Distributed machine learning systems including generation of synthetic data
A distributed, online machine learning system is presented. Contemplated systems include many private data servers, each having local private data. Researchers can request that relevant private data servers train implementations of machine learning algorithms on their local private data without requiring de-identification of the private data or without exposing the private data to unauthorized computing systems. The private data servers also generate synthetic or proxy data according to the data distributions of the actual data. The servers then use the proxy data to train proxy models. When the proxy models are sufficiently similar to the trained actual models, the proxy data, proxy model parameters, or other learned knowledge can be transmitted to one or more non-private computing devices. The learned knowledge from many private data servers can then be aggregated into one or more trained global models without exposing private data.
1 . A computer-based distributed machine learning system comprising:
at least one private data server storing local private data and having a local modeling agent; and
a non-private data server coupled with the at least one private data server over a network, the non-private data server lacking authorized access to the local private data, the non-private data server comprising at least one processor that, upon execution of software instructions stored in a computer readable memory, performs operations of:
transmitting a definition of a machine learning task to the modeling agent of the at least one private data server, the definition of the machine learning task including machine learning model instructions and local private data features;
enabling the modeling agent of the at least one private data server to generate synthetic data capable of reproducing knowledge gained from execution of the machine learning model instructions on at least some of the local private data having the local private data features, wherein the synthetic data is generated based on at least one data distribution of the local private data which includes multiple patient sample points for different patients, the synthetic data includes at least one synthetic data distribution of synthetic patient data which is different than the at least one data distribution of the local private data, and each of the multiple patient sample points includes at least one a symptom, a test result, a provider name, a diagnosis, a current procedural terminology (CPT) code, an international classification of diseases (ICD) code, or a diagnostic and statistical manual of mental disorders (DSM) code;
receiving from the modeling agent of the at least one private data server, first proxy model data representative of the knowledge gained via the synthetic data;
comparing the first proxy model data with second proxy model data received from a second private data server;
in response to the first proxy model data and the second proxy model data having at least one of a same shape and a same overall property, aggregating a combination of the second proxy model data and the first proxy model data representative of knowledge gained via the synthetic data into a global model corresponding to the machine learning task; and
flagging the second proxy model data in response to the second proxy model data having a different shape than the first proxy model data, to determine whether an underlying private data distribution set of the second proxy model data is corrupted, has missing data, or contains multiple outliers.
2 . The system of claim 1 , wherein the first proxy model data comprises compressed learned data.
3 . The system of claim 1 , wherein the first proxy model data comprises lossy compressed learned data.
4 . The system of claim 1 , wherein the first proxy model data comprises the synthetic data.
5 . The system of claim 1 , wherein the first proxy model data comprises proxy model parameters derived from the synthetic data.
6 . The system of claim 5 , wherein the operations further include duplicating the synthetic data according to the proxy model parameters.
7 . The system of claim 6 , wherein the proxy model parameters include a seed for a deterministic function able to generate the synthetic data.
8 . The system of claim 1 , wherein the first proxy model data comprises a trained proxy model trained on the synthetic data.
9 . The system of claim 1 , wherein the synthetic data comprises Monte Carlo data.
10 . The system of claim 1 , wherein the local private data comprises private patient data.
11 . The system of claim 10 , wherein the private patient data comprises at least one of healthcare data or genomic data.
12 . The system of claim 1 , wherein the local private data includes at least one of insurance data, financial data, social media profile data, human capital data, proprietary experimental data, gaming or gambling data, military data, network traffic data, or shopping or marketing data.
13 . The system of claim 1 , wherein the operations further include paying a fee in exchange for accessing the modeling agent of the at least one private data server.
14 . The system of claim 1 , wherein the machine learning model instructions comprise at least one of supervised machine learning instructions, unsupervised machine learning instructions, or machine learning clustering instructions.
15 . The system of claim 1 , wherein the machine learning model instructions comprise at least one of machine learning regression instructions or machine learning classification instructions.
