Alexander Mathis

Closed-loop imitation learning reveals muscle-centric and latent-goal codes in primate sensorimotor cortex

This paper develops a closed-loop muscle-level model of primate grasping using imitation learning and neural recordings, revealing muscle-centric and latent-goal codes in sensorimotor cortex.

DLC2Action: A Multimodal Deep Learning-based Toolbox for Automated Behavior Segmentation

DLC2Action automates behavior segmentation using video, audio, and pose data, achieving strong performance across diverse datasets and supporting self-supervised learning and extensibility.

Compressed Cortical Input Separates Control from Dynamics in Striatum

The paper develops a corticostriatal neural network model to show how cortical input compression enables distinct control and dynamic functions in the striatum across time-sensitive tasks.

Mice navigate scent trails using predictive policies

This study shows mice use predictive policies to navigate scent trails by integrating bilateral olfactory input and short-term memory.

Deep-learning models of the ascending proprioceptive pathway are subject to illusions

The paper shows that deep-learning models of proprioception can be fooled by simulated muscle-tendon vibrations, producing illusions similar to those observed in humans.

Acquiring musculoskeletal skills with curriculum-based reinforcement learning

This paper uses curriculum-based reinforcement learning to train a neural network to control a musculoskeletal hand model, revealing insights into motor control and synergy analysis.

Modeling Sensorimotor Processing with Physics-Informed Neural Networks

This paper develops a physics-informed neural network model to simulate muscle spindle function, validated across species and datasets, revealing key drivers of variability in proprioceptive responses.

Elucidating the Hierarchical Nature of Behavior with Masked Autoencoders

The paper introduces a masked autoencoder framework to analyze hierarchical behavior in motion capture data using synthetic and real benchmarks.

Task-driven neural network models predict neural dynamics of proprioception

The paper uses task-driven neural networks to predict neural activity in proprioceptive pathways during reaching tasks, finding that task-optimized models outperform other models.

Contrasting action and posture coding with hierarchical deep neural network models of proprioception

The paper uses synthetic data to compare two tasks in proprioception modeling, finding that action recognition better explains neural tuning properties than trajectory decoding.

Keypoint-MoSeq: parsing behavior by linking point tracking to pose dynamics

Keypoint-MoSeq fits a generative model that tells tracking jitter apart from real movement, cutting pose time series into unsupervised sub-second behavioural syllables in mice.

Striatal dopamine explains novelty-induced behavioral dynamics and individual variability in threat prediction

The paper uses machine learning to analyze novelty-induced behaviors and finds that dopamine in the striatum correlates with individual variability and is explained by a reinforcement learning model of threat prediction.

Real-time, low-latency closed-loop feedback using markerless posture tracking

This paper introduces a real-time, low-latency pose estimation tool with feedback capabilities for behavioral neuroscience experiments.

Using DeepLabCut for 3D markerless pose estimation across species and behaviors

The paper updates and provides a step-by-step guide for using DeepLabCut, an open-source toolbox for markerless 3D pose estimation across species and behaviors.

On the inference speed and video-compression robustness of DeepLabCut

This paper improves DeepLabCut's inference speed and demonstrates its robustness to video compression.

Reading out olfactory receptors: feedforward circuits detect odors in mixtures without demixing

The paper models how mice detect odors in mixtures using a linear decoder with sparse weights, finding that they generalize poorly without demixing.