Two-Stage Latent Sampling for Accurate Ingredient Deformulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for reverse-engineering ingredient amounts in consumer goods are inefficient and inaccurate, particularly in the cosmetics and hair care industries, where chemical methods are difficult and non-chemical methods result in waste and low accuracy.
Innovation Solution
A deformulation tool using machine learning, specifically variational auto-encoders (VAEs) with one-stage and two-stage sampling techniques, to reconstruct ingredient amounts from ordered lists by leveraging the structure within the VAE latent space for efficient processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If existing non-chemical techniques are used for deformulation, then the process can be performed without chemical complexity, but the accuracy is low and there is significant waste
Solution Approach 1:
The patent replaces traditional chemical analysis methods with a machine learning-based system that uses variational autoencoders to reconstruct ingredient amounts from ordered lists. This substitution of chemical/m mechanical systems with an information-processing system eliminates the need for complex chemical procedures while significantly improving measurement precision through computational modeling trained on formulation data.
2Device complexity
If traditional deformulation methods are used, then the process is simple to implement, but computational resources and processing time are excessive
Solution Approach 1:
The patent implements a two-stage sampling approach where the first stage performs preliminary sampling to identify promising regions in the latent space, and the second stage focuses computational resources on these refined regions. This preliminary action significantly reduces the overall computational burden and processing time while maintaining accuracy, as the system doesn't need to exhaustively search the entire formulation space.
Solution Approach 2:
The deformulation process is segmented into distinct stages: (1) training the VAE model on formulation data, (2) first-stage sampling to explore the latent space, (3) identifying promising regions, and (4) second-stage sampling to refine results. This segmentation allows each stage to be optimized independently and reduces overall computational complexity compared to a monolithic approach.
3Ease of operation
If single-stage sampling is used in the VAE latent space, then the process is straightforward, but it requires significant processing power and time to achieve accurate results
Solution Approach 1:
The sampling process is divided into two stages: first-stage sampling that broadly explores the latent space to identify promising regions, and second-stage sampling that concentrates computational effort on these refined regions. This segmentation dramatically reduces the total computational resources required compared to single-stage sampling, as the second stage operates on a much smaller, more relevant subspace.
Solution Approach 2:
The first-stage sampling serves as a preliminary action that prepares the system by identifying promising regions in the latent space before the second-stage sampling begins. This preliminary exploration prevents wasteful computation in the second stage by ensuring that computational resources are focused only on regions likely to contain accurate deformulation results.
Data Source
AI summary
A device receives an ingredient list having a sequence of ingredients ordered by relative amount, and generates formulation vectors by sampling the ingredients list. The device inputs the plurality of formulation vectors into a machine-learned model, the machine-learned model generating an encoded version of each of the plurality of formulation vectors using an encoder, and then outputting a plurality of reconstructed formulation vectors as derived using a decoder. The device identifies reconstructed formulation vectors that have an order that matches the sequence, defines a latent space using the encoded version of the matching reconstructed formulation vectors. The device iteratively samples the latent space until a threshold number of samples are derived that match an ordering constraint that corresponds to the sequence, performs a statistical aggregation of the samples, and outputs an indication of an absolute amount of each ingredient in the ingredients list.