arXane › Quantitative Biology › Quantitative Biology
Wednesday, 4 November 2026: 3 new papers, 2 cross-lists and 2 replacements in Neurons and Cognition (q-bio.NC), in arXiv's order, with full abstracts. The same papers as arXiv's own listing, set as a newspaper; every title links to arXiv.
Neuroevolution can discover compact network designs that are hard to specify by hand. The EONS framework evolves the topology, weights, thresholds and delays of neuromorphic spiking networks without gradients. We tested whether giving integrate-and-fire (IF) neurons an activity-dependent calcium current (Ca-IF neurons) improves evolved networks for EEG seizure detection, with calcium parameters fixed, shared across a network, or evolved for each neuron. On 10 held-out CHB-MIT patients never seen in training, per-neuron calcium raised the record-level Matthews correlation coefficient (MCC) from 0.592 to 0.726 over 20 independent runs (p = 0.008) and lowered the false-alarm rate from 0.17 to 0.10. Fixed or per-network calcium gave no significant benefit. IF networks with evolved synaptic delays of 1-3 ticks gave a similar gain (p = 0.005), whereas axonal delays did not. On Siena, with unchanged settings, per-neuron calcium gained +0.096 MCC over IF in one patient-disjoint split (p = 0.011) and +0.059 in the other (p = 0.052). These results suggest that calcium and delays may make the evolutionary search more effective, and motivate testing on more data and in neuromorphic hardware.
We extend FLYNN, a leaky integrator recurrent network wired on the Drosophila melanogaster connectome, with a 32-ray 2D LiDAR interface that requires no learned fusion layer beyond a scalar input gain. A 270 degree frontal scan at 8 m range is injected as inverted proximity into the 37 BM Vib head-bristle neurons. Across 24 paired comparisons in a four- state vision matrix with LiDAR on or off, FLYNN-LiDAR has 17 significant wins, 5 nonsignificant comparisons, and 2 significant losses against three LiDAR-capable baselines. Both losses are to MobileNet with the left eye off. With full vision, FLYNN- LiDAR reaches 96.6% success. Under total vision loss, the scan raises success from 45.2% to 64.6%; the baselines reach 44.6% to 56.6% with the scan on. Under out of distribution textures, it is the only LiDAR-capable model whose success improves with the scan (44.2% to 58.4%); both CNNs score 0%, and SmallWorld declines from 26.8% to 7.4%. Under periodic vision dropout, FLYNN-LiDAR achieves 93.6% and 89.6% success at 25% and 50% duty with fewer collisions. Profiling the evaluated forward paths gives 3.93M multiply accumulate operations per step; EfficientNet and MobileNet require 4.60x and 2.94x as many. Final- state analysis shows near-linear sensory superposition (cosine 0.9955).
Increasingly complex neural models predict blood-oxygen-level-dependent (BOLD) responses to naturalistic stimuli, but the predictive value of this added complexity remains unclear. We ask when pretrained stimulus representations help and how much downstream complexity is needed, with and without recent measured BOLD history. We compare linear diagnostic references and nonlinear predictors across HCP movie fMRI (154 subjects), BOLD Moments (10 subjects), and Algonauts 2025 (4 subjects). Under a shared offline protocol, models are fitted on training subjects and evaluated on held-out subjects viewing largely familiar stimuli, without refitting. Pretrained visual features generally improve prediction over raw pixels; for linear autoregression, this advantage becomes more pronounced over longer rollouts (recursive predictions without additional measured BOLD). Measured BOLD history provides substantial gains that generally diminish {as rollouts lengthen}. Nonlinear stimulus-only predictors can improve on linear readouts, whereas linear autoregression provides a strong history-conditioned baseline. Controls examine context length, regularization, fitting procedures, and model capacity. Regional response consistency across subjects is associated with held-out prediction accuracy, while less consistent regions tend to gain more from measured history. In paired simulations with fixed linear neural dynamics, slower hemodynamic responses generally sustain high temporal prediction correlations over longer rollouts, illustrating how the observation process can contribute to predictive persistence. Together, these findings support evaluating stimulus representations and downstream complexity separately, using history-conditioned linear prediction as a diagnostic reference under matched inputs and measurement availability.
During learning, humans often benefit from curricula that teach simple rules before exceptions. However, we lack a computational explanation for this phenomenon. Here, using a canonical behavioural task and modelling with neural networks, we develop and test a theory for why simple-to-complex curricula succeed. We show that during learning, both humans and neural networks show simplicity biases, a tendency to learn simple rules that are underfit to the data. In neural networks this bias is accompanied by compression of neural representations. Simple-to-complex curricula help both types of learners overcome this compression, developing representations (and policy) with appropriate dimensionality for the task. We test and confirm a further prediction from this theory: learners that have been pre-exposed to simple tasks and with more compressed initial representations benefit more from curricula. These findings suggest that curricula help with ``unlearning'' of simple rules.
Real-time phase estimation on stimulus-aligned transient EEG is critical for closed-loop neuromodulation systems targeting the theta-band, yet most real-time phase estimation methods have been developed and validated on resting-state, high-fidelity signals in the alpha band. Recent work has pursued deep learning architectures, while individualized classical bandpass filter tuning remains underexplored. This study characterizes the spectral and temporal properties of non-anticipative bandpass filtering and demonstrates the performance gains achieved through individualized tuning. We present a Bayesian-Optimized Biquad Filter-Burg (BOB) model evaluated on EEG recorded during peripheral vibrotactile stimulation of the left median nerve (theta, N=20 healthy subjects) and on resting-state recordings from the Max Planck Institute's LEMON dataset (alpha, N=20 healthy subjects). BOB was benchmarked against the Deep Dual-Branch Complex Neural Network (DBCNN), EEG Phase Prediction Network (EPN), Phastimate, and Phase Estimation with Autoregressive Padding (PEAP) under matched testing conditions: channel selection, individualized optimization, evaluation time points, and an offline reference phase obtained through zero-phase filtering and the Hilbert transform. Under these testing conditions, BOB achieved matched or better performance than these models at a fraction of the calibration time.