Introduction
In machine learning, inductive bias refers to the set of assumptions a learner uses to predict outputs for inputs it has never seen. While sequence-to-sequence (seq2seq) models like LSTMs and Transformers have revolutionized natural language processing, they are often criticized for requiring massive datasets and failing to generalize systematically. This paper investigates the "hidden personality" of these models by observing how they generalize from ambiguous, low-data scenarios—a concept known as the "poverty of the stimulus." By putting models in a state of "doubt," the authors reveal which explanations they naturally prefer: simple memorization, linear patterns, or complex hierarchical structures.
Methodology
The researchers compared three standard architectures: LSTM-based, CNN-based, and Transformer seq2seq models. To measure their biases beyond simple accuracy, they proposed a more sensitive metric:
- Description Length (DL): Rooted in Solomonoff’s theory of induction and Minimal Description Length (MDL), DL measures how "simple" a specific rule is for a model to represent. If a model can learn a rule quickly with few bits of information, that rule aligns with its inductive bias.
- Prequential Coding: This method calculates DL by training the model incrementally and measuring how well it predicts the next unseen example.
- Fraction of Perfect Agreement (FPA): A secondary metric measuring how often a model's predictions perfectly match a specific candidate rule (e.g., "always output the same length" vs. "count the input").
Experimental Setup
The authors introduced four novel synthetic tasks designed to be highly ambiguous, forcing the models to rely on their internal biases.
- Count-or-Memorization: Given $a^l \to b^l$, does the model generalize to $a^m \to b^m$ (counting) or $a^m \to b^l$ (memorizing the specific output length)?
- Add-or-Multiply: Given $a^l \to b^{2l}$, does the model see this as $x+l$, $2x$, or a constant mapping?
- Hierarchical-or-Linear: Models are trained on nested structures. Does the model pick the middle symbol (hierarchical) or a symbol at a fixed position (linear)?
- Composition-or-Memorization: Based on the SCAN benchmark, this tests if a model can combine known primitives (e.g., "thrice") with new symbols or if it just memorizes the training pairs.
Results and Discussion
The experiments revealed striking differences between architectures.
LSTMs and Arithmetic
Surprisingly, LSTM-based models can learn non-trivial arithmetic rules—like counting, addition, and multiplication—from as little as one single training example. This suggests their recurrent nature provides a strong inherent bias toward sequential counting. As shown in the results for the Count-or-Memorization task, LSTMs achieve high agreement with the counting rule at longer sequence lengths.
Architecture and Logic
There is a sharp divide in how models handle structure. Transformers and LSTMs show a strong bias toward hierarchical reasoning, which is crucial for human language syntax. In the Hierarchical-or-Linear task, Transformers achieved an FPA of 0.69 for the hierarchical rule. Conversely, CNNs are "linear thinkers," preferring rules based on absolute positions, achieving an FPA of 1.00 for the linear rule.
The Transformer's Memorization Trap
In compositional tasks, Transformers showed a massive, stubborn bias toward memorization. Even when provided with many examples, they struggled to "compose" rules compared to CNNs and LSTMs. While accuracy (FPA) is often "all or nothing," Description Length proved to be a much more sensitive measure, revealing that LSTMs were starting to "lean" toward composition while Transformers remained firmly biased toward memorization.
Conclusion
This study demonstrates that architecture dictates logic. While these results are based on synthetic, highly controlled tasks, they reveal "pure" architectural biases that may be overwhelmed by noise in real-world data. However, the findings suggest that hyperparameters like dropout can shift these preferences. Future work could involve using these diagnostic tasks to "pre-screen" new architectures for specific real-world applications, ensuring they possess the appropriate inductive biases before large-scale training.
Summary updated