Anomalies and drift detection in decentralized learning environments

1 . A computing component at a first node, comprising:

at least one processor; and

a memory operatively connected to the at least one processor, the memory including instructions that, when executed, cause the at least one processor to:

obtain a local training data reference baseline for the first node of a decentralized learning network, by training, at the first node, a local variational autoencoder, wherein the local variational autoencoder is an unsupervised artificial neural network that is configured to compress and encode training data into a continuous latent space and subsequently reconstruct the training data;

deploy at the first node, (1) the local variational autoencoder along with the local training data reference baseline, and (2) a global variational autoencoder trained across a plurality of nodes, along with a corresponding global training data reference baseline;

process production data at the first node with a local machine learning model and a global machine learning model, and derive local and global anomaly data regarding anomalous production data based on the local training data reference baseline and the global training data reference baseline, respectively, by:

extracting reconstruction losses and latent spaces from the production data,

generating clusters from the extracted reconstruction losses and latent spaces,

detecting and extracting anomalies from the generated clusters, and

detecting data drift in the generated clusters using the respective local or global training data reference baselines;

compare the local anomaly data with the global anomaly data;

assess and present impact on user ML model predictions of the anomalous production data on the local and global machine learning models; and

trigger re-training of at least one of the local and global machine learning models based on the impact on user ML model predictions.

2 . The computing component of claim 1 , wherein the clusters are generated using Mean Shift Clustering on the extracted reconstruction losses and latent spaces.

3 . The computing component of claim 1 , wherein:

to detect and to extract the anomalies, the memory includes instructions that when executed further cause the at least one processor to identify a high density cluster and detect and extract data points that are furthest from the high density cluster and above a maximum value of the respective local or global training data reference baselines; and

the high density cluster is identified based, at least in part, on a density distribution comparison that compares the density counts for the clusters with respect to one another.

4 . The computing component of claim 1 , wherein the memory includes instructions that when executed further cause the at least one processor to detect the data drift using cross-correlation values of the generated clusters to determine a rate of change in the data drift.

5 . The computing component of claim 1 , wherein the memory includes instructions that when executed further cause the at least one processor to detect the data drift by assessing density changes of the generated clusters over a period of time for each data batch across clusters per data batch, with reference to the local or global training data reference baselines.

6 . The computing component of claim 1 , wherein the memory includes instructions that when executed further cause the at least one processor to compare the local anomaly data with the global anomaly data using a density distribution comparison.

7 . A computer-implemented method, comprising:

training, at a first node of a decentralized learning network, a local variational autoencoder and obtaining a local training data reference baseline for the first node, wherein the local variational autoencoder is an unsupervised artificial neural network and the training comprises compressing and encoding training data into a continuous latent space and subsequently reconstructing the training data;

training, across a plurality of participating nodes, a global variational autoencoder and obtaining a corresponding global training data reference baseline;

deploying at the first node, (1) the local variational autoencoder along with the local training data reference baseline, and (2) the global variational autoencoder trained across a plurality of nodes, along with the corresponding global training data reference baseline;

processing production data at the first node with a local machine learning model and a global machine learning model deployed by a user, and deriving local and global anomaly data regarding anomalous production data based on the local training data reference baseline and the global training data reference baseline, respectively, by:

extracting reconstruction losses and latent spaces from the production data,

generating clusters from the extracted reconstruction losses and latent spaces,

detecting and extracting anomalies from the generated clusters, and

detecting data drift in the generated clusters using the respective local or global training data reference baselines;

comparing the local anomaly data with the global anomaly data;

assessing and presenting impact on user ML Model predictions of the anomalous production data on the local and global machine learning models; and

triggering (1) discarding of the anomalous production data or (2) re-training of at least one of the local and global machine learning models, based on the impact on user ML Model predictions.

8 . The method of claim 7 , wherein the clusters are generated using Mean Shift Clustering on the extracted reconstruction losses and latent spaces.

9 . The method of claim 7 , wherein

the detecting and the extracting anomalies comprises identifying a high density cluster and detecting and extracting data points that are furthest from the high density cluster and above the maximum value of the respective local or global training data reference baselines; and

the high density cluster is identified based, at least in part, on a density distribution comparison that compares density counts for the clusters with respect to one another.

10 . The method of claim 7 , further comprising detecting the data drift using cross-correlation values of the generated clusters to determine a rate of change in the data drift.

11 . The method of claim 7 , further comprising detecting the data drift by assessing density changes of the generated clusters over a period of time for each data batch across clusters per data batch, with reference to the local or global training data reference baselines.

12 . The method of claim 7 , in which the comparing of the local anomaly data with the global anomaly data is performed using a density distribution comparison.

13 . The method of claim 7 , further comprising:

processing the production data at the first node with the local machine learning model and the global machine learning model deployed by a user, and deriving local and global drift data regarding drifting production data based on the local training data reference baseline and the global training data reference baseline, respectively;

comparing the local drift data with the global drift data;

assessing and presenting impact on user ML Model predictions of the drifting production data on the local and global machine learning models; and

triggering (1) discarding of the drifting production data or (2) re-training of at least one of the local and global machine learning models, based on the impact on user ML Model predictions.

14 . A node operating in a decentralized learning network, comprising:

at least one processor; and

a memory operatively connected to the at least one processor, the memory including instructions that, when executed, cause the at least one processor to:

obtain a local training data reference baseline for the node, by training, at the node, a local variational autoencoder, wherein the local variational autoencoder is an unsupervised artificial neural network that is configured to compress and encode training data into a continuous latent space and subsequently reconstruct the training data;

deploy at the node (1) the local variational autoencoder along with the local training data reference baseline and (2) a global variational autoencoder trained across a plurality of nodes in the decentralized learning network, along with a corresponding global training data reference baseline;

process production data at the node with a local machine learning model and a global machine learning model, and derive local and global drift data regarding drifting production data based on the local training data reference baseline and the global training data reference baseline, respectively, by:

extracting reconstruction losses and latent spaces from the production data,

generating clusters from the extracted reconstruction losses and latent spaces,

detecting and extracting anomalies from the generated clusters, and

detecting data drift in the generated clusters using the respective local or global training data reference baselines;

compare the local drift data with the global drift data;

assess and present impact on user ML Model predictions of the drifting production data on the local and global machine learning models; and

implement a decision with respect to at least one of the local and global machine learning models based on the impact on user ML Model predictions.

15 . The node of claim 14 , wherein the decision includes either (1) discarding the drifting production data or (2) re-training of at least one of the local and global machine learning models.