PaperlayerView exact revision on arXiv

Source-linked AI summary

FairCompressAgent: An Agentic Framework for Fairness-Aware Model Compression for FPGA Deployment

Yuanbo Guo, Yiyu Shi

arXiv:2609.17786v1cs.AIcs.AR

TL;DR

Compression decisions conventionally balance prediction quality and computational cost without explicitly considering demographic disparities, although compression can disproportionately affect bias even when accuracy changes little. FCA selects and composes pruning, quantization, and sparse low-rank factorization through a common operator interface, using model profiles and measured outcomes to guide choices under user-defined requirements. The selected configuration achieved a 59.54% storage reduction from Vanilla in inference tensor storage.

  • Problem

    Compression decisions conventionally balance prediction quality and computational cost without explicitly considering demographic disparities, although compression can disproportionately affect bias even when accuracy changes little.

    Source: §I · p. 1

  • Method

    FCA selects and composes pruning, quantization, and sparse low-rank factorization through a common operator interface, using model profiles and measured outcomes to guide choices under user-defined requirements.

    Source: §I · p. 2, E. Operator Information for Planning · p. 4, A. Planning, Execution, and Evaluation · p. 4

  • Results

    The selected configuration achieved a 59.54% storage reduction from Vanilla in inference tensor storage.

    Source: B. Results of the Compression Operators · p. 6

  • Takeaways & Limitations

    FCA connects measured model execution with explicit constraint reporting and online adaptation to revised requirements in its Fitzpatrick-17k experiments.

    Source: §VII · p. 8

Abstract

from arXiv · show

Fairness-aware model compression requires selecting methods and configurations that balance accuracy, fairness, and deployment cost. These decisions become more difficult when compression methods are composed or the user's requirements change. In this paper, we propose FairCompressAgent (FCA), an agentic framework that integrates fairness-aware pruning, incremental quantization, and sparse low-rank factorization through a common operator interface. A language-model planner uses model profiles and measured outcomes to select compression configurations, while an execution layer performs compression, fine-tuning, evaluation, and constraint-based selection. FCA also supports requirement updates and reports the remaining violation when a request cannot be satisfied. Experiments on Fitzpatrick-17k with VGG-11 compare four search methods over 40 measured configurations. Under the accuracy-constrained request, FCA selects a compressed model with 59.54% less inference tensor storage, while validation average precision increases from 0.5141 to 0.5233 and equalized opportunity (EOpp) decreases from 0.2251 to 0.2168. It reaches the same final selection as one-shot planning with 7.33 versus 12 candidate evaluations on average, under their respective stopping policies. Repeated fine-tuning, held-out testing, and online requirement updates characterize the stability and interactive use of this compression workflow. The results demonstrate how measured feedback and explicit constraints support the selection and interactive refinement of fairness-aware compression configurations.

I. INTRODUCTION

FairCompressAgent (FCA) selects and composes fairness-aware compression operators under user-defined requirements, using model profiles and measured outcomes to guide its choices. The work addresses the need to weigh accuracy, fairness, and deployment efficiency together and supports updated requirements during search.

I. INTRODUCTION

  • Fairness and efficiency are often treated separately, although compression can disproportionately affect bias even when accuracy changes little.Compression decisions conventionally balance prediction quality and computational cost without explicitly considering demographic disparities.Source: §I · p. 1
  • Fairness-aware compression methods offer different operator choices, but configuration decisions still include compression strength, fine-tuning, layer coverage, and method order.The cited approaches include group-sensitive pruning, incremental quantization, and sparse low-rank factorization.Source: §I · p. 1
  • FCA draws on model profiles and evaluation results to guide compression choices, while users can revise requirements during the search.The introduction frames operator compatibility and explicit accuracy, fairness, and efficiency criteria as requirements for coordinating these decisions.Source: §I · p. 2
  • FCA organizes pruning, quantization, and sparse low-rank factorization through a common operator interface for constrained configuration selection.Its agent architecture checks constraints and supports requirement updates.Source: §I · p. 2
  • On Fitzpatrick-17k, FCA selected a model with 59.54% less inference tensor storage under the accuracy-constrained request and reported the unmet constraint under a stricter fairness request.The evaluation used 40 measured configurations and four search methods.Source: §I · p. 2

II. PROBLEM FORMULATION

The formulation treats compression as selecting a compressed model that meets user requirements for accuracy, fairness, and efficiency. It defines group-based average precision and fairness measures, and specifies inference storage as an efficiency metric.

