11 articles · sorted by date
When speech is gone, the brain may still be holding more language clues than today’s decoders use. That is the promise behind MoDAl, a new framework for speech neuroprosthesis, which tries to read intended speech from neural activity without audible output. Instead of relying mainly on motor cortex, it trains several parallel brain encoders and matches them to text embeddings from a pretrained large language model. A second objective keeps those encoders from collapsing into the same representation, so the system can uncover different neural signals rather than duplicates. On the Brain-to-Text Benchmark ’24, this setup lowered word error rate from 26.3% to 21.6% compared with the previous best end-to-end method. The paper reports that the improvement from adding previously discarded area 44 input came entirely from the decorrelation mechanism. The discovered area 44 encoders specialized in structural and syntactic features, including sentence length, grammatical voice, and wh-words, matching what neurolinguistics already expects from Broca’s area.
If you want brain-reading software to get smarter, this paper says the fastest path is more data, not a bigger model. The team built OmniMouse from 3.1 million neurons in the visual cortex of 73 mice, collected over 323 sessions and more than 150 billion neural tokens. Those recordings came from natural movies, images, parametric stimuli, and behavior. OmniMouse is a multi-modal, multi-task model that can flexibly handle neural prediction, behavioral decoding, neural forecasting, or any combination of the three at test time. It reached state-of-the-art performance and beat specialized baselines across nearly all evaluation settings. The surprise was in how it improved: adding more data kept helping, while increasing model size eventually hit a wall. That flips the usual AI scaling story from language and vision, where bigger models often carry the day. Here, even in the mouse visual cortex, the models still look data-limited. The authors say this kind of scaling pattern could hint at phase transitions, where larger and richer datasets unlock new capabilities.
If a neural circuit takes too long to answer, it can miss the pattern it was built to detect. That timing problem is the focus of this paper on lateral predictive coding, a simple framework for feature detection in biological neural circuits. The authors show that the system’s characteristic response time can be pushed down close to the lower bound while keeping the mean predictive error, which they treat as energetic cost, and the robustness of signal transmission intact. They also report that networks organized in a modular structure, with far fewer lateral interactions, perform just as well as fully connected networks. On feature detection, response time, energetic cost, and information robustness, the modular designs stay on equal footing with all-to-all networks. The result is a cleaner picture of how recurrent neural systems can be both fast and efficient, without giving up the tradeoffs that make them reliable.
If two vibrations arrive in a row, the first one can change how strong the second feels. That order effect has long been a puzzle in haptic perception, where people judge pairs of sequential stimuli differently depending on which comes first. This paper introduces a dynamical Bayesian model, meaning a model that blends noisy sensory input with an internal estimate that keeps evolving over time. Tested on vibrotactile discrimination experiments, it reproduced both the direction and the size of these time-order effects across subjects, along with the differences between individuals, using only a small number of parameters. The fitted parameters also gave a compact readout of each person’s prior expectations and noise characteristics. Beyond matching the data, the model reshaped stimulus space itself, revealing a subject-dependent geometry of perceived touch. In that transformed space, the judgments showed approximate symmetries that are missing in the raw physical stimulus coordinates, suggesting that temporal bias is built into how the mind infers touch.
Why do brains and neural network (a software system inspired by how brain cells connect) models sometimes look so similar? This paper says the answer may not be a single universal code hidden in all minds and machines. Instead, it argues that alignment between artificial neural networks and biological brains can emerge when different systems face overlapping ecological constraints: the pressures of the environments they grow up in. The authors call this the Umwelt Representation Hypothesis, or URH. They review evidence that representational differences across species, individuals, and neural networks are systematic and adaptive, which is hard to fit with the idea of universality. They also note that recent comparisons can show similar matches to neural data across very different models, especially when the mapping from model to brain is flexible. The payoff is a new way to think about model comparison: not as a hunt for one perfect world model, but as a way to map clusters of alignment in ecological constraint space.
A choice is rarely a frozen snapshot; it can feel like a tug-of-war that settles, wobbles, or suddenly flips. This paper argues that decision making is better captured as a moving process shaped by an informational environment, not just a static yes-or-no state. The authors survey quantum-like models of cognition and then build around the Gorini-Kossakowski-Sudarshan-Lindblad (GKSL) master equation, a tool from open quantum systems that describes how a mental state evolves through dissipation. They separate two dynamical regimes, called passive and active Hamiltonians, and show that non-commutation with decision projections marks cognitive agency and what they call quantum escape from classical equilibria. The framework is also used to stabilize non-Nash outcomes in strategic games such as the Prisoner’s Dilemma. A second signature they highlight is “cognitive beats”: slow modulations that appear when two competing flows of mind run at nearly the same frequency, creating peaks of readiness and hesitation. In their picture, those beats offer a spectral diagnostic for the depth of deliberation and the complexity of the underlying thought process.
When brain differences are scattered across the cortex, a single thickness score can miss them. That is the problem the paper tackles in juvenile myoclonic epilepsy, where structural abnormalities are described as subtle and spatially distributed. Instead of collapsing the brain’s shape into one local measure, the authors introduce a Poisson flow model built from gradients of the mean curvature field on the cortical surface. The method turns that information into a smooth scalar field by solving a Poisson equation, and the surface gradient of that field becomes a flow representation of folding organization. In plain terms, it offers a way to trace how sulci and gyri fit together across the cortex, rather than judging each spot in isolation. The paper presents this as a principled geometric framework for studying distributed cortical alterations in JME, where conventional morphometric measures such as cortical thickness have limited sensitivity.
Today’s AI can talk and predict, but it still struggles with the basics of acting in the real world. A workshop report from August 2025 says current systems face three big gaps: they cannot reliably interact with the physical world, they learn in ways that make them brittle, and they use too much energy and data. The paper points to neuroscience for clues: designing bodies and controllers together, learning through prediction and interaction, using multi-scale learning with neuromodulatory control, building hierarchical distributed architectures, and relying on sparse event-driven computation. It also lays out a roadmap for near-, mid-, and long-term research. The authors argue that making this happen will take a new generation of researchers trained across neuroscience and engineering, plus the right support: interdisciplinary training, hardware access, community standards, and ethics. In their view, NeuroAI could help fix current AI’s limits while also revealing more about how biological neural computation works.
When signals loop instead of flowing neatly in one direction, classic causal models can miss the action. That is a big problem for brain networks, where recurrent and higher-order interactions are common. This paper replaces the usual acyclic view with a variational causal framework based on a minimum energy principle. In this picture, causality is directional energy flow from high-energy to low-energy states along network connections. Using Hodge theory, the authors split network flows into dissipative parts and a persistent harmonic part that captures stable cyclic interactions. Applied to resting-state fMRI connectivity, the framework reveals robust cyclic causal patterns that conventional causal models do not detect. The result points to a different way of thinking about causality: not just as a chain of causes, but as an energy landscape where loops can be real and persistent.
If you've ever tried to run the same psychology task on different lab computers, the software can be the real obstacle. Goxpyriment tackles that problem by compiling a whole behavioral or cognitive experiment into a single executable file, with zero runtime dependencies. Because Go can embed graphics, audio files, and stimulus lists directly into the program, collaborators and testing machines do not need a separate runtime setup. The framework is inspired by Expyriment and includes text, shapes, images, Gabor patches, motion clouds, WAV playback, and tone generation. It also takes timing seriously: the operating system timestamps input events at hardware-interrupt time, so reaction times come from comparing two OS-level timestamps instead of continuous polling. Go's garbage collector can be disabled to reduce unpredictable pauses that could disrupt stimulus timing. The authors also provide more than forty psychology experiments built with Goxpyriment, and they say these examples help both humans learning the system and AI-assisted coding tools.
When a model predicts brain activity, the hard part is judging whether it is truly right or just matching noise. That problem matters because the signal scientists want is buried inside noisy electro- and magneto-encephalography, or MEEG, measurements. This paper tackles that by comparing model predictions to a ground-truth approximation, built with canonical correlation analysis, a way to align two data streams, and with participant averaging. The new score, called CPA-PA, improved single-participant evaluations by about 300–1000% on synthetic EEG data and by about 250% on 34 real MEEG datasets, covering 818 datapoints. The gains came from better sensitivity to stimulus-relevant neural activity and less dependence on signal-to-noise ratio. In plain terms, the paper offers a sturdier way to tell whether an encoding model is capturing the brain’s response to a stimulus, instead of being fooled by measurement noise.
