relationships within a group, without needing perfect experimental control over every single trial.
Ines: Exactly. The paper says this decoder can decode the preferred orientation of mouse visual cortex neurons from in-vivo MICrONS responses with a mean error of twenty five degrees when it has no labels or reference population in its input. They also say it can read the receptive field position at R squared of zero point two three on their digital twin substrate and zero point four eight.
Marcus: So, the key improvement they point out is how they recover the absolute frame of orientation by looking at a departure from circulant symmetry in the class relations, which they call an anisotropy of the signal correlations. That means it doesn't just know if something is horizontal or vertical, but it figures out where "up" or "left" actually is in space.
Yuki: That idea about the anisotropy being a property of tuning similarity rather than trial-to-trial co-fluctuation really connects to how we think about neural coding across different samples. If the orientation information is carried by those relations between tuning profiles, it should hold up even if you're looking at neurons from different sessions or even different animals.
Ines: That's a huge implication for understanding how these fundamental visual properties are encoded in the brain, suggesting that the structure of these correlations themselves contain the spatial information we want to extract. But they also mention some limits, like how beyond two hundred fifty to five hundred twelve neurons, orientation error moves by at most two degrees up or down.
Marcus: They also note that for orientation, a model population with only that anisotropy property recovers the frame because the anisotropy factor is one of tuning similarity, not of trial-to-trial co-fluctuation. That means if you can find those structural similarities in the correlation matrix, you can recover the absolute frame.
Yuki: And they also say that for receptive field position, what crosses animals is where each animal’s imaged patch looks on the screen rather than the layout of neurons within it, which is a really interesting way to think about cross-species transfer of this positional data.
Title and authors: Ines: So, we're not just getting a single number like orientation; we're getting a structured understanding of the neuron's tuning and position based on its relational structure alone. The paper shows that this method can actually improve upon standard decoders, which get an error of thirty seven point seven degrees when they only use the three row statistics per neuron.
Marcus: And when they subtract the whole class structure, which should destroy decoding, the error improves to seventeen point two raw and fifteen point six aligned. That shows that isolating that specific structural information is what makes it work better than just looking at individual neuron statistics.
Yuki: What this means for us is that we can start to infer things about neural organization from correlation matrices in biological data, which is a big step because it bypasses the need for labeled neurons or controlled stimulus presentation during the initial decoding stage.
Ines: It definitely opens up a new way to study these systems, moving from stimulus-based mapping to purely relational analysis. We should probably look closer at how this relates to other works we've seen on population activity patterns in the background material.
Marcus: Right, so the next thing is how they handle scaling and what it says about population size. They show that beyond five hundred twelve neurons, orientation error moves by at most two degrees in either direction, and for receptive fields, the error also changes by at most two degrees when you go past two hundred fifty to five hundred twelve neurons.
Yuki: So the power of this method seems tied to having a sufficiently large and diverse population sample that still retains enough structural information about those orientation classes. It's not just about adding more neurons; it's about the structure they form.
Ines: And they do mention that the ordering is recoverable because those relations between orientation classes are close to circulant, which fixes the class structure up to rotation and reflection. That’s a neat mathematical property of how these tuning profiles are related in space.
Marcus: So, to bring it all together, this paper on "Reading a neuron's tuning from its correlations with other neurons" shows that you can decode orientation and position from correlation matrices without labels by looking at the structure of those relations, specifically exploiting the anisotropy in the signal correlations to find the absolute frame.
Yuki: It suggests that we might be able to infer fundamental properties of neural tuning just by analyzing how neurons interact statistically, which is a pretty big shift in how we approach population dynamics.
Ines: It's definitely worth considering this as a way to bypass some of the traditional hurdles in studying these complex systems where controlled stimulus presentation isn't always possible.
Marcus: Yeah, I think the next step is seeing if we can apply this framework to other types of neural data or even different biological systems where those structural symmetries exist.
The paper's summary: Ines: So we're looking at what the researchers actually found about these correlation matrices of neurons today. It turns out they can figure out two major things about a neuron from just looking at how it correlates with all the others in its group, even when you don't have any labels or any specific stimulus data in the input.
