Daily Summary for 2026-11-09
daily
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: It's the ninth of November, twenty twenty-six, and this is the day's research.
Marcus: 18 new papers came out today.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We'll take the day in one pass, then pull out the papers we're staying with.
Marcus: Start us off with Self-Supervised Learning of Local Predictive Directions as a Neural Computation.
The day's research: Ines: Today we look at neural systems discovering local geometry directly from experience.
Marcus: Natural sensory dynamics often show simple local patterns despite being globally nonlinear.
Yuki: Zenari et al introduces Local Predictive Directions, a self-supervised framework that learns local directions by predicting near-future trajectory evolution.
Ines: This means the system learns the underlying shape of movement by looking at how trajectories evolve over time.
Marcus: For natural sensory streams, this process yields receptive fields resembling those found in early visual neurons.
Yuki: Neuro-Dist distillation shows transferring low tangling state from cortical dynamics improves artificial controllers' robustness to sensory feedback noise.
Ines: This suggests learning geometric features is a powerful mechanism for improving embodied AI control.
Marcus: Understanding how information geometry structures neural populations is relevant here.
Yuki: A multi-scale information geometry reveals the structure of mutual information in neural populations, showing how distances contract when stimulus resolution is lost.
Ines: This leads to a multi-scale extension of the Fisher information metric that captures encoding structure from fine details to coarse distinctions.
Marcus: BrainPFN addresses traditional kernel ridge regression limitations by meta-training on a large prior dataset to predict behavior from functional MRI.
Yuki: This method can outperform standard methods even with small cohorts.
Ines: The most significant development concerns using deep generative modeling at the agent level in discrete-event simulations.
Marcus: This fundamentally changes how we run computationally expensive agent-based simulations by replacing complex modeling steps with trained generative models.
Yuki: Leveraging deep generative modelling at the agent level in discrete-event simulations drastically reduces run time while preserving full simulation framework and dynamics.
Ines: This approach is crucial because it moves us closer to enabling full-model emulators for very intensive simulations.
Lucky paper: 2611.07809: Ines: So we’re looking at Self-Supervised Learning of Local Predictive Directions as a Neural Computation, and the core idea is that neural circuits can discover local geometry just by looking at how trajectories evolve over time.
Marcus: The paper introduces Local Predictive Directions, or LPD, which learns local directions by predicting near-future trajectory evolution using past experience.
Yuki: It uses past-future canonical correlation analysis on temporal differences to find these directions and it recovers local predictive geometry in stochastic nonlinear systems.
Ines: That means for natural sensory streams, it produces center–surround spatial receptive fields and ON/OFF filters that look like early visual neurons.
Marcus: The methodology involves a non-negative similarity-matching formulation where the objective function is min Z greater than or equal to zero times G transpose minus Z times Z times F.
Yuki: They solve that objective via gradient descent, alternating between an inference step for Z and updating the weights W and M from empirical covariances.
Ines: The paper states that they validated this on stochastic nonlinear systems with known local structure, like affine saddle-point dynamics, where it correctly identifies unstable directions even when the fixed point location is unknown.
Marcus: They also showed it works on generalized lotka–volterra systems by capturing the predictive direction associated with each transition between saddles in a stable heteroclinic channel.
Yuki: When applied to naturalistic stimuli, LPD learned temporal filters qualitatively resemble electrophysiologically measured neuronal receptive fields, including center–surround structures and sustained and transient filters.
Ines: The paper also proposes an internal criterion for selecting the number of output components k by observing where the objective function plateaus, which gives a natural estimate of the number of dynamical modes.
Marcus: They found that precision quantifies quality while penalizing spurious components, and recall measures how many ground-truth directions are recovered.
Yuki: The paper also mentions that LPD achieves higher normalized mutual information at low and intermediate noise levels, converging with finite differences only at high noise.
Ines: It’s interesting they found that the learned LPDs correspond to modes where small perturbations influence future trajectories, which suggests a natural interface with control.
Marcus: The authors also mentioned code availability for reproducing all experiments at https://github.com/MarcoZenari/local-predictive-directions and the arXiv identifier is 2611 point 07809v1 q-bio.NC.
Yuki: And they pointed out that the temporal differencing removes the translation, while whitening by Cpp and Cf f removes the remaining linear part, leaving outputs invariant and filters equivariant.
Lucky paper: 2611.06934: Ines: Okay, let's look at BrainPFN, which is paper number two of five on the list. So it’s about predicting brain behavior using prior data fitted networks to solve small sample size problems in studies.
Marcus: The core innovation there seems to be amortizing inference across multiple scores and cohorts within a single forward pass instead of fitting separate models for each one.
Yuki: They use a prior built from seven thousand real functional connectivity matrices to sample one billion synthetic training samples, which lets the network learn a mapping from those matrices to behavioral outcomes.
