Source-linked AI summary
Why shared attention vectors fail: a case for outcome-indexed tuning
Lenard Dome
TL;DR
Globally shared attention vectors can fail in multi-outcome learning because summing outcome-specific gradients can destabilize or erase feature saliences. The paper analyzes this failure and introduces outcome-indexed attention weights; it reports that this representation supports outcome-specific feature importance and is structurally incapable of the described instability. The authors retain the core gradient-descent principles and note that their architecture is indistinguishable from the Mackintosh tradition when K = 1.
Problem
Method
Results
Outcome-indexed attention weights support outcome-specific feature importance, and the attention matrix is structurally incapable of the described shared-salience instability.
Source: §6 · pp. 6–7Takeaways & Limitations
The architecture extends gradient-based attentional learning while remaining indistinguishable from the Mackintosh tradition when K = 1.
Source: §6 · p. 7The authors do not present evidence that prior models are inadequate, and the matrix becomes indistinguishable from the Mackintosh tradition under K = 1.
Source: §6 · p. 7
Abstract
from arXiv · showhide
Dimensional attention in learning is often implemented as a globally shared attention vector, where each stimulus dimension corresponds to a single scalar. These scalars are learned by models through gradient-descent on error, where predictive features acquire more salience. We show that under multi-outcome learning, where models predict more than one outcome, this shared vector becomes unstable; it collapses to its bounds and prevents the models from learning meaningful attentional tunings for learning and generalization. We address this by introducing an outcome-indexed attentional matrix that converts globally shared attentional tuning into an outcome-indexed representation. We present an analysis of the unstable shared vectors and derive the conditions under which it holds. Empirically, three synthetic experiments benchmark the proposed attention matrices and show that they converge to meaningful representations, something shared attention vectors fail to do. These results suggest that outcome-indexed attentional matrices are a general fix for gradient-based attentional processes, which improves models of learning under multi-outcome conditions.
1 Introduction
The introduction frames attention as selecting diagnostic stimulus information and describes a tradition of models that learn feature salience through prediction. It argues that globally shared attention becomes unstable in multi-outcome learning and proposes outcome-indexed weights as a solution.
1 Introduction
- Attention directs learning toward diagnostic information while filtering irrelevant stimulus features.Selective attention to stimulus features has also been proposed as a formal requirement for categorization.Source: §1 · p. 1
- Mackintosh (1975) formalized attention as feature-specific salience acquired through predictive success, a framework later models adjust by gradient descent on error.These models store saliences in a globally shared attention vector, with each feature's salience constrained to [0, ∞).Source: §1 · p. 1
- Under multi-outcome learning, summing gradients on a shared vector can make its weights unstable and prevent models from learning what features to attend.The introduction identifies summation across outcomes as the primary cause, even when a feature is diagnostic for one outcome but not others.Source: §1 · p. 1
- The paper proposes outcome-indexed attention weights to represent richer attentional mappings and improve models’ capacity to account for complex real-world behavior.The proposed change separates attentional tuning by outcome rather than retaining a globally shared vector.Source: §1 · p. 2
2 Preliminaries
The preliminaries describe models that combine binary feature inputs with salience-derived attention gains to predict multiple outcomes. Saliences are adjusted through gradient-based learning, with non-negativity constraints and iterative attention settling.
2 Preliminaries
- Each stimulus is a binary feature vector, and each feature has a non-negative scalar salience used to generate attention gains.The model combines stimulus presence-or-absence inputs with the corresponding salience values at the start of each trial.Source: §2 · p. 2
- Normalized attention strengths are combined with connection weights to produce predictions across K outcomes.The p-norm parameter controls the degree of competition among currently present input dimensions.Source: §2 · p. 2
- After feedback, gradient descent increases salience for predictive stimuli and decreases it for unpredictive stimuli, while a hard projection keeps salience non-negative.The update uses a clamp to enforce the non-negativity constraint.Source: §2 · p. 2
- Attention settling is nonlinear, so the model reiterates Equation 6 ten times, recalculating intermediate values and clamping updated attention gains to [0, +∞).The final salience update uses state displacement, with α as its learning rate.Source: §2 · p. 2
3 Problem Statement
The problem statement explains how summing outcome-specific gradients onto one shared attention state can erase selective information or drive saliences to their boundary. Under the stated multi-outcome conditions, the magnitude of attention updates grows with the number of active outcomes.
3 Problem Statement
- For models satisfying the stated conditions, attention-update magnitude scales as O(K) when more than one outcome is simultaneously active.The conditions include additive loss across outcomes, gradient descent on attention, hard non-negativity, same-sign reinforcement for non-predicted outcomes, and multi-outcome activation.Source: §3 · p. 3
- When a feature predicts only a minority of co-active outcomes, gradients from the non-predicting majority can drive its shared salience downward until the clamp resets it to zero.The suppressing group's magnitude grows with the number of outcomes the feature does not inform, so selectively informative cues can be punished.Source: §3 · p. 3
- The boundary-reset mechanism has a distinct single-outcome case: at K = 1, dropping a feature discards a genuinely uninformative feature.The passage contrasts this with multi-outcome learning, where a cue may be punished for being informative about only some outcomes.Source: §3 · p. 3
- Opposing outcome-specific gradients can cancel in the aggregated update, leaving a shared salience frozen despite large individual demands.This cancellation can occur even when the non-negativity boundary is never approached.Source: §3 · p. 3
- The failure is not limited to hard non-negativity: collapsing signed, outcome-specific gradients onto one shared scalar can erase the attention a feature needs for different outcomes.The shared state aggregates the loss across outcomes, which can combine conflicting update demands.Source: §3 · pp. 2–3
4 Architectural Fix
Outcome-indexed attention treats a cue’s predictiveness as specific to an outcome and cue configuration, rather than globally fixed. The model replaces shared salience with outcome-specific weights and updates each outcome’s weights using its own error signal.
