Research

Observers discriminated the motion direction of a briefly presented perifoveal drifting stimulus under two difficulty manipulations — cardinal versus oblique directions, and large versus small tilt offsets. Increased task difficulty strengthened and prolonged microsaccade inhibition, producing delayed rebounds, and peak pupillary responses were both larger in amplitude and delayed for harder conditions. Discrimination response times correlated with both microsaccade rebounds and peak pupil responses. These delays are consistent with a prolonged period of sensory evidence accumulation, and the correlated dynamics suggest a shared neural mechanism underlying microsaccade and pupil responses.

Perceptual thresholds for motion coherence were significantly lower under natural gravity than under reversed-gravity conditions, regardless of variations in stimulus parameters and visual context. The results suggest that the human visual system inherently extracts the gravitational-acceleration cues conveyed by local motion signals and integrates them into a unified global motion percept, thereby facilitating the perception of complex motion patterns in natural environments.

Localizing visual areas on the cortical surface usually requires substantial time and expertise from human raters. We trained convolutional neural networks (U-Net with a ResNet backbone) to predict the boundaries and iso-eccentric regions of V1, V2, and V3 from combinations of anatomical, diffusion-weighted, and functional MRI data. CNNs given the same functional data available to human raters predicted these maps with accuracy statistically indistinguishable from inter-rater reliability, while CNNs lacking functional data were less accurate. The results show that CNNs can replace the manual effort of defining early retinotopic maps, but not yet the acquisition of functional data.

Using parallel behavioral experiments and artificial neural network models, this work examines how response bias shapes both accuracy and confidence in perceptual tasks with more than two alternatives, asking whether humans and ANNs show comparable signatures of biased confidence computation.

How are working-memory representations transformed between encoding and retrieval — through stable codes, dynamically updated subspaces, or their interplay? Combining high-resolution 7T fMRI from the Natural Scenes Dataset with recurrent neural networks trained on a naturalistic 1-back task, and using representational similarity, cross-decoding, and subspace-geometry analyses, the study finds convergent evidence for a mixture mechanism: early visual regions (V1–hV4) undergo large representational changes across encoding and retrieval, including both rotational and non-rotational transformations.

Delineating visual field maps and iso-eccentricities from fMRI is important but time-consuming. We trained U-Net CNNs with ResNet18 backbones to predict V1/V2/V3 boundaries or five iso-eccentricity regions from different combinations of input data (T1-weighted anatomy, T2*-weighted anatomy, diffusion-weighted tract endpoints, and functional retinotopy). All CNNs using functional data reached cross-validated accuracy indistinguishable from inter-rater reliability (dice coefficient ≈ 92%), while those without functional data were lower (~75%) but still exceeded existing non-CNN methods. This is the earlier conference version of the bioRxiv preprint above.

Microsaccades occur ~1–2 times per second and, though largely involuntary, are modulated by task demands. Using a 2AFC task in which observers judged the drift direction of a perifoveal Gabor (8 reference directions, 8 iso-eccentric locations), we varied difficulty via cardinal-versus-oblique directions and tilt offset. Microsaccade rates were suppressed prior to stimulus onset and were further modulated by difficulty: they decreased as the angular offset shrank and when stimuli drifted in oblique rather than cardinal directions. Greater fixational stability for harder trials may mitigate blur and prolong evidence accumulation, so microsaccades may index cognitive effort for perifoveal visual tasks.

We added a gain-based, feature-similarity attention mechanism to one layer at a time in matched visual (VGG16) and auditory (VGGish) networks, across detection and read-out tasks. In the visual models, attention reliably improved performance — more so when applied to higher layers — whereas the same mechanism did not improve performance in the auditory models, suggesting that bio-inspired attention does not transfer uniformly across sensory modalities.

Using psychophysics and eye tracking, we measured behavioral differences across motion directions (e.g., radial versus tangential) at different locations in the visual field, fitting psychophysical functions to compare sensitivity for different motion directions.

Using the hippoCorpus dataset of ~6,800 diary-like stories labeled as recalled, imagined, or retold, we built machine-learning classifiers (best model: XGBoost, ~76% accuracy) and asked which features best separate the three story types. Story-construction time and lexical detail (e.g., noun counts) were the most diagnostic features — supporting the view that recalled stories draw directly on episodic memory while imagined stories rely more on semantic memory, and connecting model feature-importance back to cognitive theories of memory.

We extended semi-supervised learning from parameter space into function space, combining Function-Space Variational Inference (FSVI) with Maximum Uncertainty Regularization, and compared three strategies for locating the most-uncertain "virtual" points used for consistency regularization — greedy search, a first-order entropy approximation, and a K-NN approximation. On SVHN and CIFAR-10, the K-NN variant gave the best trade-off between accuracy and computational cost.