NOTE-250: MIMI: the objective that needs the knowledge it assumes away

Contribution

An answer to a genuinely hard problem: how do you fit a human-machine interface when there is no prior mapping, no action labels, no reward signal, and no list of what the person is trying to do? The proposal is that intuitiveness has a task-free signature — a good interface makes a person’s commands informative about what happens — so the mutual information between command and induced state transition can stand in for a task metric nobody can observe.

Two uses: ranking existing interfaces offline, and maximizing the score directly by human-in-the-loop RL to build one from scratch.

Key insight

You cannot see what the user wants, but you can see whether their inputs explain the outcomes. A channel the operator cannot use produces commands uncorrelated with what follows, whatever they were for. That makes interface quality measurable without ever naming the task — and it makes what counts as “what follows” the whole question, which is where the paper’s most useful result lives.

Assumptions

• The interface can observe the state s_t. The paper notes this fails when the user can see things the interface cannot (they look around; it has no camera) and calls that an extension.
• A parameterized interface family small enough for Bayesian-optimization RL: the largest tested has 8 parameters.
• The user adapts too — this is co-adaptation, and the objective is estimated on a moving target.
• A horizon Δ is fixed before scoring.

Key results

• Offline validation. 540K examples of keyboard and eye-gaze interfaces for typing, simulated robot control and games. In 4 of 5 domains the MI score is predictive of ground-truth task reward, average Spearman ρ = 0.43. The X2T eye-gaze arm alone gives ρ = 0.45 over 36 points (12 users × 3 conditions).
• The inversion. In the shared-autonomy Lunar Lander (SAvDRL) data, the one-step objective gave “a strong negative correlation” with reward. Cause, in the authors’ words: the assistant prevents crashes, “which necessarily involves ignoring a large fraction of the user’s commands” — lowering I(x_t, (s_t, s_{t+1})) while increasing the user’s influence over later states. Generalizing to I(x_t, (s_t, s_{t+Δ})) and setting Δ to the maximum episode length (10³) restores a positive correlation.
• The admission. “Choosing the value of Δ requires prior knowledge of the timescale of the user’s desired influence over the system… This is the primary limitation on the generality of our method.” Repeated in §6.1.
• Online learning. 12 participants, 2D cursor control through a perturbed mouse. MIMI substantially beats randomly parameterized interfaces and approaches an oracle that moves straight to the target, in under 30 minutes. MI reward and true task reward correlate at ρ = 0.87.
• What it converges to. The learned perturbation angle θ over 12 users is sharply bimodal: θ ≈ 0 (no perturbation) and θ ≈ π (exact inversion), against a uniform random baseline. Trajectories go from curved and wandering to straight after 150 episodes.
• Lunar Lander by hand gesture, one expert user: the learned interface works, and “does not enable the user to fire the left thruster in isolation” — every gesture that fires it also fires the main engine.

Claims

id	claim	strength	support
C1	MI between command and induced transition ranks interfaces without task knowledge	moderate	540K examples, five domains — four of which worked, at ρ = 0.43
C2	The objective can be maximized to learn an interface from scratch	moderate	12-participant study, 8-parameter interface, one task family
C3	The correlation’s sign depends on the horizon Δ	strong	the SAvDRL sweep, with a mechanism that explains it
C4	Choosing Δ needs task knowledge	strong	stated twice by the authors as the primary limitation
C5	The objective is “completely unsupervised”	weak	true except for Δ, and Δ is where the task knowledge went
C6	It scales beyond 8 parameters	—	not claimed; named as future work

Method

Estimate I(x_t, (s_t, s_{t+Δ})) from logged interaction. To learn: randomly initialize the interface parameters, let the user attempt their own tasks, estimate the score over the resulting episodes, and update the parameters with Bayesian-optimization reinforcement learning against that score as reward. No labels, no reward feedback, no task list — and Δ set in advance.

Concepts

• MIMI — the method: mutual information maximizing interfaces.
• generalized MI objective — I(x_t, (s_t, s_{t+Δ})), the one-step version at Δ = 1. The generalization is presented as a fix and is better read as the discovery.
• co-adaptation — the user learns the interface while the interface learns the user. It is why the bimodal convergence to inversion is not simply a failure: users adapt to a consistent mapping.
• infomax — the older principle the authors connect to in future work.

Connections

The nearest prior work the paper names is PCA-based unsupervised co-adaptation, which requires a linear interface and does not optimize intuitiveness. The record holds neither that line nor anything else on assistive interfaces, so no relation is declared and none is missing.

Recommendations

• R1 — score a control interface by command-to-outcome mutual information when labels and rewards are unavailable, and choose the horizon deliberately. Filed as SOTA-309, with the horizon in the title because the sign depends on it.
• R2 — do not read a one-step influence metric as a measure of how well an assistant is helping. Folded into R1’s Conditions rather than filed separately: it is the same measurement fact stated from the assistant’s side rather than the interface’s, and one source should not produce two practices about one table.

Bearing on the record

The horizon result joins the four measurement traps the record picked up today: SOTA-305 (reconstruction FID needs its rate), SOTA-307 (generation FID needs seeds), SOTA-308 (recall bought with relevance), and this. It is the sharpest of the set, because the measurement choice flips the sign of the correlation rather than shifting its magnitude, and because the flip happens exactly when the system under test is working — an assistant that successfully prevents a crash looks worst.

Four instances now, from four unrelated fields, all of the form “the reported number is sensitive to a choice the report does not state”. I am recording the count and not writing the principle: DP-009, and the four are not obviously one mechanism.

Limitations

• Four of five domains, and the abstract says “a variety of domains” with the average taken over the four.
• Δ is the load-bearing hyperparameter and its correct value is task knowledge. The paper is admirably direct about this; the abstract’s “completely unsupervised” is not.
• ρ = 0.43 is moderate. Adequate for ranking, not for close calls. The 0.87 online figure comes from an easier setting.
• Eight parameters. No evidence the objective survives a high-dimensional interface, and the data efficiency of the RL is the binding constraint.
• Small n everywhere: 12 participants in the study, 12 per offline evaluation, and a single expert user for the gesture experiment.
• Consistency versus intuitiveness. The convergence to θ ≈ π shows the objective rewarding a mapping the user can learn rather than one they find natural. That may be the right target — the paper does not distinguish them.

Open questions

• How wrong can Δ be? The measurement that would move this from Proposed: correlation against a held-out task metric as a curve over Δ, in a second setting, so somebody picking Δ blind can see the cost. The paper has this curve for SAvDRL (Fig. 2d) and for one domain only.
• Does it survive a high-dimensional interface? Named as future work, with specific algorithms suggested. Until then the method is demonstrated on 8 parameters.
• Is the bimodal convergence a feature? If the objective reliably finds some consistent mapping rather than the natural one, then it is a learnability score and should be described as one — which would be a stronger claim than the paper makes and a different one.