4 Architectural Fix
- A cue can predict some co-active outcomes but be uninformative for others, so its attentional value depends on the outcome and cue configuration.The paper therefore indexes attention by the connection it modulates, rather than assigning each cue one globally shared scalar.Source: §4 · p. 3
- The architecture replaces each shared state η_i with an outcome-indexed state η_ki and normalizes attention gains separately for each outcome.Each outcome has its own p-norm normalization.Source: §4 · p. 3
- Each attention-matrix row is updated independently, removing the sum over outcomes so its gradient magnitude no longer scales with the number of active outcomes.The loss and attention shift are defined without summing gradients across outcomes.Source: §4 · pp. 3–4
5 Simulations
Across three simulations of increasing input-output complexity, shared attention vectors collapsed while outcome-indexed matrices retained structured, outcome-specific attention. In the overlapping-outcome case, the matrix preserved opposing demands that the shared salience could not represent.
5 Simulations
- The simulations pretrained nonzero input-output weights, then compared shared vectors and outcome-indexed matrices across a step-size sweep using boundary hits as the stability measure.The three stimulus sets procedurally increased mapping complexity; Table 1 pairs each stimulus with its teaching vector.Source: §5 · p. 4, §5.1 · p. 4, Table 1 · p. 5
5.1 Distinctive Singular: minimal case with no conflict
- The shared vector reset all feature scalars to the floor, whereas the matrix retained nonzero weights that encoded differential informativeness.The matrix elevated 8 of 12 feature-outcome mappings and left the rest at zero.Source: §5.1 · p. 4
5.2 Distinctive Multi-Outcome: no outcome overlap with multiple excited output units
- With five disjoint outcomes, the shared vector again collapsed to the floor, while attention matrices retained robust representations of where features were informative.Stimulus A’s features retained informativeness longer at smaller step sizes than in the prior simulation.Source: §5.2 · pp. 4–5, Figure 1 · p. 5
5.3 Shared Outcome Spaces: overlapping co-active outcomes
- In the shared-outcome case, opposing within-trial demands leave shared salience with no net updates, while matrix rows store those demands separately by outcome.The stimulus sets include outcomes shared across teaching vectors and cues with opposite-sign demands within a trial.Source: §5.3 · pp. 5–6, Figure 2 · p. 6
6 Summary
The simulations identify multi-outcome gradient summation as a source of instability in shared attention vectors and show that outcome-indexed matrices remove this failure mode. The matrix also supports outcome-specific feature importance, with implications for settings where inputs predict multiple outcomes.
6 Summary
- Across three simulations, collapse occurred throughout the explored range of ρ > 0, sign-cancellation was independent of non-negativity constraints, and decoupling gradients by outcome removed the failure modes.The proposed matrix is applied independently per outcome k for η ∈ R^K×J.Source: §6 · p. 6
- Features can carry different importance for different outcomes, so attentional reallocation depends on the outcome being predicted rather than cue-specific associability alone.Features may diagnose some outcomes but not others, and their information is used depending on which outcomes are excitable.Source: §6 · p. 6
- When outcomes are shared between inputs, features connect to multiple output nodes, a setting in which scalar salience breaks down but the attention matrix is structurally incapable of that instability.The cited simulations use shared outcomes between inputs, with implications for probabilistic feature-outcome mappings.Source: §6 · pp. 6–7
- The attention-to-connection-weight mapping amplified inhibitory connections in almost all simulations, while excitatory connections were less likely to gain salience.With close-to-zero initial weights, features could become predictive of outcome absence early; the authors note that initialization is a choice and alternatives exist.Source: §6 · p. 7
- The proposed architecture preserves gradient-descent attention principles and is indistinguishable from the Mackintosh tradition when K = 1; the paper does not show that prior models are inadequate.The authors frame the change as extending models to more challenging environments, where multiple outcomes may be anticipated.Source: §6 · p. 7
7 Conclusion
Globally shared salience vectors become unstable under multi-outcome learning because outcome-indexed gradients are summed onto one shared state. Indexing attention by outcome restores stable learning and represents feature informativeness specific to the outcome and its configuration.
7 Conclusion
- Shared salience vectors become unstable in multi-outcome learning because gradients indexed by outcome are summed onto one shared state.The instability persists across step-sizes and is not visible in single-outcome paradigms.Source: §7 · p. 7
- Indexing attention by outcome restores stable learning with a relatively small architectural change.The paper identifies this change as a precondition for extending attentional learning to richer environments.Source: §7 · p. 7
- Outcome-indexed weights represent feature informativeness specific to the outcome and the configuration in which the feature appears.Source: §7 · p. 7
A Attention shift equations
The appendix introduces attention-shift definitions for globally shared salience vectors and outcome-indexed attention matrices. The shared-vector update uses a positive step size and a tanh function to constrain updates.
A Attention shift equations
- The shared-vector attention shift uses ρ, a positive gradient-descent step size called the attention shift rate.Source: A Attention shift equations · p. 8
- The tanh function further constrains updates to lie between -1 and 1.Source: A Attention shift equations · p. 8
- The appendix presents separate attention-shift definitions for globally shared salience vectors and outcome-indexed attention matrices.Source: A Attention shift equations · p. 8