Abstract
Long Short-Term Memory (LSTM) networks are widely used for temporal data analysis, yet they often face challenges in effectively fusing information across extended sequences. This paper introduces a "stepped sampler" for LSTM training, grounded in the psychological principle of "repeated input" for memory reinforcement. By presenting overlapping sub-batches of video data, the proposed method enhances the model's ability to capture temporal correlations. Experimental results on the UCF101 action recognition dataset demonstrate that the stepped sampler achieves faster convergence, superior training stability, and higher classification accuracy compared to traditional sequential sampling methods.
Introduction
While LSTMs are designed to handle sequential dependencies, their performance is heavily influenced by how data is presented during training. Standard data loading techniques in modern frameworks typically treat sequences as discrete, non-overlapping chunks. This approach lacks the reinforcement mechanisms found in biological learning, such as "adjacent repetition" and "timely review," which are essential for long-term memory formation in humans. This research proposes a novel data-loading algorithm that translates these psychological principles into a computational strategy for video temporal detection.
Methodology
The primary innovation is the Stepped Sampler, which utilizes a "Wheel Tactic" to structure input data. Instead of moving through a sequence in disjoint blocks, the sampler creates overlapping sub-batches.
As shown in the figure above, the sampler defines a stride ($n$) between sub-batches of size $m$. This intentional redundancy ensures that the end of one sub-batch serves as the beginning of the next, reinforcing the temporal context. The authors justify this approach using Bayes' Theorem and Mutual Information to quantify frame-to-frame correlations.
The overall system architecture employs a four-layer CNN backbone for spatial feature extraction, which feeds into a three-layer LSTM network for temporal modeling.
Experimental Results
The method was evaluated on the UCF101 dataset, comprising over 13,000 videos across 101 action categories. The dataset's diversity in camera motion and lighting provides a rigorous test for temporal detection.
Performance and Stability
The stepped sampler demonstrated clear advantages in both accuracy and training dynamics:
- Accuracy Gains: With a batch size of 15 and a stride of 5, the stepped sampler reached a test accuracy of 0.637, outperforming the standard PyTorch batch sampler's 0.616.
- Convergence Speed: The proposed method reached a stable state in significantly fewer epochs.
- Training Stability: Traditional samplers often exhibit "jittery" loss curves. In contrast, the stepped sampler produced a much smoother convergence, with training loss stabilizing after epoch 60, whereas the baseline showed significant spikes through epoch 100.
Discussion and Conclusion
The study highlights that the "pedagogy" of data presentation is as critical as the network architecture itself. By finding a "Goldilocks" balance in the stride—mimicking human-like repetition—the model achieves more stable local minima.
However, the method introduces a trade-off: the intentional redundancy increases the total data processed per epoch. Future work should investigate the raw wall-clock time efficiency and the sensitivity of the stride hyperparameter across different datasets. Ultimately, incorporating biological memory effects into deep learning training pipelines represents a promising path toward more robust and efficient artificial intelligence.
Summary updated