Sensory transformer method of generating ingredients and formulas
Techniques to suggest one or more sets of ingredients that can be used to recreate or mimic a target sensory description using artificial intelligence are disclosed. An artificial intelligence model includes a transformer inspired neural network architecture that learns from ingredients, recipes, and sensory profiles. The artificial intelligence model includes a sensory transformer model that generates a probability distribution of source ingredients based on an embedding associated with first digital data representing ingredients and the second digital data representing sensory description, and a selector that selects at least one candidate ingredient from the probability distribution of source ingredients for the embedding. A complete set of ingredients generated based on the at least one candidate ingredient when combined become a food product that has or achieves the sensory description.
1. A computer-implemented method comprising
applying an artificial intelligence model to first digital data representing base ingredients and second digital data representing a target sensory description, wherein the artificial intelligence model comprises:
a sensory transformer model generating a probability distribution of source ingredients based on an embedding associated with the first digital data representing the base ingredients and the second digital data representing the target sensory description, wherein the sensory transformer model comprises
an encoder that includes a first neural network language-based model outputting the first digital data representing the base ingredients and a second neural network language-based model outputting the second digital data representing the target sensory description, and
a decoder that includes a neural attention-based model using the first digital data and the second digital data to determine the embedding for generating the probability distribution of source ingredients; and
a selector selecting at least one candidate ingredient from the probability distribution of source ingredients;
in response to applying the artificial intelligence model, identifying the at least one candidate ingredient, wherein a complete set of ingredients generated based on the at least one candidate ingredient when combined becomes a food product having the target sensory description.
2. The method of claim 1 , wherein the first digital data is obtained by applying the first neural network language-based model to names of the base ingredients, and the second digital data is obtained by applying the second neural network language-based model to the target sensory description.
3. The method of claim 1 , wherein the first neural network language-based model comprises a language model and a dense layer, wherein [CLS] token representations generated by the language model are used to train the dense layer.
4. The method of claim 1 , wherein the neural attention-based model comprises a plurality of attention layers and a classification layer.
5. The method of claim 4 , wherein the embedding is generated by applying the plurality of attention layers to the first digital data representing the base ingredients and the second digital data representing the target sensory description.
6. The method of claim 4 , wherein the probability distribution of source ingredients is generated by the classification layer for the embedding.
7. One or more non-transitory computer-readable storage media storing one or more instructions programmed, when executed by one or more computing devices, cause:
applying an artificial intelligence model to first digital data representing base ingredients and second digital data representing a target sensory description, wherein the artificial intelligence model comprises:
a sensory transformer model generating a probability distribution of source ingredients based on an embedding associated with the first digital data representing the base ingredients and the second digital data representing the target sensory description, wherein the sensory transformer model comprises
an encoder that includes a first neural network language-based model outputting the first digital data representing the base ingredients and a second neural network language-based model outputting the second digital data representing the target sensory description, and
a decoder that includes a neural attention-based model using the first digital data and the second digital data to determine the embedding for generating the probability distribution of source ingredients; and
a selector selecting at least one candidate ingredient from the probability distribution of source ingredients;
in response to applying the artificial intelligence model, identifying the at least one candidate ingredient, wherein a complete set of ingredients generated based on the at least one candidate ingredient when combined becomes a food product having the target sensory description.
8. The one or more non-transitory computer-readable storage media of claim 7 , wherein the first digital data is obtained by applying the first neural network language-based model to names of the base ingredients, and the second digital data is obtained by applying the second neural network language-based model to the sensory description.
9. The one or more non-transitory computer-readable storage media of claim 7 , wherein the first neural network language-based model comprises a language model and a dense layer, wherein token representations generated by the language model are used to train the dense layer.
10. The one or more non-transitory computer-readable storage media of claim 7 , wherein the neural attention-based model comprises a plurality of attention layers and a classification layer.
11. The one or more non-transitory computer-readable storage media of claim 10 , wherein the embedding is generated by applying the plurality of attention layers to the first digital data representing the base ingredients and the second digital data representing the target sensory description.
12. The one or more non-transitory computer-readable storage media of claim 10 , wherein the probability distribution of source ingredients is generated by the classification layer for the embedding.
13. A computing system comprising:
one or more computer systems comprising one or more hardware processors and storage media; and
instructions stored in the storage media and which, when executed by the computing system, cause the computing system to perform:
applying an artificial intelligence model to first digital data representing base ingredients and second digital data representing a target sensory description, wherein the artificial intelligence model comprises:
a sensory transformer model generating a probability distribution of source ingredients based on an embedding associated with the first digital data representing the base ingredients and the second digital data representing the target sensory description, wherein the sensory transformer model comprises
an encoder that includes a first neural network language-based model outputting the first digital data representing the base ingredients and a second neural network language-based model outputting the second digital data representing the target sensory description, and
a decoder that includes a neural attention-based model using the first digital data and the second digital data to determine the embedding for generating the probability distribution of source ingredients; and
a selector selecting at least one candidate ingredient from the probability distribution of source ingredients;
in response to applying the artificial intelligence model, identifying the at least one candidate ingredient, wherein a complete set of ingredients generated based on the at least one candidate ingredient when combined becomes a food product having the target sensory description.
14. The computing system of claim 13 , wherein the first digital data is obtained by applying the first neural network language-based model to names of the base ingredients, and the second digital data is obtained by applying the second neural network language-based model to the target sensory description.
15. The computing system of claim 13 , wherein the first neural network language-based model comprises a language model and a dense layer, wherein token representations generated by the language model are used to train the dense layer.
16. The computing system of claim 13 , wherein the neural attention-based model comprises a plurality of attention layers and a classification layer.
17. The computing system of claim 13 , wherein the embedding is generated by applying the plurality of attention layers to the first digital data representing the base ingredients and the second digital data representing the target sensory description, wherein the probability distribution of source ingredients is generated by the classification layer for the embedding.