Human intelligence is characterized by the ability to combine known concepts in novel ways. If you know what "jump" means and you know what "twice" means, you immediately understand the command "jump twice," even if you have never heard that specific combination before. This ability is known as systematic compositionality.
While modern deep learning models excel at pattern matching, they often lack this "algebraic" capacity. As Brenden M. Lake and Marco Baroni demonstrated in Still not systematic after all these years, standard recurrent neural networks (RNNs) fail catastrophically when asked to apply a learned rule to a new primitive. This "compositional gap" suggests that neural networks often memorize specific training patterns rather than learning the underlying logical rules.
The Problem of Entanglement
The primary reason standard architectures fail to generalize is semantic entanglement. In a typical sequence-to-sequence model, the representation of a word is often "polluted" by its neighbors. Hao Zheng argues in DANGLE that models fail because they don't treat components as independent units; instead, the internal representation of "jump" changes depending on whether it is followed by "left" or "twice."
When representations are entangled, the model cannot perform "zero-shot" mapping. It treats every unique combination as a distinct task rather than a composition of known building blocks. This leads to a lack of functional reuse, where the network allocates independent weights for identical operations in different contexts rather than using a shared, modular rule.
Architectural Interventions
To bridge this gap, researchers have proposed architectures that explicitly separate the "what" from the "how."
- Syntactic and Semantic Separation: Jacob Russin and Yoshua Bengio introduced Syntactic Attention, which mimics the human brain by processing syntax and semantics in separate streams. By forcing the model to learn the structure of a command (the syntax) independently of the specific actions (the semantics), the model can apply a "twice" rule to any action it has seen, regardless of whether they appeared together in training.
- Modular Sub-networks: Instead of a monolithic block of weights, researchers like Simon Schug explore modular solutions that discover independent, reusable components. These architectures, such as the Multi-Task Recurrent Modular Network by Dongkuan Xu, dynamically route data through shared modules, ensuring that a learned function is available for any new task configuration.
- Relational Inductive Biases: Models like Hidden Schema Networks by Ramsés J. Sánchez impose explicit relational structures—such as latent graphs—onto model outputs. This forces the model to interpret sentences as paths on a graph, ensuring that the relationship between entities remains consistent even when the entities themselves change.
Compositionality as a Learned Skill
A more recent perspective suggests that systematicity is not just an architectural property, but a skill that can be trained. Marco Baroni introduced Meta-learning for Compositionality (MLC), a training regime where the model is explicitly rewarded for its ability to compose.
Instead of just learning to predict the next word, the model is trained on a series of "episodes" that require it to apply new rules to new primitives on the fly. This "Incentive and Practice" approach allows standard Transformers to achieve human-like systematicity, suggesting that the failure of earlier models was due to a lack of compositional training data rather than an inherent flaw in the connectionist architecture.
The Scaling Debate
The rise of Large Language Models (LLMs) has complicated this debate. While Anders Søgaard and Ruchira Dhar find that scaling up model size improves compositional abilities, they also note that "alignment" techniques like instruction tuning can actually degrade them.
Furthermore, even state-of-the-art models like DALL-E 2 still struggle with basic relational logic. As Tomer D. Ullman shows in Testing Relational Understanding, these models often rely on statistical "gluing" of objects rather than a true understanding of relationships like "X under Y." This suggests that while scaling provides a better approximation of compositionality, achieving human-level systematicity may still require the explicit structural interventions and meta-learning strategies currently being explored in the field.