Traditional machine learning workflows treat model evaluation and model fixing as two entirely separate stages. Explainable AI (XAI) and visualization tools typically act as one-way mirrors: they show you why a model made a mistake—such as a misclassified image or a biased recommendation—but leave you powerless to fix it without collecting new data or restarting an expensive training loop from scratch. As noted by Hammer et al. et al. in Preface: Intelligent interactive data visualization, there is a significant gap between static data representations and interfaces that allow human experts to actively guide and tune models.
To bridge this gap, a growing body of research focuses on interactive model steering. The core insight is simple yet powerful: turn explanations into control knobs. Instead of treating transparency as a passive output, systems can use visual interfaces to establish a bidirectional flow of information, allowing human operators to directly manipulate internal model parameters, attention weights, or optimization spaces in real-time.
The Mechanism of Semantic Interaction
How do you map a human's high-level conceptual intent to low-level mathematical adjustments? This is the core challenge of semantic interaction. Instead of forcing a user to manually tweak abstract hyperparameters like regularization constants or kernel widths, the system translates direct manipulations of data or visual explanations into algorithmic updates.
Consider the task of dimensionality reduction. In a typical setup, an algorithm projects high-dimensional data onto a 2D scatterplot, and the user passively observes the clusters. In contrast, Tefas et al. demonstrate a system where users can interactively drag data points across the screen to define target positions based on domain knowledge. The system treats these movements as a supervision signal, applying gradient descent to adjust the underlying projection parameters so that local user interactions result in a globally consistent reorganization of the dataset. Similarly, Huang et al. integrate user-defined pairwise constraints (such as "must-link" or "cannot-link" relationships) directly into the training process as additional loss terms, forcing the visualization to adapt to human intuition.
[Passive XAI] Model ───────> Static Explanation ───────> User
[Interactive] Model <─────── Semantic Interaction <───── User
(Gradients, Loss Terms, Priors)
This translation mechanism can also be applied to error distributions. In Performance and Preferences: Interactive Refinement of Machine Learning Procedures, Kapoor et al. et al. allow users to interact directly with a confusion matrix. If a model is making too many false-positive errors in a critical category, the user simply clicks an arrow to demand fewer errors. The interface mathematically translates this preference into an optimized adjustment of multiple kernel weights, steering the classifier's decision boundaries without requiring the user to understand the underlying optimization math.
Closing the Loop via Attention and Rules
Another prominent strategy for interactive steering involves manipulating the intermediate representations that drive model decisions, such as attention maps or rule-based surrogates.
While attention mechanisms provide transparency by highlighting which features a model considers important, they are typically passive outputs. Heo et al. introduce Interactive Attention Learning, a framework where users manually adjust attention weights on-the-fly to correct a model's focus. By utilizing Neural Attention Processes, the model incorporates this feedback instantly without expensive retraining. In the context of computer vision, Zhang et al. allow users to provide visual "prompts"—simple sketches indicating which image regions to prioritize or ignore—to steer a model's attention in high-stakes scenarios like medical imaging.
User edits human-readable rules/prototypes
│
▼
[Surrogate Interface] ──(Synthetic Data / Fine-Tuning)──> [Black-Box Model]
▲ │
└─────────────────── Explains current state ───────────────┘
For complex black-box models, surrogate representations can serve as the control interface. In AIMEE: Interactive model maintenance with rule-based surrogates, Cornec et al. extract human-interpretable rules from a black-box model. When business requirements or regulations change, users edit these rules directly. The system then generates synthetic data reflecting the updated rules and retrains the black-box model to adopt the new decision boundaries. A parallel approach is taken by Ming et al. for sequence learning: domain experts manually add, remove, or modify a set of human-readable "prototype" sequences, and the model is fine-tuned to align its predictive logic with human intuition.
Steering the Optimization Landscape
Interactive steering is particularly effective when applied to automated search processes, such as Hyperparameter Optimization (HPO) and Neural Architecture Search (NAS). Standard automated tools often operate as "one-shot" black boxes, which can alienate practitioners when they observe sub-optimal trends early in a run, as documented by Xin et al..
By introducing interactive steering, systems like HyperTendril allow users to pause an automated optimization run, refine hyperparameter ranges based on real-time visual trends, and restart the search. To do this dynamically without losing progress, Fehring et al. introduce dynamic priors into Bayesian Optimization. This allows human experts to inject emerging observations into an ongoing search, shifting the algorithm's focus toward high-potential regions of the configuration space.
This collaboration fundamentally alters optimization dynamics. Rather than acting as passive oracles, humans often act strategically. Colella et al. demonstrated that when humans interact with a Bayesian Optimization loop, they intentionally provide biased feedback to exploit the algorithm's logic, guiding it to find global maxima significantly faster than a fully automated search.
Why It Works: The Rashomon Effect
The mathematical justification for why interactive steering can succeed without destroying a model's predictive accuracy relies heavily on the Rashomon Effect.
As explained by Liu et al., for many complex datasets, there exists a massive space of near-optimal models that achieve virtually identical predictive performance but rely on completely different internal logic or combinations of variables.
Predictive Loss Landscape
│
│ ┌─► Model A (Accurate, but uses biased features)
│ │
└───┼─► Model B (Accurate, and aligns with human intuition) ◄── Selected via Steering
│
└─► Model C (Accurate, but too complex)
Because an automated algorithm is indifferent to which model it selects within this optimal bracket, it may easily choose a model that relies on spurious correlations. Interactive steering allows a domain expert to navigate this "Rashomon set" visually. By injecting post-hoc constraints, adjusting weights, or deferring rigid parameter selections to an interactive deployment phase—such as the runtime conservatism slider designed for reinforcement learning in User-Interactive Offline Reinforcement Learning—the user guides the system toward a model that is both highly accurate and aligned with human intent.
Key Foundational Work
Human-Guided Simple Search (2000) — contributed 1 concept elements to this direction.
Human-Guided Tabu Search (2002) — contributed 1 concept elements to this direction.