A. Inputs and Accuracy Metrics

  • The objective is to find a compressed model that satisfies user requirements on accuracy, fairness, and efficiency.The task starts from a pre-trained model and uses binary sensitive attributes to divide annotated samples into groups.Source: A. Inputs and Accuracy Metrics · p. 2
  • Average precision is the main accuracy measure and is the mean of the macro precision values of the two demographic groups.The formulation follows FairPrune, FairQuantize, and FairLRF in using this measure.Source: A. Inputs and Accuracy Metrics · p. 2
  • For each group and class, Precg,k denotes one-versus-rest precision, set to zero when no sample in that group is predicted as that class.Source: A. Inputs and Accuracy Metrics · p. 2

B. Fairness and Efficiency Metrics

  • Equalized opportunity is defined using summed differences in group true-negative or true-positive rates within each target class; equalized odds sums true-positive- and false-positive-rate differences.Source: B. Fairness and Efficiency Metrics · p. 2
  • Lower EOpp0, EOpp1, and EOdd values are preferred.Source: B. Fairness and Efficiency Metrics · p. 2
  • The paper uses EOpp1 as its fairness representation, denoted EOpp, because fairness for positive cases is often more important in scenarios such as disease diagnosis.Source: B. Fairness and Efficiency Metrics · p. 2
  • Storage S(ρ) counts bytes in inference parameter and buffer tensors, excluding optimizer state, training masks, and file headers.The experiments use dense FP32 tensors, including low-rank layer factors; execution cost also depends on mapping the representation to the target architecture.Source: B. Fairness and Efficiency Metrics · p. 2
  • A compressed model is represented by z = (θ, ρ), where ρ specifies its executable form, including factor layers and constrained weights.Storage therefore measures the representation used for inference, while execution cost additionally depends on its mapping to the target architecture.Source: B. Fairness and Efficiency Metrics · p. 2

C. Constrained Configuration Selection

Compression candidates are defined by ordered operators and settings, then filtered by user-specified constraints and ranked by a fixed priority among feasible observations.

III. FAIRNESS-AWARE COMPRESSION OPERATORS

FCA organizes fairness-aware pruning, quantization, and factorization around a common group-sensitive score and evaluation interface. Pruning and quantization modify weights according to that score, with constraints preserved during fine-tuning.

A. A Common Group-Sensitive Score

  • FairPrune, FairQuantize, and FairLRF are organized as compression operators with a common scoring and evaluation interface.Their group-sensitive scores guide which weights or factors to modify.Source: A. A Common Group-Sensitive Score · p. 3
  • The diagonal quadratic score ranks perturbations using group-specific curvature, with β adjusting the local ranking between reference groups.The experiment fixes Light and Dark as the underperforming and better-performing reference groups.Source: A. A Common Group-Sensitive Score · p. 3
  • Full evaluation after fine-tuning, not the local score alone, determines whether a resulting model satisfies user requirements.Source: A. A Common Group-Sensitive Score · p. 3

B. Pruning and Incremental Quantization

  • FairPrune removes eligible weights with the lowest scores within each target tensor and keeps them zero through subsequent fine-tuning.Removing a weight sets it to zero, and a cumulative mask preserves that constraint.Source: B. Pruning and Incremental Quantization · p. 3
  • FairQuantize scores weight changes by squared displacement, then incrementally freezes selected power-of-two weights while remaining weights are fine-tuned.Previously frozen entries remain fixed during later increments; the resulting weights can map to a compatible quantized execution backend.Source: B. Pruning and Incremental Quantization · p. 3

C. Sparse Low-Rank Factorization

Sparse low-rank factorization approximates a linear layer with retained singular directions, then uses group-sensitive scores to sparsify its factors. The resulting representation uses two linear maps and targets fully connected layers.