16 . The system of claim 1 , wherein the operations further include receiving private data metadata about the local private data from the local modeling agent.
17 . The system of claim 16 , wherein the operations further include generating the definition of the machine learning task based on the private data metadata.
18 . The system of claim 17 , wherein the private data metadata comprises attribute space information relating to the local private data.
19 . The system of claim 1 , wherein synthetic data is generated such that similarity score between a trained proxy model trained on the synthetic data and a trained actual model trained on at least some of the local private data, is below a threshold.
20 . The system of claim 19 , wherein the first proxy model data is received upon satisfaction of transmission requirements defined based on the similarity score.
21 . A method of computer-based distributed machine learning, the method comprising:
transmitting, by at least one processor of a non-private data server coupled with at least one private data server over a network, a definition of a machine learning task to a modeling agent of at least one private data server,
the at least one private data server storing local private data and having a local modeling engine,
the non-private data server lacking authorized access to the local private data, and
the definition of the machine learning task including machine learning model instructions and local private data features;
enabling the modeling agent of the at least one private data server to generate synthetic data capable of reproducing knowledge gained from execution of the machine learning model instructions on at least some of the local private data having the local private data features, wherein the synthetic data is generated based on at least one data distribution of the local private data which includes multiple patient sample points for different patients, the synthetic data includes at least one synthetic data distribution of synthetic patient data which is different than the at least one data distribution of the local private data, and each of the multiple patient sample points includes at least one a symptom, a test result, a provider name, a diagnosis, a current procedural terminology (CPT) code, an international classification of diseases (ICD) code, or a diagnostic and statistical manual of mental disorders (DSM) code;
receiving from the modeling agent of the at least one private data server, first proxy model data representative of the knowledge gained via the synthetic data;
comparing the first proxy model data with second proxy model data received from a second private data server;
in response to the first proxy model data and the second proxy model data having at least one of a same shape and a same overall property, aggregating a combination of the second proxy model data and the first proxy model data representative of knowledge gained via the synthetic data into a global model corresponding to the machine learning task; and
flagging the second proxy model data in response to the second proxy model data having a different shape than the first proxy model data, to determine whether an underlying private data distribution set of the second proxy model data is corrupted, has missing data, or contains multiple outliers.
22 . A non-transitory computer-readable medium comprising computer-executable instructions configured to, when executed by at least one processor, cause the processor to perform operations including:
transmitting, by at least one processor of a non-private data server coupled with at least one private data server over a network, a definition of a machine learning task to a modeling agent of at least one private data server,
the at least one private data server storing local private data and having a local modeling engine,
the non-private data server lacking authorized access to the local private data, and
the definition of the machine learning task including machine learning model instructions and local private data features;
enabling the modeling agent of the at least one private data server to generate synthetic data capable of reproducing knowledge gained from execution of the machine learning model instructions on at least some of the local private data having the local private data features, wherein the synthetic data is generated based on at least one data distribution of the local private data which includes multiple patient sample points for different patients, the synthetic data includes at least one synthetic data distribution of synthetic patient data which is different than the at least one data distribution of the local private data, and each of the multiple patient sample points includes at least one a symptom, a test result, a provider name, a diagnosis, a current procedural terminology (CPT) code, an international classification of diseases (ICD) code, or a diagnostic and statistical manual of mental disorders (DSM) code;
receiving from the modeling agent of the at least one private data server, first proxy model data representative of the knowledge gained via the synthetic data;
comparing the first proxy model data with second proxy model data received from a second private data server;
in response to the first proxy model data and the second proxy model data having at least one of a same shape and a same overall property, aggregating a combination of the second proxy model data and the first proxy model data representative of knowledge gained via the synthetic data into a global model corresponding to the machine learning task; and
flagging the second proxy model data in response to the second proxy model data having a different shape than the first proxy model data, to determine whether an underlying private data distribution set of the second proxy model data is corrupted, has missing data, or contains multiple outliers.