Ines: So they're using that synthetic data generation process to teach the model how to handle those small sample sizes without needing per-dataset optimization at inference time.
Marcus: The derivation shows that the required noise variance sigma squared needed for a target correlation r is calculated by the formula (one minus r squared) over r squared.
Yuki: That formula helps determine how much noise you need to expect for a given correlation, which is pretty concrete.
Ines: They claim BrainPFN consistently outperforms Kernel Ridge Regression in the small-sample regime across all tested targets and contexts.
Marcus: But they also noted that as the context size N grows to approximately two hundred subjects, which is several times the median OpenNeuro cohort size, Ridge regression gradually closes that gap.
Yuki: That’s an interesting finding because it suggests the advantage isn't guaranteed at larger sample sizes either.
Ines: They also made a public release of the meta-training pool, which is a fraction of the size and free of restrictive data use agreements, which is pretty significant for researchers.
Marcus: It shows that a modest, openly available pool can match or exceed models trained at massive scales like those from UK Biobank.
Yuki: The network was trained for ten thousand steps using the AdamW optimizer with a learning rate of three times ten to the negative four and weight decay of ten to the negative two on a single V100 or A100 GPU.
Ines: That training setup gives us some concrete details about how much compute is needed for that initial meta-training phase.
Marcus: The analysis validated BrainPFN's performance across four independent cohorts and various cognitive measures like IST fluid, IST intelligence total, NEO A, and BAS fun across different brain regions.
Yuki: So they tested it on several things simultaneously to see if the geometric prior held up generally.
Ines: While they found a consistent advantage over Ridge regression in small samples, they also noted that no of the tested foundation models consistently maintained superiority as N grew past two hundred subjects.
Lucky paper: 2611.06936: Ines: So we are looking at From Maps to Metabolism, which tackles whether consciousness needs more than just computation to exist.
Marcus: The paper synthesizes several ideas, drawing from thinkers like Korzybski and Searle to propose a non-computational physicalism.
Yuki: Essentially, the authors argue that consciousness depends on intrinsically organized, metabolically grounded physical dynamics instead of abstract computation.
Ines: They specifically critique mechanistic computationalism by saying that grounding computation in neural mechanisms doesn't give it intrinsic significance or the normativity we see in conscious systems.
Marcus: The central argument is that computation is just an abstract, observer-dependent description of physical processes, while consciousness is an intrinsic feature of certain kinds of physically organized systems.
Yuki: They suggest that for a system to be conscious, it needs to be self-maintaining and genuinely invested in its own continued existence through metabolic grounding and autopoietic organization.
Ines: That means the formal structure of computation itself isn't what matters; whether the physical system is self-maintaining is what determines if it has intrinsic normativity.
Marcus: The paper concludes that computation cannot be the vehicle through which consciousness acts on the world, especially when you consider large language models lacking that embodied, metabolically invested existence.
Yuki: That leads to a crucial point: artificial general intelligence understood only as a computational system can't replicate the full causal profile of a conscious human agent due to this lack of grounded existence.
Ines: It shifts the question from what computation is implementing to what kind of physical thing the system actually is, which requires metabolic physicalism.
Marcus: I wonder how this connects back to what we discussed on embodied AI control earlier when we talk about improving artificial controllers?
Yuki: The connection is that if we want robust controllers, they need that grounding in self-maintaining dynamics, not just a good set of learned weights or an abstract computational graph.
Ines: It’s about moving beyond just the computation to focusing on the physical reality of what's running the process.
Marcus: I see how that ties into Capturing Membrane Dynamics and Spike Timing of Human Neurons at Scale using Neural Operators, where they learn shared operators instead of separate models.
Yuki: That framework helps because it models large populations by learning these operators, suggesting a more unified physical organization underneath the individual neuron details.
Ines: So when you look at the results, what specific evidence do they use to support this idea that metabolic grounding is key?
Marcus: They focus on demonstrating that consciousness requires self-maintaining dynamics and autopoiesis rather than just simulating those dynamics through computation.
Yuki: The paper doesn't give a single measurement, but it uses these philosophical concepts—self-maintaining, metabolically grounded—to define the necessary physical properties for intrinsic normativity.
Ines: That’s a big conceptual leap from just training a model to defining what kind of system qualifies as conscious in the first place.
Marcus: And this has implications for the future of AI, doesn't it? It suggests that impressive intelligence alone isn't enough if you lack that lived, physically grounded existence.
Yuki: Precisely; so artificial general intelligence, defined purely computationally, can never replicate the full causal profile of a conscious human agent because it misses that vital grounding.
Ines: So the paper’s main contribution is restoring gravity to the mind question by tying it firmly to self-maintaining physical dynamics.
