Dimitra Maoutsa

Nicely written piece that merges many fascinating frontiers of science and thought @SuryaGanguli amacad.org/publication/da…

i’ve seen this new therapist 4 times and every time we talked about how tired i am of academia. today she asked what my research was about should i find a new therapist

Normalise disrespecting people who disrespect you .

Me: entering Japanese cat café. employee stops me immediately. Employee: One rule. Me: Okay. Employee: Orange cat bites people from France. Me: I’m not French. Employee: Good. I sit down, cats everywhere, peaceful, adorable, therapeutic. then gigantic orange cat jumps onto table, built like retired wrestler. Cat staring at me aggressively. Employee watching nervously from distance. Me: I thought he only hated French people. Employee: He improvises. cat slowly pushes my drink off table while maintaining eye contact. Me: THIS IS TARGETED. small child nearby points at cat. Child: That one evil. Employee: No no, He just passionate. cat suddenly climbs onto my lap, starts purring violently. Me: …wait he likes me? Employee shocked. Employee: Impossible. another worker comes over, then another, entire staff now observing me like chosen prophet. Manager arrives. Manager: He has never trusted customer before. Me: What does that mean. Manager bows slightly. Manager: You must take him. Me: TAKE HIM WHERE. orange cat already asleep on me. Employee quietly bringing adoption papers. Me: I CAME HERE FOR COFFEE.

A tour de force by my student Kaito Iwasaki and collaborators for nonlinear non-Gaussian Bayesian filtering, SKF. Moment representation + score matching -> solving linear systems. It exactly recovers information from KF for linear-Gaussian systems. arxiv.org/abs/2605.16644

A neat takeaway from this paper: Wasserstein distance is not just used to compare distributions, but to define the geometry of learning itself. The drifting model is reinterpreted as a Wasserstein gradient flow, meaning samples move along the steepest descent direction in distribution space. In short, optimal transport becomes the rule that tells the generator where to move. On the Wasserstein Gradient Flow Interpretation of Drifting Models Paper: arxiv.org/abs/2605.05118

Neuron Adaptive reorganization of history encoding in the retrosplenial cortex supports flexible decision-making strategies This study investigates how the brain flexibly adjusts how much past experience is used during decision-making. Mice performed tasks with either rapidly changing or stable environments, and they adapted by relying on shorter or longer reward histories. Neurons in the retrosplenial cortex (RSC) reorganized their activity to encode different timescales of past choices and rewards depending on the task. The results suggest that the brain dynamically controls the timescale of memory integration to support adaptive decision-making. cell.com/neuron/abstrac…

CAT ADVISORY: It's very hot across the UK this weekend. All cats should find the coolest spot in their home and stay there. Tell your staff to bring chilled snacks and water every 15 minutes.

Episode #41 in #TheoreticalNeurosciencePodcast: On functional effects of neuronal heterogeneity – with David Dahmen theoreticalneuroscience.no/thn41 While network modelers often assume all neurons to be the same, biology says different. Can this hetereogeneity be a good thing?

Academics to students: "always aim for clarity" Academics to editors: "apologies for the nonsensical title but I couldn't resist the pun"

Go big or go home. Bigger reward leads to better learning is not exactly news. What's striking is that much of the systems neuroscience field persevered for decades with tiny rewards, based on old work suggesting dopamine responses saturate at modest reward sizes (more trials is better for recordings though). Worth flagging that this wasn't universal. The addiction field, including our lab, has long used 100 µL rewards for water, alcohol, and sweeteners in self-administration. Looks like we were onto something. Mice learn navigation, motor, and decision tasks ~10x faster with a few large rewards vs many small. Sustained NAc dopamine scales with reward size, drives engagement, drives learning. doi.org/10.1126/scienc…

A critical initialization for biological neural networks Spontaneous brain activity is often treated as noise: the background hum of a nervous system waiting for a task. But large-scale recordings in mice have shown something more structured. Even in darkness, without explicit stimuli, thousands of neurons display coordinated activity patterns that extend across the brain and persist far longer than the fast biophysical timescales of individual neurons. Marius Pachitariu and coauthors ask a simple question: could this macroscopic structure emerge from a simple kind of network initialization? Their answer connects neuroscience, random matrix theory and machine learning. They model spontaneous neural activity as linear dynamics governed by a random connectivity matrix, stabilized by a global inhibitory-like normalization. When this matrix is symmetric and critically normalized, with its largest eigenvalue very close to one, the network naturally produces high-dimensional activity modes with a power-law covariance spectrum. This is not just a mathematical curiosity. The same spectral structure appears in large-scale mouse recordings from cortex and brainwide Neuropixels data, with power-law exponents around 0.7–0.85. Hippocampal CA1 is the striking exception: its activity looks less correlated, closer to an efficient, high-capacity code for information storage. The ML perspective is especially interesting. In artificial neural networks, initialization is often treated as a technical detail: Xavier, He, orthogonal schemes, and so on. But this paper reframes initialization as a computational substrate. A critically initialized recurrent system can generate slow, global, high-dimensional modes before task-specific learning. In simulations, these dynamics support time-dependent computations, including zero-shot working memory tasks. The biological implication is powerful: spontaneous activity may not be random noise, but a preconfigured dynamical scaffold on which learning and computation can operate. The brain may start from an initialization already close to useful temporal memory, with learning then shaping readouts or task-specific pathways. For R&D teams building ML systems in drug discovery, materials development, energy research or biotechnology, the lesson is broader than neuroscience. Initialization, architecture and dynamics define what kinds of scientific signals a model can preserve, combine and retrieve before training. In applied research pipelines where data are scarce, noisy and time-dependent, designing the right dynamical substrate may be as important as choosing the loss function. Source: Pachitariu et al., Nature (2026) — CC BY 4.0 | doi.org/10.1038/s41586…

You CANNOT manipulate a human who studies patterns !!!

value-driven transport! a new framework for generative modeling, combining elements of * optimal control / RL * optimal transport * stochastic primal-dual optimization thread about our new work with @pablorenoz (@UPFBarcelona) & @adrianlmueller (@ETH_en) 1/

Your drifting model is secretly a fixed point for the Wasserstein gradient flow on... ...the KL? ...an approximation to the Sinkhorn? ...Is it even a Wasserstein gradient flow at all? arxiv.org/abs/2605.05118 @liwenliang @agalashov @JamesTThorn @ValentinDeBort1 @ArnaudDoucet1

Result 4: If you train on all 7 tasks at once, and the PCA projection of the residual stream just... gives you the world map. No digging for linear directions for world representations. Why multi-task training surfaces linear representations with higher magnitudes is an open question! 14/n

Result 3: Multi-task learning drives representational convergence. We now train models on 2 tasks jointly. We find that world representations become more aligned both qualitatively and CKA. This is true even across models trained on completely disjoint task sets. 11/n

The qualitative geometry is reliable across seeds: Same task → similar shape, regardless of initialization. 9/n

Result 2: The training task controls world representation geometry. Different tasks on the same underlying world produce qualitatively distinct geometries: distance → thread-like angle → 2D manifold compass → fragmented clusters inside → diffuse 8/n

Interestingly, once representational structure forms, it stays largely fixed for the rest of training. The geometry freezes early in training, while loss continues to drop and accuracy keeps rising. This suggests there are critical learning periods for representations during pretraining. 7/n