Marcus: Right, so they took this big set of five hundred twelve neurons and tried to decode their preferred orientation and where their receptive field is on a screen. The main thing is that it works surprisingly well with just those statistical relationships, even when there's no prior knowledge about what the neuron is tuned to.
Yuki: From a population genetics angle, that’s interesting because it suggests we can infer internal neural structure from just observing the statistical relationships within a group without needing perfect experimental control over every single trial. It implies that these fundamental visual properties are encoded in the structure of those correlations themselves.
Ines: Exactly. They found this decoder can decode the preferred orientation with a mean error of twenty five degrees when it has no labels or reference populations in its input. And for receptive field position, they get a score of zero point two three on their digital twin substrate and zero point four eight.
Marcus: The real kicker is how they recover the absolute frame of orientation. They do this by looking for a departure from circulant symmetry in the class relations, which they call an anisotropy of the signal correlations. That means it doesn't just know if something is horizontal or vertical, but it figures out where "up" or "left" actually is in space.
Yuki: That idea about the anisotropy being a property of tuning similarity rather than trial-to-trial co-fluctuation really connects to how we think about neural coding across different samples. If the orientation information is carried by those relations between tuning profiles, it should hold up even if you're looking at neurons from different sessions or even different animals.
Ines: But they also have some limits. They point out that beyond two hundred fifty to five hundred twelve neurons, the orientation error moves by at most two degrees up or down. It’s not infinitely scalable like we might hope for with more data points.
Marcus: And they mention that for receptive field position, the error also changes by at most two degrees when you go past two hundred fifty to five hundred twelve neurons. So, population size beyond this range doesn't just give you perfect decoding anymore; it starts introducing some noise in the measurement record.
Yuki: It’s a bit like how population structure eventually hits a saturation point in evolution, where adding more individuals doesn't change the fundamental patterns of relatedness you can see.
Ines: What this means for us is that we can start to infer things about neural organization from correlation matrices in biological data without needing labeled neurons or controlled stimulus presentation during the initial decoding stage. It opens up a new way to study these systems, moving from stimulus-based mapping to purely relational analysis.
Marcus: Yeah, it's definitely a big step toward figuring out how these complex systems organize themselves naturally, bypassing some of the traditional hurdles we face when we try to map neurons based on external input. But that structure they find—the anisotropy—that’s what really sets this method apart for finding that absolute frame.
Yuki: So, the next thing to consider is how this relational property transfers across different animals or even different substrates, because if those tuning profiles are truly universal, the implications for understanding species-specific visual processing become huge.
The paper's improvements: Tom: So we're looking at how these researchers actually improved their method for decoding neuron properties from correlation matrices today. They found that they can get much better results by adding a relational term to the decoder’s input, rather than just using the basic row statistics.
Ines: Right, so instead of just feeding the decoder three row statistics per neuron without any relational term, which gave them an error of thirty seven point seven degrees, they added that whole class structure information. That improved it down to seventeen point two raw and fifteen point six aligned.
Marcus: That subtraction process is key because it shows that isolating that specific structural information is what makes the method work better than just looking at individual neuron statistics in isolation. It suggests the relationship between neurons is more important than any single neuron's statistic alone.
Yuki: And from a population genetics perspective, this implies that you need to look at how the whole group relates to each other, not just treating each neuron as an independent data point in your analysis. The sample structure matters a lot here for getting that better accuracy.
Ines: They also showed that they can achieve high accuracy in decoding orientation by training on populations drawn with equal numbers of neurons from each of the eight orientation classes. That gave them a result of twenty two point zero degrees on the digital twin and twenty six point six degrees in vivo when using two seeds.
Marcus: So, having a balanced representation across those eight classes seems to be a necessary condition for getting that level of accuracy on both platforms. It’s not just about having more neurons, it's about making sure the population is structurally representative enough.
Yuki: It supports the idea that the underlying structure of the tuning profiles dictates what you can recover in terms of orientation, which is a strong statement for how we model visual processing across different biological contexts.