Lucky paper: 2605.06304: Ines: So we're looking at A multi-scale information geometry reveals the structure of mutual information in neural populations today.
Marcus: It introduces a unique Riemannian geometry that emerges from how distances contract as stimulus resolution gets lost through coarse-graining.
Yuki: This results in a multi-scale extension of the Fisher information metric that captures encoding structure from fine stimulus details to coarse global distinctions, which is really interesting because it connects directly to how information is encoded.
Ines: It says the resulting geometry is exactly related to the mutual information encoded by the population, where well-encoded stimulus directions get stretched more strongly and poorly encoded ones get contracted.
Marcus: That means mutual information equals the expected rate at which squared distances increase under infinitesimal perturbations of the stimulus.
Yuki: And that identity I R X equals one over two times the trace of G(x) is shown to be equivalent to equation twelve in the main text, which links mutual information directly to that expected rate of distance increase.
Ines: The paper proposes using diffusion models to estimate this geometry, and it depends only on differences between two samples from the same conditional model, which avoids mismatch bias.
Marcus: That makes it practical for large neural populations and high-dimensional stimuli because the estimator depends only on those specific differences.
Yuki: The leading eigenvectors of the multi-scale metric tensor G(x) identify stimulus variations that contribute most strongly to encoded information, revealing which stimulus variations contribute most to mutual information.
Ines: These directions yield interpretable image features that differ systematically across visual areas, which is a nice practical result for understanding neural coding.
Marcus: Comparing different geometries showed that the subspace alignment of A and B only constrains the angles between eigenvectors and is insensitive to the magnitude of the associated eigenvalues.
Yuki: They argue that the Bures–Wasserstein distance is a more appropriate choice for decodable information notions because it has a direct optimal-transport interpretation between those two representational geometries.
Ines: But they do flag a limitation, saying that like the Fisher information, it requires learning a probabilistic encoding model p of r x to describe how neurons respond to stimuli.
Marcus: And another limitation is computational, because estimating the geometry is more expensive than estimating the Fisher information because you need training a diffusion model and performing multiple evaluations for Monte Carlo estimates of the metric tensor.
Lucky paper: 2611.06926: Ines: We're moving into paper five of five today, and it’s about Foundation Models as Modular Tools for fMRI Analysis.
Marcus: It looks like this work is trying to solve a problem where different brain data aspects need different models but there isn't a good way to combine them without retraining everything.
Yuki: The authors propose making each Foundation Model an independent tool server that a fixed language agent calls next to classical tools and literature tools, which avoids the need for retraining entirely.
Ines: So instead of aligning all these different models, you just add a new model to the server list?
Marcus: Exactly. They set up four main servers: piop1 for time-series and graph metrics, pubmed for literature tools, brainlm for reconstruction error and forecasting, and omni fmri for patch embeddings.
Yuki: The gain they report is a holistic answer quality improvement of plus zero point seven zero on a rubric scored from one to five across thirty-four runs.
Ines: So what does that actually mean for the person listening who just wants to know how much better the answer is?
Marcus: It lands in how the agent selects, parameterizes, and grounds its tool calls, and this gain peaks where the agent has to cross between different model vocabularies.
Yuki: Specifically, tool selection improves by plus one point zero nine across all questions tested.
Ines: That sounds like a lot of improvement in how the AI decides what to do next, but they also mentioned some challenges?
Marcus: Yeah, there are two failure modes identified where the two models answer different questions well, but an agent using both scores the same as one using just brainlm alone.
Yuki: Plus, a voxel-addressed model ends up running seven point eight times more without a gradeable answer compared to zero point six two percent when listing coordinates.
Ines: So that context window issue where listing coordinates just fills up the agent's memory is a real problem for deployment?
Marcus: It is critical because an FM only reaches the final answer through the calls an agent can make and the tokens those calls consume, so that interface property really matters as much as its representation.
Yuki: The main open problem they flag right now is orchestration, which involves creating a policy that picks the right model for each question and treats disagreement between models as a signal.
Ines: So if you only listen to the show, what does that orchestration mean in practice?
Marcus: It means figuring out how to make the system intelligently decide which tool server is needed for any given question without needing a human to pre-program every single path.
Yuki: The paper concludes that an FM providing all three necessary interface properties can work alongside other models as soon as it's released.
Ines: And what about the benchmark they used? How robust was the test?
Marcus: They used a seventy six question benchmark with ten families along three axes: single-FM workflow, interface crossing, and multi-FM synthesis, scored by an independent LLM judge.
Yuki: The holistic score that was the headline outcome across all those criteria is what they use throughout the paper.
Ines: I see they also pointed out a detail about naming conventions between piop1 and brainlm regions, which is a practical thing to watch for when building these things.
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