C. Sparse Low-Rank Factorization

  • For a linear layer with m inputs and n outputs, truncated singular value decomposition represents its weight matrix using a retained rank r.Source: C. Sparse Low-Rank Factorization · p. 3
  • FairLRF sparsifies factor entries using group-sensitive scores, retaining a prefix of singular directions and scoring tails of factor rows and columns.The retained prefix is q = ⌊αr⌋; curvature is computed in factor coordinates with singular values fixed during scoring.Source: C. Sparse Low-Rank Factorization · p. 3
  • The factor matrices require r(m + n) FP32 elements versus mn for the original layer, and inference applies two consecutive linear maps.The original bias is applied at the output, and the evaluated factorization operator targets fully connected layers.Source: C. Sparse Low-Rank Factorization · p. 3

D. Composition of Operators

FCA composes compatible compression operators and exposes their representation and evaluation information to a planner. The planner selects supported configurations, while an executor performs transformations and fine-tuning and an evaluator measures outcomes.

D. Composition of Operators

  • FCA supports P → Q and L → Q, preserving pruning zeros or fixed factor sparsity during quantization and subsequent fine-tuning.Reverse compositions are excluded because factorization may not preserve a quantization codebook or pruning zeros.Source: D. Composition of Operators · pp. 3–4

E. Operator Information for Planning

  • For linear layers, dense factors reduce matrix storage when r < mn/(m + n), and the planner receives retained-rank storage, evaluation metrics, and representation constraints.Pruning masks and quantized values specify constraints for subsequent transformations.Source: E. Operator Information for Planning · p. 4

A. Planning, Execution, and Evaluation

  • The language-model planner selects from a finite catalog using model profiles, user constraints, search budget, structural information, and queried evaluation results.The executor transforms and fine-tunes models, while the evaluator measures average precision, EOpp, group coverage, inference tensor storage, and execution cost.Source: A. Planning, Execution, and Evaluation · p. 4

B. State and Decision Cycle

  • FCA checks configuration support and budget before evaluation, reuses records for repeated requests, and records outcomes and remaining budget after new online executions.For a new configuration, the executor performs compression, fine-tuning, and validation from the supplied baseline.Source: B. State and Decision Cycle · p. 4

C. Measured Feedback and Requirement Updates

FCA uses measured outcomes to compare search methods and can reapply its selection rule when requirements change. It also imposes evaluation and request budgets and reports the smallest normalized violation when no candidate is feasible.

C. Measured Feedback and Requirement Updates

  • Search methods query a shared pre-evaluated candidate set, with each distinct query adding its recorded execution time to accumulated model evaluation time.The planner sees only queried outcomes, keeping candidate measurements identical across search methods.Source: C. Measured Feedback and Requirement… · p. 4
  • A requirement update changes the current constraints and reapplies selection to existing evaluation records without further model execution.Each selection retains the requirements under which it was made.Source: C. Measured Feedback and Requirement… · p. 4

D. Stopping and Unmet Constraints

  • Sessions limit distinct evaluations, accumulated model evaluation time, and language-model requests, though the planner may stop earlier when observed trade-offs suffice.Source: D. Stopping and Unmet Constraints · p. 4
  • When no observed candidate is feasible, FCA reports the unmet request and smallest normalized violation, with component violations identifying unsatisfied requirements.Changing a bound requires a user instruction.Source: D. Stopping and Unmet Constraints · p. 4

1) Dataset and Backbone:

Experiments use Fitzpatrick-17k with a fixed VGG-11 checkpoint and compare compression configurations under two predefined requests and several search methods.

1) Dataset and Backbone:

  • Fitzpatrick-17k groups skin types 1–3 as Light and 4–6 as Dark; experiments use a fixed pre-trained VGG-11 adapted to 112 × 112 inputs and 114 outputs.Evaluation uses resize and center-crop preprocessing; fine-tuning adds random cropping and horizontal flipping.Source: 1) Dataset and Backbone · p. 4

2) Compression Configurations:

  • The candidate set contains 40 configurations spanning five operator sequences, two layer sets, two strengths, and two group-weighting settings.All configurations receive two fine-tuning epochs, providing a shared comparison budget.Source: 2) Compression Configurations · pp. 4–5

3) User Requirements and Search Methods:

  • Request U limits average-precision loss to one percentage point and EOpp to no worse than Vanilla; request F permits three points of loss, requires EOpp ≤0.20, and caps storage at 80% of Vanilla.Random search, TPE, and one-shot planning are compared with FCA using three shared initial observations; one-shot commits to nine further candidates while FCA adapts and may stop early.Source: 3) User Requirements and Search Methods · p. 5

B. Results of the Compression Operators

Across the measured configurations, feasibility depends on operator sequence and settings; under U, low-rank factorization yields the best feasible candidate, while the stricter F request remains unmet.

