Tag
This paper shows that predictive coding networks compute the same gradients as backpropagation in the limit of width much larger than depth, bridging biological learning and standard neural network training.
This paper investigates the developmental conditions under which a minimal predictive neural system (a 192-dimensional GRU) can distinguish self-caused changes from world-caused changes, identifying four necessary conditions for agency and introducing a metric called agency gain.
This paper tracks how different learning rules (backprop, feedback alignment, predictive coding, STDP) affect the alignment of CNN representations with human fMRI across training. It finds that backprop destroys V1 alignment in one epoch, while local rules preserve it, suggesting a trade-off between building higher-level representations and retaining early visual features.
This paper tracks how supervised training with different learning rules (backpropagation, feedback alignment, predictive coding, STDP) degrades alignment between neural network representations and early visual cortex fMRI data, finding that untrained networks often match or exceed trained ones in V1 alignment.
AdaCodec reduces video encoding redundancy in multimodal LLMs by transmitting full visual tokens only when scene prediction fails, otherwise using compact inter-frame change descriptions. It outperforms per-frame RGB baselines at matched token budgets and achieves better or comparable results with significantly fewer tokens, reducing time-to-first-token from 9.26s to 1.62s.
The paper introduces closed-form predictive coding via hierarchical Gaussian filters that restore precision-weighted prediction errors, yielding faster and more efficient training without global error signals, outperforming backpropagation on certain tasks.
Swift Sampling is a training-free algorithm that uses Taylor expansion to identify high-information moments in long-form videos by detecting deviations from predicted feature trajectories, improving accuracy on video QA tasks with minimal computational overhead.
Explores how close a biologically plausible Hebbian agent can get to PPO on Pong, finding only a 2% gap but identifying catastrophic forgetting under self-play as a key bottleneck.
RoboMemArena introduces a large-scale benchmark for evaluating robotic memory across 26 complex tasks with real-world validation, alongside PrediMem, a dual-system vision-language-action model that improves memory management through predictive coding.