[Paper Review] Optimal allocation of attentional resource to multiple items with unequal relevance

Nuwan de Silva, Wei Ji|arXiv (Cornell University)|Feb 18, 2018

Cognitive Science and Mapping3 citations

TL;DR

This paper proposes a normative model for optimal attentional resource allocation across multiple items with unequal probing probabilities. Using a concave utility function and a dimensionality-reduction algorithm, it shows that optimal allocation often deviates from proportional sharing—sometimes assigning zero attention to non-negligible targets when resources are scarce—offering a principled framework for studying attentional control in perception.

ABSTRACT

In natural perception, different items (objects) in a scene are rarely equally relevant to the observer. The brain improves performance by directing attention to the most relevant items, for example the ones most likely to be probed. For a general set of probing probabilities, it is not known how attentional resources should be allocated to maximize performance. Here, we investigate the optimal strategy for allocating a fixed resource budget E among N items when on each trial, only one item gets probed. We develop an efficient algorithm that, for any concave utility function, reduces the N-dimensional problem to a set of N one-dimensional problems that the brain could plausibly solve. We find that the intuitive strategy of allocating resource in proportion to the probing probabilities is in general not optimal. In particular, in some tasks, if resource is low, the optimal strategy involves allocating zero resource to items with a nonzero probability of being probed. Our work opens the door to normatively guided studies of attentional allocation.

Motivation & Objective

• To determine the optimal distribution of limited attentional resources across multiple items when their relevance (probing probability) varies.
• To develop a computationally feasible method for solving the high-dimensional optimization problem of attention allocation.
• To challenge the assumption that attention should be allocated proportionally to relevance.
• To provide a normative benchmark for evaluating biological attentional mechanisms in perception.

Proposed method

• Formalizes attention allocation as a constrained optimization problem maximizing expected utility under a fixed resource budget E.
• Uses a concave utility function to model diminishing returns of attention on perceptual performance.
• Applies a mathematical transformation to reduce the N-dimensional optimization problem into N one-dimensional subproblems.
• Employs a Lagrangian relaxation technique to derive necessary conditions for optimality.
• Derives a fixed-point iteration algorithm that the brain could plausibly implement for real-time resource allocation.
• Validates the algorithm's efficiency and optimality through numerical simulations across various probing probability distributions.

Experimental results

Research questions

• RQ1 What is the optimal allocation of attentional resources when items have unequal probabilities of being probed?
• RQ2 Can attention be optimally allocated even when some items with non-zero probing probability receive zero attention?
• RQ3 How does the optimal allocation strategy change as a function of total resource budget and number of items?
• RQ4 Is proportional allocation of attention always optimal under general utility functions?
• RQ5 Can the brain implement optimal attention allocation using a biologically plausible algorithm?

Key findings

• Optimal attention allocation is not proportional to probing probabilities; in fact, it can assign zero attention to items with non-zero probing probability when resources are limited.
• The optimal strategy involves concentrating attention on a subset of the most relevant items, especially when the resource budget is low.
• The proposed algorithm reduces the N-dimensional optimization problem to N one-dimensional problems, enabling efficient computation.
• For low resource budgets, the optimal solution may involve allocating zero attention to items with moderate or high relevance, depending on the distribution.
• The model demonstrates that diminishing returns (concave utility) lead to non-uniform, sparse allocation strategies that outperform proportional allocation.
• The framework provides a normative basis for testing whether biological attentional systems approximate optimal behavior.