B. Results of the Compression Operators

  • 10 of 40 configurations satisfy U, including five from L, three from P → Q, and two from L → Q; individual P and Q configurations are outside the feasible region.The validation plot places higher precision and lower EOpp toward the lower right; its dashed line marks the Vanilla EOpp upper bound used by U.Source: B. Results of the Compression Operators · pp. 5–6, Figure 3 · p. 6
  • 59.54% less inference tensor storage accompanies Pavg = 0.5233 and EOpp = 0.2168 for the best U-feasible candidate, low-rank factorization on both hidden classifier layers at rank 512 and β = 1.The candidate's validation precision increases by 0.918 percentage points and EOpp decreases by 0.00829.Source: B. Results of the Compression Operators · p. 6
  • Adding Q to the best L configuration yields Pavg = 0.5146 and EOpp = 0.2365 at unchanged tensor storage, making that L → Q candidate infeasible under U.P → Q nevertheless supplies three U-feasible configurations, showing that outcomes vary with the complete operator sequence and settings.Source: B. Results of the Compression Operators · p. 6

C. Comparison of Search Methods

  • Under U, FCA and one-shot planning select the best feasible candidate in all three runs, with median storage ratio 0.4046 versus 0.5953 for Random and TPE; FCA averages 7.33 evaluations versus 12.The corresponding accumulated model evaluation times are 13.66 and 22.11 minutes; the comparison reflects FCA's adaptive stopping and one-shot planning's committed list.Source: C. Comparison of Search Methods · p. 6
  • Under F, every method reports an unmet request because the candidate-set minimum EOpp is 0.2168, above the 0.20 bound; FCA, one-shot planning, and TPE have median violation 0.0839 versus Random's 0.1059.The 0.8 storage bound also excludes dense P, Q, and P → Q representations based on their known storage requirements.Source: C. Comparison of Search Methods · p. 6

D. Repeated Fine-Tuning and Held-Out Testing

Held-out testing reports a storage reduction alongside average-precision and EOpp comparisons, while planning costs and online updates show how FCA handles different requirements. The online cases report both revised selections and an unmet fairness constraint.

F. Online Execution and User Feedback

  • After the user raised U’s permitted average-precision loss from one to two percentage points, FCA selected L with A/strong/5/9: validation Pavg 0.5184, EOpp 0.2186, and storage ratio 0.7225.Source: F. Online Execution and User Feedback · p. 6
  • After the user tightened F’s EOpp bound from 0.20 to 0.19, FCA reported L with B/strong/1 infeasible, with normalized violation 0.1410 and EOpp identified as unsatisfied.Source: F. Online Execution and User Feedback · p. 6

D. Repeated Fine-Tuning and Held-Out Testing

E. Planning Cost

  • Under U, FCA used 22 requests and 75,421 completion tokens versus one-shot planning’s 3 requests and 69,653 tokens; request durations were 15.80 versus 18.82 minutes.Both methods produced the same final U selection.Source: E. Planning Cost · p. 6, Table IV · p. 7
  • Under F, FCA used 35 requests and 154,014 completion tokens versus one-shot planning’s 3 requests and 67,774 tokens, though both reached the same least-violating model.Request and model-time measurements characterize the complete search workflow separately.Source: E. Planning Cost · p. 6, Table IV · p. 7

VI. DISCUSSION

The discussion treats the complete compression configuration as the comparison unit because compositions can alter accuracy–fairness trade-offs. It connects representation-aware operator design with FCA’s measured, constraint-based selections.

A. Implications for Compression Selection

  • Composed operators can change an existing configuration’s accuracy–fairness trade-off, so selection should compare complete configurations and associate recommendations with measured metrics and constraints.Source: A. Implications for Compression… · p. 7

B. Representation-Aware Design

  • Pruning, quantization, and factorization yield different representations whose suitability depends on target resources and execution patterns; FCA encodes these distinctions through compatible operator sequences.Source: B. Representation-Aware Design · p. 7

VII. CONCLUSION

  • On Fitzpatrick-17k, FCA reports 59.54% less inference tensor storage under the accuracy-constrained request, alongside explicit unmet-constraint reporting and adaptation to revised requirements.Source: §VII · p. 8
Loading 2609.17786v1…