Research map / Continual learning / Online and replay-based
Thalamus: a brain-inspired algorithm for biologically-plausible continual learning and disentangled representations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use latent updates to adapt task context while keeping model weights frozen | Thalamus · Split MNIST, all 5 tasks | end average accuracy: 97.71 (± 0.13); RtF: 97.31 (± 0.11) | Helped |
| Use latent updates after learning the first four tasks, with weights frozen | Thalamus · Split MNIST, unseen task 5 | accuracy: 91.14±3.73 | Helped |
| Use latent updates to handle flipped-label contextual tasks | Thalamus · Contextual split MNIST | accuracy: 96.09 ±0.83; other methods: around 59% | Helped |
| Use SGD with 10^-4 learning rate instead of the usual Thalamus update balance (ablation) | Thalamus · Split MNIST, all tasks and unseen task 5 | accuracy: 83%; unseen fifth task: 67% | Hurt |
| Remove the accuracy criterion and alternate weight and latent updates every trial (ablation) | Thalamus · NeuroGym, unlabeled task stream | accuracy change after latent updates | Hurt |
| Replace Bernoulli latent-to-network gates with rectified Gaussian gates (ablation) | Thalamus · NeuroGym, 5 tasks | performance | No clear effect |
| Add Bernoulli sparsification to rectified Gaussian latent-to-network gates (ablation) | Thalamus · NeuroGym, 5 tasks | performance | Helped |
| Allow latent-to-network gate weights Wz to be trainable rather than fixed (ablation) | Thalamus · NeuroGym, 5 tasks | performance | No clear effect |
| Use task identifiers, training to criterion, and rehearsal together during pretraining (ablation) | Pretrained RNN · 15 NeuroGym tasks | average accuracy | Helped |
Closest papers by meaning
- Rapid Learning without Catastrophic Forgetting in the Morris Water Maze (ICML 2024 · similarity 0.88)
- Continuous Thought Machines (NeurIPS 2025 · similarity 0.88)
- SPICED: A Synaptic Homeostasis-Inspired Framework for Unsupervised Continual EEG Decoding (NeurIPS 2025 · similarity 0.87)
- InTAct: Interval-based Task Activation Consolidation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- Predictive Attractor Models (NeurIPS 2024 · similarity 0.86)
- Task-Free Continual Generation and Representation Learning via Dynamic Expansionable Memory Cluster (AAAI 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.