Репост из: the last neural cell
The most tasty neuro papers in 2025
Мой личный, абсолютно субъективный список лучших статей по нейро-декодингу за 2025 год. Тренд года: кидай больше данных и молись чтобы стало лучше. Но главный вопрос по-прежнему в том, как объединить данные между животными, областями мозга и типами электродов в одной модели.
Разбил 10 статей на 4 секции: инвазивные сигналы, неинвазивные, нейро софт и напоследок принес статью из нашей лабы (я не соавтор, работа началась ещё до меня).
01. Invasive signals
▸ Foundation model of neural activity predicts response to new stimulus types
Foundation model trained on mouse V1 generalizes to completely new stimulus types (motion, noise, Gabor) and predicts cell types from MICrONS.
Nature 2025
▸ Neural Encoding and Decoding at Scale (NEDS)
One model does both encoding and decoding across 83 mice on Neuropixels/IBL data. Multi-task masking, SOTA on both directions, predict brain activity and behavior.
ICML 2025
▸ A cross-species neural foundation model for end-to-end speech decoding (BIT)
End-to-end brain-to-text: transformer pretrained on 367 hours of Utah array data (human + monkey), paired with audio-LLM. WER drops from 24.69% to 10.22%, transfers to imagined speech.
ICLR 2026
▸ A Generalist Intracortical Motor Decoder (NDT3)
Autoregressive transformer pretrained on 2000 hours of intracortical spiking from 30+ subjects. Beats from-scratch models when you have less than 1.5 hours of task data.
NeurIPS 2025
▸ OMiSO: Adaptive optimization of state-dependent brain stimulation
Closed-loop stimulation that uses pre-stim brain state to pick which electrodes to zap on a Utah array in monkey PFC. Adapts online, outperforms state-blind methods.
NeurIPS 2025
02. Noninvasive signals
▸ REVE: A Foundation Model for EEG
MAE on 60,000 hours of EEG from 25,000 subjects across 92 datasets. 4D positional encoding handles any electrode config. SOTA on 10 downstream tasks.
NeurIPS 2025
▸ LSM-2: Learning from incomplete wearable sensor data
Foundation model for wearables pretrained on 40M hours of day-long sensor data. Learns directly from incomplete recordings via learnable mask tokens, no imputation.
arXiv 2025
03. NeuroTools
▸ Kilosort4
Graph-based clustering with merging trees for spike sorting. Best across all conditions, handles low-amplitude neurons under high drift.
Nature Methods 2024
▸ Jaxley: differentiable simulation of biophysical neural models
JAX framework for biophysical neuron models with autodiff and GPU acceleration. Fit compartmental models orders of magnitude faster, trained a 100k-param morphological network on vision.
Nature Methods 2025
04. Our lab
⚡️ Neural correlates of phosphene perception in blind individuals: A step toward a bidirectional cortical visual prosthesis
Science Advances 2025
p.s.
Пишите в комментарии, какие работы я пропустил и должен был добавить в этот топ.
Мой личный, абсолютно субъективный список лучших статей по нейро-декодингу за 2025 год. Тренд года: кидай больше данных и молись чтобы стало лучше. Но главный вопрос по-прежнему в том, как объединить данные между животными, областями мозга и типами электродов в одной модели.
Разбил 10 статей на 4 секции: инвазивные сигналы, неинвазивные, нейро софт и напоследок принес статью из нашей лабы (я не соавтор, работа началась ещё до меня).
01. Invasive signals
▸ Foundation model of neural activity predicts response to new stimulus types
Foundation model trained on mouse V1 generalizes to completely new stimulus types (motion, noise, Gabor) and predicts cell types from MICrONS.
Nature 2025
▸ Neural Encoding and Decoding at Scale (NEDS)
One model does both encoding and decoding across 83 mice on Neuropixels/IBL data. Multi-task masking, SOTA on both directions, predict brain activity and behavior.
ICML 2025
▸ A cross-species neural foundation model for end-to-end speech decoding (BIT)
End-to-end brain-to-text: transformer pretrained on 367 hours of Utah array data (human + monkey), paired with audio-LLM. WER drops from 24.69% to 10.22%, transfers to imagined speech.
ICLR 2026
▸ A Generalist Intracortical Motor Decoder (NDT3)
Autoregressive transformer pretrained on 2000 hours of intracortical spiking from 30+ subjects. Beats from-scratch models when you have less than 1.5 hours of task data.
NeurIPS 2025
▸ OMiSO: Adaptive optimization of state-dependent brain stimulation
Closed-loop stimulation that uses pre-stim brain state to pick which electrodes to zap on a Utah array in monkey PFC. Adapts online, outperforms state-blind methods.
NeurIPS 2025
02. Noninvasive signals
▸ REVE: A Foundation Model for EEG
MAE on 60,000 hours of EEG from 25,000 subjects across 92 datasets. 4D positional encoding handles any electrode config. SOTA on 10 downstream tasks.
NeurIPS 2025
▸ LSM-2: Learning from incomplete wearable sensor data
Foundation model for wearables pretrained on 40M hours of day-long sensor data. Learns directly from incomplete recordings via learnable mask tokens, no imputation.
arXiv 2025
03. NeuroTools
▸ Kilosort4
Graph-based clustering with merging trees for spike sorting. Best across all conditions, handles low-amplitude neurons under high drift.
Nature Methods 2024
▸ Jaxley: differentiable simulation of biophysical neural models
JAX framework for biophysical neuron models with autodiff and GPU acceleration. Fit compartmental models orders of magnitude faster, trained a 100k-param morphological network on vision.
Nature Methods 2025
04. Our lab
⚡️ Neural correlates of phosphene perception in blind individuals: A step toward a bidirectional cortical visual prosthesis
Science Advances 2025
p.s.
Пишите в комментарии, какие работы я пропустил и должен был добавить в этот топ.