Research map / Label-efficient and robust learning / Self-supervised and contrastive
PhiNets: Brain-inspired Non-contrastive Learning Based on Temporal Prediction Hypothesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PhiNet, a SimSiam extension with an extra CA1 predictor inspired by the temporal prediction hypothesis, plus X-PhiNet with an EMA long-term encoder, showing more robust learning dynamics and better online/continual learning on CIFAR-5m.
Problem
Whether the temporal prediction hypothesis of hippocampal sequence memory is a reasonable computational model, tested through non-contrastive self-supervised learning, focusing on representational collapse, robustness, and adaptivity in online and continual learning.
Approach
Extend SimSiam with an additional predictor (mapped to CA1) on top of the original predictor (CA3) and a second loss (Sim-2), so that StopGradient-induced time delay mirrors the hippocampal CA3-CA1 prediction; X-PhiNet adds an EMA long-term encoder mapped to neocortex under complementary learning systems theory.
Key results
- On CIFAR-5m online learning, X-PhiNet (cos) reaches 87.46% linear probing accuracy versus 77.71% for SimSiam and 81.05% for BYOL.
- On Split CIFAR-5m (wd=5e-4), X-PhiNet (cos) achieves 92.83% average accuracy versus 90.84% for SimSiam, with lower forgetting (1.95 vs 2.45).
- Linear probing on CIFAR10 shows PhiNet (MSE) at 49.89% with wd=0 versus 25.41% for SimSiam, indicating better stability under weight-decay changes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the CA1 predictor and Sim-2 loss to SimSiam | SimSiam · CIFAR10 | accuracy: 25.41_{0.02} -> 49.89_{0.35}; 2.63_{0.18} -> 55.90_{1.57}; 60.82_{1.57} -> 33.92_{7.42}; 44.51_{33.64} -> 66.73_{0.03}; | Mixed |
| Use MSE rather than negative cosine loss for Sim-2 (ablation) | PhiNet · CIFAR10 | accuracy | Mixed |
| Use a separate CA1 predictor instead of setting g=h or g=I (ablation) | PhiNet · CIFAR10 | accuracy | Helped |
| Add a momentum encoder with EMA to PhiNet | PhiNet · CIFAR-5m | accuracy: 76.74_{1.82} -> 87.46_{0.14} | Helped |
| Add a momentum encoder with EMA to PhiNet | PhiNet · Split CIFAR-5m (wd=5e-4) | Acc: 90.69_{0.11} -> 92.83_{0.12} | Helped |
| Add a momentum encoder with EMA to PhiNet | PhiNet · Split CIFAR-5m (wd=2e-5) | Fg: 0.05_{0.19} -> -1.03_{0.41} | Helped |
| Use X-PhiNet (Cos) rather than X-PhiNet (MSE) (ablation) | X-PhiNet · Split CIFAR-5m (wd=5e-4) | Acc: 91.02_{0.36} -> 92.83_{0.12} | Helped |
| Apply data augmentation to the input of the long-term encoder (ablation) | X-PhiNet · CIFAR-5m | accuracy: 87.30_{0.13} -> 86.45_{0.25}; 87.46_{0.19} -> 86.96_{0.17} | Hurt |
| Replace the separate CA1 predictor with the identity, g=I (ablation) | X-PhiNet (Cos) · CIFAR-5m | accuracy: 87.46_{0.19} -> 86.71_{0.27} | Hurt |
| Train for one epoch rather than the standard CIFAR-5m training schedule (ablation) | X-PhiNet · CIFAR-5m | accuracy: 69.65_{0.29} -> 74.76_{0.52} | Helped |
| Increase training from one to two epochs per task (ablation) | X-PhiNet (Cos) · Split CIFAR-5m | Acc: 92.83_{0.12} -> 92.83_{0.06} | No clear effect |
Surprising findings
- PhiNet with MSE Sim-2 can underperform SimSiam on CIFAR-5m (76.74% vs 77.71%) without the EMA long-term encoder, and X-PhiNet with batch-size-matched g=I is worse than full X-PhiNet.
- Under two epochs per task, replay-free baselines such as BYOL, Barlow Twins and RM-SimSiam catch up to X-PhiNet, narrowing the advantage.
Limitations
- Relies on backpropagation, unlike biological learning; predictive coding without backprop is not yet available for ResNet-scale models.
- Model is feedforward with a single time step; recurrent structure for time series is future work.
- Learning dynamics analysis uses linear networks, isotropic Gaussian data and MSE loss rather than the cosine loss used in experiments.
Experimental setup
Datasets: CIFAR-5m CIFAR10 STL10 ImageNet Split CIFAR-5m Split CIFAR10 Split CIFAR100
Benchmarks: CIFAR10 STL10 ImageNet CIFAR-5m Split CIFAR-5m
Metrics: linear probing accuracy average accuracy forgetting
Baselines: SimSiam BYOL Barlow Twins RM-SimSiam CaSSLe
Builds on: SimSiam BYOL Complementary Learning Systems theory temporal prediction hypothesis
Closest papers by meaning
- Predictive Attractor Models (NeurIPS 2024 · similarity 0.86)
- Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks (ICML 2024 · similarity 0.86)
- A Stable, Fast, and Fully Automatic Learning Algorithm for Predictive Coding Networks (ICLR 2024 · similarity 0.85)
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.85)
- Learning Delays in Spiking Neural Networks using Dilated Convolutions with Learnable Spacings (ICLR 2024 · similarity 0.85)
- Divide-and-Conquer Predictive Coding: a structured Bayesian inference algorithm (NeurIPS 2024 · similarity 0.85)
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.