Abstract:
We introduce Online-CBGT-Net, a neuromimetic architecture for interpretable, online prediction in streaming environments. Inspired by the cortico-basal ganglia-thalamic (CBGT) circuit in the mammalian brain, our model integrates evidence accumulation and a reset mechanism for continuous, online decision making. At each time step, the model receives input and incrementally accumulates evidence. A decision is made only when the accumulated evidence surpasses a threshold, enabling the model to defer predictions in low-information conditions and act decisively when confident.

A key innovation is the incorporation of a reset factor, that scales down accumulated evidence after each decision. This allows the model to preserve temporal context across predictions to influence future predictions by past evidence. We further enhance interpretability by modeling separate Go and No-Go pathways using dual encoder branches, which produce supporting and opposing evidence for each class. This explicit separation allows users to inspect not only what the model favors, but also what it opposes, mirroring inhibitory competition in the CBGT circuit of mammalian brain.

We evaluate Online-CBGT-Net on the CalMS21 behavior dataset, demonstrating improvements over episodic CBGT-Net and LSTM baselines across F1 score, latency, and persistence. Additionally, we analyze the internal evidence distributions, revealing how the Go/No-Go streams highlight the model’s confidence in each decision. Our findings show that neuromimetic structures such as thresholded accumulation, reset mechanisms and dual-pathway encoding can produce transparent, confident, and biologically grounded decisions, advancing the goal of trustworthy AI in real-time sequential tasks.

Committee:
Prof. Katia Sycara (advisor)
Prof. Steven Chase
Renos Zabounidis