Paper-Conference | Jiyi Wang

Learning Reusable Motor Motifs for Continuous Animal Behavior Modeling

Sun, 27 Sep 2026 00:00:00 +0000

Learning the building blocks of behavior

Complex movements often combine simpler components: an animal can turn while grooming, or sniff while moving forward. Motif-based Continuous Dynamics (MCD) learns these reusable motor motifs directly from behavioral trajectories and models behavior as smoothly changing combinations of them.

Method

Learn from offline trajectories. Use transition-based spectral representation learning to discover a shared basis for dynamics, policies, and rewards.
Compose skills continuously. Represent behavior through time-varying motif weights, allowing multiple motifs to be active together.
Infer policies and reward structure. Connect the learned representation to maximum-entropy policies and inverse reinforcement learning.

Experimental findings

Evaluated on simulated gridworlds, maze navigation, and real animal behavior data. On the animal behavior benchmark, MCD outperforms Keypoint-MoSeq, SemiSeg, and OPAL in human-annotated behavior decoding accuracy and trajectory discrimination AUC. The learned motifs capture both transitions and simultaneous movements, such as locomotion and sniffing.

Research connection

This work studies how reusable, compositional skills can be extracted from observed behavior. It informs my broader interest in learning embodied skills from demonstrations. The experiments in this paper concern simulated tasks and animal behavior; transferring this approach to human motion and robots is a future research direction.

Leveraging Dale’s Principle as an Inductive Bias in Recurrent Neural Networks

Sun, 27 Sep 2026 00:00:00 +0000

A biological principle for more stable learning

Biological neurons send either excitatory or inhibitory outputs. EISep RNN incorporates this constraint, known as Dale’s principle, into recurrent networks while allowing each neuron’s excitatory or inhibitory identity to be learned.

Method

The architecture combines a learnable sign assignment using Gumbel–Softmax. These choices let the network adapt its excitatory–inhibitory organization during training while improving optimization stability.

Experimental findings

Multi-task learning: competitive performance on a 20-task cognitive benchmark, with more stable gradients and recurrent dynamics.
Continual learning: reduced task interference and stronger resistance to catastrophic forgetting across sequential cognitive tasks.
Reinforcement learning: improved recurrent advantage actor–critic performance on Lunar Lander and HalfCheetah, covering discrete and continuous control.
Interpretable structure: sparse, modular connectivity emerges from the excitatory–inhibitory constraint.

Research connection

Reliable policy learning depends on stable optimization and representations that can be reused across tasks. This work explores architectural inductive biases for those properties.

Inverse Reinforcement Learning with Switching Rewards and History Dependency for Characterizing Animal Behaviors

Wed, 01 Jan 2025 00:00:00 +0000

Inferring the intentions behind behavior

SWIRL (Switching Inverse Reinforcement Learning) infers changing goals from behavioral trajectories. It models long sequences as transitions between hidden decision-making modes, each with its own reward function and policy.

Method

Goal switching: transitions between hidden modes depend on the previous mode and the animal’s current state.
Behavioral memory: rewards and policies can depend on recent states, capturing how past experience shapes actions.
Joint inference: expectation–maximization alternates between inferring behavioral modes and learning their transitions, rewards, and policies.

Experimental findings

Gridworld: combining both levels of history dependence improves reward recovery and behavioral segmentation.
Mouse navigation: recovers interpretable water-seeking, home-seeking, and exploration modes in a 127-node labyrinth.
Spontaneous behavior: outperforms autoregressive HMM baselines in held-out likelihood. State-dependent switching helps, while adding action-level history does not improve this dataset.

Research connection

This work connects learning from demonstrations with interpretable reward inference, offering a way to study changing intentions under the complicated behavior.