Abstract
Recurrent Neural Networks (RNNs) are favored for sequential data due to their ability to capture temporal relationships. However, their reliance on large datasets to learn these correlations limits their effectiveness in low-resource scenarios. This paper explores the k-FFNN, a Feed-forward Neural Network (FFNN) architecture that incorporates a priori temporal knowledge during training. By injecting known temporal patterns into training labels, k-FFNN achieves performance comparable to or better than RNNs in data-sparse environments. We evaluate this approach using the MediaEval 2016 "Emotional Impact of Movies" task, demonstrating that explicit knowledge infusion can effectively substitute for architectural complexity.
Introduction: The Data Hunger of Recurrence
In deep learning, sequential tasks like speech processing or video analysis typically demand Recurrent Neural Networks (RNNs) or Long Short-Term Memory (LSTM) units. These architectures utilize internal feedback loops to maintain a "memory" of previous inputs, allowing them to automatically discover temporal correlations.
While powerful, this automatic learning comes with a significant overhead: RNNs require vast amounts of data to tune their recurrent weights. In "low-resource" scenarios where data is scarce, RNNs often fail to converge on the underlying temporal structure. Conversely, standard Feed-Forward Neural Networks (FFNNs) are easier to train with limited data but are fundamentally "blind" to time, treating each input segment as an independent event. The k-FFNN addresses this gap by asking if a priori knowledge can be used to "teach" a simple FFNN about time.
Methodology: Knowledge Infusion via Label Shaping
The k-FFNN does not modify the network architecture itself. Instead, it alters the training targets to reflect known temporal dynamics.
- Segmentation: A long sequence (e.g., a 10-second audio clip) is divided into smaller 1-second segments.
- Knowledge Infusion: Instead of assigning the global sequence label to every segment, labels are modified by a function $f(i)$ representing a priori knowledge. For movie clips, emotions often follow a "fade-in/fade-out" pattern.
- Knowledge Functions: The authors tested several functions:
- Fn1 & Fn2: Smoothly scale labels at the start and end of a sequence to mimic natural emotional transitions.
- Fn3: A poorly-fitted control function to test the sensitivity of the model to incorrect assumptions.
- Training and Inference: A standard FFNN is trained on these shaped labels. During inference, the segment-level outputs are decoded back into a single sequence-level value using the inverse of the knowledge function.
Experimental Setup and Dataset
The study utilizes the MediaEval 2016 dataset, specifically the "Emotional Impact of Movies" task. This task requires predicting the shift in a viewer's emotional state—defined by Valence (positive/negative) and Arousal (exciting/calm)—after watching a short film clip.
The authors extracted 384 audio features using the openSMILE toolkit, eventually reducing them to 21 key features via WEKA for training.
Key Results and Insights
The k-FFNN was compared against standard FFNNs, RNNs, LSTMs, and Bidirectional LSTMs (BLSTMs) across varying training set sizes.
- Superiority in Low-Resource Scenarios: With only 200 training samples, k-FFNN (using Fn1) achieved a Mean Squared Error (MSE) of 0.927, significantly outperforming the standard RNN (0.977). Its Pearson Correlation Coefficient (PCC) was 0.16, nearly double that of the RNN (0.079).
- Convergence at Scale: As the training set size increased to the full 6814 samples, the performance gap closed. RNNs eventually learned the temporal patterns that k-FFNN was "told" at the start. This confirms that k-FFNN is a highly efficient alternative when data is limited.
- Sensitivity to Knowledge Quality: The success of k-FFNN is tied to the accuracy of the a priori knowledge. When the "incorrect" Fn3 function was used, performance dropped below that of a standard FFNN, highlighting the importance of domain expertise.
Conclusion
The k-FFNN demonstrates that explicit mathematical knowledge can serve as a powerful substitute for architectural complexity. By "baking" temporal expectations into the training process, researchers can deploy simpler, more efficient models that excel in data-constrained environments. Future work may explore learnable knowledge functions that can adapt to diverse data sequences without requiring manual tuning.
Summary updated