Fine-tuning Generative Model Using Efficient Decoding Instances

Solution Overview

Problem

Existing decoding methods for generative models, which achieve improved accuracy, are less computationally efficient, leading to significant resource utilization and latency, particularly when processing large volumes of inputs.

Innovation Solution

Utilizing a less computationally efficient decoding method to generate training instances for fine-tuning a student generative model, and then employing a more computationally efficient decoding method at inference time to generate generative predictions, thereby balancing accuracy and computational efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a less computationally efficient decoding method (such as QE reranking or MBR decoding) is used to decode probability distributions, then the accuracy of generative content is improved, but computational resource utilization increases and latency increases

Engineering Contradiction:

Improveaccuracy of generative contentVSAvoidcomputational resource utilization

Core Design Contradiction:

Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system performs preliminary action by using the less computationally efficient decoding method to generate training instances and fine-tune the student generative model before actual inference. This preliminary fine-tuning process embeds the accuracy benefits of sophisticated decoding methods into the model itself, so that during subsequent inference, simpler and faster decoding methods can achieve comparable accuracy without the computational cost of the original sophisticated methods.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system creates a student generative model that copies the desirable accuracy characteristics from a teacher model fine-tuned with sophisticated decoding methods. The student model is trained on training instances generated using QE reranking or MBR decoding, allowing the student to inherit these accuracy properties while being deployed with more efficient decoding methods.

Inventive Principle:

Principle #26Copying

2Measurement precision

If a less computationally efficient decoding method is used to decode probability distributions, then the accuracy of generative content is improved, but latency increases

Engineering Contradiction:

Improveaccuracy of generative contentVSAvoidlatency

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary action by using the less computationally efficient decoding method to generate training instances and fine-tune the student generative model before actual inference. This preliminary fine-tuning process embeds the accuracy benefits of sophisticated decoding methods into the model itself, so that during subsequent inference, simpler and faster decoding methods can achieve comparable accuracy without the computational cost of the original sophisticated methods.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system creates a student generative model that copies the desirable accuracy characteristics from a teacher model fine-tuned with sophisticated decoding methods. The student model is trained on training instances generated using QE reranking or MBR decoding, allowing the student to inherit these accuracy properties while being deployed with more efficient decoding methods.

Inventive Principle:

Principle #26Copying

3Productivity

If a more computationally efficient decoding method is used to decode probability distributions, then computational resource utilization is reduced and latency is reduced, but the accuracy of generative content deteriorates

Engineering Contradiction:

Improvecomputational efficiencyVSAvoidaccuracy of generative content

Core Design Contradiction:

ProductivityVSMeasurement precision

Solution Approach 1:

The system performs preliminary action by using the less computationally efficient decoding method to generate training instances and fine-tune the student generative model before actual inference. This preliminary fine-tuning process embeds the accuracy benefits of sophisticated decoding methods into the model itself, so that during subsequent inference, simpler and faster decoding methods can achieve comparable accuracy without the computational cost of the original sophisticated methods.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system creates a student generative model that copies the desirable accuracy characteristics from a teacher model fine-tuned with sophisticated decoding methods. The student model is trained on training instances generated using QE reranking or MBR decoding, allowing the student to inherit these accuracy properties while being deployed with more efficient decoding methods.

Inventive Principle:

Principle #26Copying

Data Source

PatentUS20250077850A1Fine-tuning generative model utilizing instances automatically generated from less computationally efficient decoding and subsequent utilization thereof with more computationally efficient decoding

Publication Date: 2025.03.06 GOOGLE LLC

AI summary

Implementations disclose utilizing a less computationally efficient decoding method in automatically generating corresponding single generative content predictions for training instances and fine-tuning a student generative model based on those automatically generated training instances. Those implementations are further directed to then utilizing, in an inference time environment, the fine-tuned student generative model and a more computationally efficient decoding method in generating generative predictions—and without any utilization of the less computationally efficient decoding method in generating the generative predictions.