Task Dependent Representation Selection

24 concepts17 papers1 proposal

In machine learning, we often chase the ideal of a "universal representation"—a single, all-encompassing vector space that perfectly captures the essence of a text, a molecule, an image, or a protein. The intuition behind foundation models suggests that if a representation is sufficiently rich, it should seamlessly generalize to any downstream task.

However, a cross-disciplinary wave of research demonstrates that a truly universal representation is an empirical mirage. Whether analyzing natural language, molecular structures, or contrastive visual features, the optimal way to represent data depends entirely on the specific downstream objective.

To understand why task-dependent representation selection is necessary, we must look at representations not as absolute descriptions of reality, but as strategic acts of compression.

The Core Intuition: Representation as Lossy Compression

Every representation must discard information to be useful. If a model retained every raw pixel, atom coordinate, or character token without abstraction, it would fail to generalize. A representation succeeds by filtering out "noise" and amplifying "signal."

The fundamental conflict arises because what constitutes noise for one task is often the vital signal for another.

Consider contrastive learning in computer vision. To learn unsupervised representations, models train on different "views" (mutated versions) of the same image. In What Makes for Good Views for Contrastive Learning?, Yonglong Tian et al. formalize this via the InfoMin principle: different views should share the minimal amount of mutual information necessary to solve the downstream task. If you are training a model to classify shapes, color is noise; an augmentation that strips color helps the model focus on geometry.

However, as Alex Tamkin et al. demonstrate in Feature Dropout: Revisiting the Role of Augmentations in Contrastive Learning, designing a universal augmentation is impossible. A transformation that is safely "label-preserving" for one task might completely destroy the features required for another. If the downstream task changes from identifying object shapes to identifying fine-grained flower species, stripping color removes the exact signal required for success.

The Chemical and Biological Divide

This tension is acutely visible in molecular machine learning and drug discovery, where molecules must be mapped to computational descriptors. As reviewed by Laurianne David et al. in Molecular representations in AI-driven drug discovery: a review and practical guide, there is no single representation that fits all chemical AI applications.

The divergence stems from the different physical scales and properties being predicted:

  • Geometry vs. Reactivity: In Machine learning accelerates pharmacophore-based virtual screening of MAO inhibitors, Marcin Cieślak et al. show that topological descriptors excel at predicting physics-based geometric docking scores, whereas substructural fingerprints are far more effective for predicting biological activity. One captures physical fit; the other captures chemical reactivity.
  • Inductive Biases: Agnieszka Wojtuch et al. demonstrate in Extended study on atomic featurization in graph neural networks for molecular property prediction that explicitly including "redundant" features like hydrogen counts provides a vital inductive bias that boosts solubility predictions, yet those same features add unhelpful noise when predicting other quantum or metabolic properties.
  • Objective-Driven Definitions: Even the fundamental units of chemistry—moieties (functional groups like alcohols or benzenes)—cannot be defined universally. Jonas Lederer et al. show in Automatic identification of chemical moieties that a model optimized for energy prediction identifies entirely different structural boundaries for moieties than a model optimized for geometric reconstruction.

Large-scale empirical benchmarks confirm this fragmentation. When Shuangjia Zheng et al. evaluated 16 state-of-the-art models across millions of experiments in MolRep: A Deep Representation Learning Library for Molecular Property Prediction, they found that no single representation consistently outperformed the others. Similar task-dependent trade-offs emerge when predicting drug combination synergy (Bulat Zagidullin et al. in Comparative analysis of molecular fingerprints in prediction of drug combination effects) and protein fitness landscapes (Peter Mørch Groth et al. in FLOP: Tasks for Fitness Landscapes Of Protein wildtypes), where complex structure-based embeddings are required for intricate fitness predictions, but simple descriptors suffice for basic activity levels.

Parallel Patterns in Language and Software

This phenomenon is not unique to physical sciences; it is deeply embedded in how models process language and code.

In syntax analysis, Lilja Øvrelid argues in Downstream use of syntactic analysis: does representation matter? that there is no single optimal way to represent sentence structure. Different downstream NLP applications prioritize entirely different aspects of linguistic information.

Similarly, when evaluating representations for technical software engineering data, Junda He et al. show in Representation Learning for Stack Overflow Posts: How Far are We? that models optimized to recommend tags fail to accurately predict post relatedness. Because the semantic requirements of these tasks conflict, practitioners must select or tune representations based on the specific target objective.

Engineering Around Task Dependency

If a single universal representation is mathematically and empirically impossible, how do we build robust, generalizable systems? Modern architectures handle this by shifting from fixed representations to adaptive selection frameworks.

1. Automated Hyperparameter Optimization

Instead of manually guessing which features matter, we can treat the choice of representation as a hyperparameter to be optimized alongside model weights. In Bayesian Optimization of Text Representations, Dani Yogatama et al. demonstrate that using Bayesian optimization to automatically select text representations (such as n-gram ranges and weighting schemes) allows simple logistic regression models to match or outperform massive, computationally expensive neural networks tailored to specific datasets.

2. Dynamic Multi-View Fusion

Rather than forcing a model to pick a single representation, frameworks can maintain multiple distinct "views" of the data and fuse them dynamically based on the task context.

[Raw Input Data]
       │
       ├─► View 1: Structural Geometry ──┐
       ├─► View 2: Task Context ─────────┼─► [Dynamic Fusion Layer] ─► Task-Specific Representation
       └─► View 3: Expert Rules ─────────┘            ▲
                                                      │
                                            (Downstream Objective)

Jiaxin Ju et al. implement this strategy in M^2LLM: Multi-view Molecular Representation Learning with Large Language Models. Their framework integrates structural embeddings, task contexts, and scientific rules. Crucially, the model alters the weights of these views depending on the objective: structural embeddings dominate for categorical classification, while expert rules and context are prioritized when predicting continuous physical properties like solubility.

3. Flexible Multi-Architecture Routing

When properties are fundamentally incompatible, unified platforms must act as routers rather than monolithic encoders. For instance, the peptide property platform developed by Yinuo Zhang et al. in PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction avoids enforcing a uniform model architecture, instead matching distinct peptide properties to the specific deep learning architectures best suited for them.

A similar routing philosophy underpins the interatomic potential library developed by Christoph Brunken et al. in Machine Learning Interatomic Potentials: library for efficient training, model development and simulation of molecular systems, which allows users to swap architectures depending on whether a simulation prioritizes energy accuracy or force accuracy.

Conclusion

The shift toward task-dependent representation selection represents a maturing phase in machine learning. Accepting that no single vector space can represent all facets of an object allows engineers and researchers to stop searching for the perfect universal encoder. Instead, the focus has turned to building flexible systems that can isolate, weight, and select the specific information required to solve the problem at hand.

Key Foundational Work

Bayesian Optimization of Text Representations (2015) — contributed 2 concept elements to this direction.
Evaluating Word Representation Features in Biomedical Named Entity Recognition Tasks (2014) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_task_dependent_representation_selection_80f914b8,
  title = {task-dependent representation selection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/task-dependent-representation-selection/rdr_80f914b87f3c4452ad9293117a40d401},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_task_dependent_representation_selection_80f914b8_v1,
  title = {task-dependent representation selection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/task-dependent-representation-selection/rdr_80f914b87f3c4452ad9293117a40d401/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake