Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AI Model Modulation with Logits Redistribution

Zihan Wang, Zhongkui Ma, Xinguo Feng, Zhiyang Mei, Ethan Ma, Derui Wang, Minhui Xue, Guangdong Bai

Abstract

Large-scale models are typically adapted to meet the diverse requirements of model owners and users. However, maintaining multiple specialized versions of the model is inefficient. In response, we propose AIM, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements. AIM enables two key modulation modes: utility and focus modulations. The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model's focused input features. AIM introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner. We establish a formal foundation to ensure AIM's regulation capability, based on the statistical properties of logits ordering via joint probability distributions. Our evaluation confirms AIM's practicality and versatility for Al model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.

AI Model Modulation with Logits Redistribution

Abstract

Large-scale models are typically adapted to meet the diverse requirements of model owners and users. However, maintaining multiple specialized versions of the model is inefficient. In response, we propose AIM, a novel model modulation paradigm that enables a single model to exhibit diverse behaviors to meet the specific end requirements. AIM enables two key modulation modes: utility and focus modulations. The former provides model owners with dynamic control over output quality to deliver varying utility levels, and the latter offers users precise control to shift model's focused input features. AIM introduces a logits redistribution strategy that operates in a training data-agnostic and retraining-free manner. We establish a formal foundation to ensure AIM's regulation capability, based on the statistical properties of logits ordering via joint probability distributions. Our evaluation confirms AIM's practicality and versatility for Al model modulation, with tasks spanning image classification, semantic segmentation and text generation, and prevalent architectures including ResNet, SegFormer and Llama.
Paper Structure (36 sections, 3 theorems, 12 equations, 7 figures, 1 table)

This paper contains 36 sections, 3 theorems, 12 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Our work
  3. Contributions
  4. Availability
  5. Problem Formulation
  6. Deep Neural Networks
  7. Motivation
  8. Defining Model Modulation
  9. Utility Modulation
  10. Focus Modulation
  11. Our Approach -- Aim
  12. Logits Redistribution
  13. Model Logits
  14. Logits Redistribution
  15. Utility Modulation
  16. ...and 21 more sections

Key Result

Theorem 1

Let $\hat{y} = (\hat{y}_1, \hat{y}_2, \ldots, \hat{y}_n)$ be a vector of logits with an ordering $\hat{y}_{\tau_1} \leq \hat{y}_{\tau_2} \leq \cdots \leq \hat{y}_{\tau_n}$, where $\tau$ is a permutation of ${1, 2, \ldots, n}$. Let $\epsilon = (\epsilon_1, \epsilon_2, \ldots, \epsilon_n)$ be a vector where $\Delta_i = \hat{y}_{\tau_{i+1}} - \hat{y}_{\tau_i}$ and $\Phi(\cdot)$ is the cumulative dist

Figures (7)

  • Figure 1: An illustration of Aim's logits redistribution.
  • Figure 2: Classification and semantic segmentation performance under varying noise levels ($\sigma$) for utility modulation.
  • Figure 3: Performance of Llama-3.1-8B on GSM8K and MMLU datasets with different noise levels ($\sigma$), accompanied by a sample MMLU question.
  • Figure 4: Segmentation of pedestrians improves progressively with moderate noise levels (c-e) compared to no noise ($\sigma = 0$), where pedestrians are partially or not detected (b).
  • Figure 5: Focus modulation enhances targeted class accuracy but risks reducing overall mIoU if adjustments are excessive.
  • ...and 2 more figures

Theorems & Definitions (3)

  • Theorem 1
  • Theorem 2
  • Theorem 3