Adapting AI models from one domain to another
A method for adapting to a new domain an AI model pre-trained for a current domain. Having at least one main block of the AI model for modeling a target variable and having at least one covariates block of the AI model for modeling covariates effect on the target variable in the current domain, the method comprises: replacing the covariates block with a new covariates block adapted to the new domain, the new covariates block modifying one or more first layers compared to the covariate block, the target variable in the new domain being affected differently by at least one of the one or more covariates; training the new covariates block of the AI model using a new-domain-specific dataset from the new domain; and fine-tuning the at least one main block of the AI model using the new-domain-specific dataset from the new domain.
1 . A method for adapting an artificial intelligence (AI) model pre-trained for a current domain to a new domain, the method comprising:
providing at least one main block of the AI model for modeling a target variable and at least one covariates block of the AI model for modeling effects of one or more covariates on the target variable in the current domain, the AI model being pre-trained to forecast future values of the target variable using past values thereof in the current domain, the future values of the target variable being affected by the one or more covariates wherein the one or more covariates are independent from the target variable, and in order to adapt the AI model to the new domain:
replacing a covariates block of the at least one covariates block with a new covariates block adapted to the new domain, wherein the covariates block comprises one or more first layers comprising a number of hidden units with respective weights, with the new covariates block modifying the number of the hidden units, the respective weights, or both, of the one or more first layers based on covariates of the new domain, the target variable in the new domain being affected differently by at least one of the one or more covariates;
training the new covariates block of the AI model using a new-domain-specific dataset from the new domain; and
fine-tuning the at least one main block of the AI model using the new-domain-specific dataset from the new domain.
2 . The method of claim 1 , wherein the at least one main block of the AI model models the target variable by producing a forecast of the future values of the target variable.
3 . The method of claim 1 , wherein the target variable in the new domain is affected by at least one covariate different from the one or more covariates affecting the target variable in the current domain.
4 . The method of claim 1 , wherein the new covariates block is chosen to structurally accommodate the covariates of the new domain.
5 . The method of claim 1 , wherein training the new covariates block of the AI model using the new-domain-specific dataset is performed by:
freezing the at least one main block; and
training the AI model using the new-domain-specific dataset.
6 . The method of claim 5 , wherein freezing the at least one main block is performed to prevent the at least one main block from fitting the new-domain-specific dataset.
7 . The method of claim 1 , wherein before fine-tuning the at least one main block of the AI model using the new-domain-specific dataset the method includes:
freezing the covariates block; and
unfreezing the at least one main block.
8 . The method of claim 1 , wherein the main block is a neural network based model for univariate time series forecasting.
9 . The method of claim 1 , wherein fine-tuning the at least one main block of the AI model on data from the new domain is performed using incremental moment matching algorithms.
10 . The method of claim 1 , wherein fine-tuning the at least one main block of the AI model on data from the new domain is performed using transfer learning based fine-tuning.
11 . The method of claim 1 , further comprising computing a backcast of the past values of the target variable.
12 . An artificial intelligence server configured for adapting an artificial intelligence (AI) model pre-trained for a current domain to a new domain, the artificial intelligence server comprising:
a memory module for storing a new-domain-specific dataset and a current-domain-specific dataset;
a processor module that, having at least one main block of the AI model for modeling a target variable and at least one covariates block of the AI model for modeling effects of one or more covariates on the target variable in the current domain, the AI model being pre-trained to forecast future values of the target variable using past values thereof in the current domain, the future values of the target variable being affected by the one or more covariates wherein the one or more covariates are independent from the target variable, and in order to adapt the AI model to the new domain, the processor module is configured to:
replace a covariates block of the at least one covariates block with a new covariates block adapted to the new domain, wherein the covariates block comprises one or more first layers comprising a number of hidden units with respective weights, with the new covariates block modifying the number of the hidden units, the respective weights, or both, of the one or more first layers based on covariates of the new domain, the target variable in the new domain being affected differently by at least one of the one or more covariates;
train the new covariates block of the AI model using a new-domain-specific dataset from the new domain; and
fine-tune the at least one main block of the AI model using the new-domain-specific dataset from the new domain.
13 . The artificial intelligence server of claim 12 , wherein the at least one main block of the AI model models the target variable by producing a forecast of the future values of the target variable.
14 . The artificial intelligence server of claim 12 , wherein the target variable in the new domain is affected by at least one covariate different from the one or more covariates affecting the target variable in the current domain.
15 . The artificial intelligence server of claim 12 , wherein the new covariates block is chosen to structurally accommodate the covariates of the new domain.
16 . The artificial intelligence server of claim 12 , wherein training the new covariates block of the AI model using the new-domain-specific dataset is performed by:
freezing the at least one main block; and
training the AI model using the new-domain-specific dataset.
17 . The artificial intelligence server of claim 16 , wherein freezing the at least one main block is performed to prevent the at least one main block from fitting the new-domain-specific dataset.
18 . The artificial intelligence server of claim 12 , wherein the processor module is configured to before fine-tuning the at least one main block of the AI model using the new-domain-specific dataset:
freeze the covariates block; and
unfreeze the at least one main block.
19 . The artificial intelligence server of claim 12 , wherein the main block is a neural network based model for univariate time series forecasting.
20 . The artificial intelligence server of claim 12 , wherein fine-tuning the at least one main block of the AI model on data from the new domain is performed using incremental moment matching algorithms.
21 . The artificial intelligence server of claim 12 , wherein fine-tuning the at least one main block of the AI model on data from the new domain is performed using transfer learning based fine-tuning.
22 . The artificial intelligence server of claim 12 , wherein the processor module is further configured to compute a backcast of the past values of the target variable.