Ines: They also confirmed that they can distinguish between orientation and receptive field readout by observing where each animal’s imaged patch looks on the screen rather than the layout of neurons within it. That really helps separate what you're decoding from which specific neuron is doing what at a given moment.
Marcus: That distinction is important because it tells us that for receptive field position, you can use correlations with "each animal’s imaged patch looks on the screen" instead of focusing on how the neurons are laid out inside that image.
Yuki: It reinforces the idea that the positional information is tied to a more fundamental spatial relationship between different viewing frames across animals.
Ines: So, what this changes for someone listening is that you don't just get one number for orientation; you get a structured understanding of the neuron's tuning and position based on its relational structure. It moves us away from stimulus-based mapping toward purely relational analysis.
Marcus: Yeah, it’s a big step toward figuring out how these complex systems organize themselves naturally, bypassing some of the traditional hurdles we face when we try to map neurons based on external input. But that structure they find—the anisotropy—that’s what really sets this method apart for finding that absolute frame.
Yuki: So, the next thing to consider is how this relational property transfers across different animals or even different substrates, because if those tuning profiles are truly universal, the implications for understanding species-specific visual processing become huge.
Conclusion: Ines: So we're wrapping up on "Reading a neuron's tuning from its correlations with other neurons." The paper shows that you can decode orientation and receptive field position from just looking at the structure of how neurons talk to each other, even without any labels.
Marcus: Exactly. It’s about using the statistical relationships between neurons as a proxy for their underlying biological tuning, which is a pretty cool way to bypass needing perfect stimulus control in these recordings.
Yuki: From a population perspective, it means that the structure of those correlations carries information about how those populations evolved their visual processing capabilities across different conditions. It supports the idea that species have certain inherent patterns in how they organize their sensory systems.
Ines: They also showed that this method can actually improve upon standard decoders, which get an error of thirty seven point seven degrees when they only use the three row statistics per neuron. That improvement comes from incorporating that whole class structure information we talked about earlier.
Marcus: And for the receptive field position, it’s not just a number; it’s a spatial relationship. They found that what matters is where each animal’s imaged patch looks on the screen, not which specific neuron inside that image is looking where.
Yuki: That really frames the receptive field as something dependent on the viewing context of the whole animal, rather than just an internal coordinate of a single neuron in isolation.
Ines: The method works best when you have a population with equal numbers of neurons from each orientation class, which gives them twenty two point zero degrees on their digital twin and twenty six point six degrees in vivo when using two seeds. That suggests structural balance is key for high accuracy.
Marcus: It’s a strong result, but we gotta remember the caveats they mentioned earlier: beyond five hundred twelve neurons, the orientation error starts moving by at most two degrees up or down. So it doesn't work perfectly if you just keep adding more data points past that threshold.
Yuki: That limitation makes me think about how much structural information a sample really can support before the noise from random fluctuations starts overwhelming the signal. It’s like how population samples eventually lose their ability to reveal deeper evolutionary patterns.
Ines: But overall, what this paper does is establish that we can infer fundamental properties of neural tuning just by analyzing how neurons interact statistically, which is a big step toward figuring out how these complex systems organize themselves naturally.
Marcus: It’s a solid piece of work for showing that you don't need perfect labeling to start decoding the structural features of neural populations. It opens up a new way to study these systems by focusing on their internal organization rather than just external input.
Yuki: So, while this paper focuses on orientation and receptive fields in mouse visual cortex, the idea that tuning profiles can be inferred from correlation matrices might apply to other sensory systems across different species too. It’s a powerful way to look at the commonalities in neural architecture.
More episodes
- 2610.33881-Sensorimotor behavior from an untrained brain-and-nerve-cord connectome
- 2610.40344-Morphological Features of Murine Pulmonary Arterial Networks for Diagnosing Vascular Disease
- Daily Summary for 2026-11-19
- 2609.10193-Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
- 2610.38236-Generative discovery of chemical reaction networks
- 2507.11176-An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2504.03590-Towards a Unified Framework for Determining Conformational Ensembles of Disordered Proteins
- 2603.15217-A multiscale discrete-to-continuum framework for structured population models