Multi-machine based collaborative learning
Individual nodes (e.g., edge machines) in an overlay network each build local machine learning (ML) models associated with a particular behavior of interest. Through a communication mechanism, nodes exchange some portion of their ML models between or among each other. The portion of the local model that is exchanged with one or more other nodes encodes or encapsulates relevant knowledge (learned at the source node) for the particular behavior of interest; in this manner, relevant transfer learning is enabled such that individual node models become smarter. Sets of machines that collaborate converge their models toward a solution that is then used to facilitate another overlay network function or optimization. The local knowledge exchange among the nodes creates an emergent behavioral profile used to control the edge machine behavior. Example functions managed with this ML front-end include predictive pre-fetching, anomaly detection, image management, forecasting to allocate resources, and others.
1 . A method carried out in an overlay network comprising a shared infrastructure, the shared infrastructure comprising a set of edge network machines, comprising:
collecting data at each of first and second edge network machines, the first and second edge machines comprising a first subset of the set of edge network machines, the collected data representing information observed by the respective edge network machine;
at each of the first and second edge network machines, building a machine learning model from the data collected;
at the first edge network machine, receiving information from the second edge network machine, the information associated with the machine learning model built by the second edge network machine;
at the first edge network machine, augmenting its machine learning model with the information received from the second edge network machine to generate an augmented machine learning model that converges to a first emergent knowledge solution uniquely associated to the first subset of the set of edge network machines, the augmented machine learning model being distinct from a variant machine learning model built by a second subset of the set of edge network machines, the variant machine learning model built from the collected data representing the information as observed by the respective edge network machines in the second subset and that converges to a second emergent knowledge solution uniquely associated to the second subset of the set of edge network machines; and
performing an overlay network optimization using the augmented machine learning model, wherein the overlay network optimization is one of: predictive pre-fetching, anomaly detection, image management, and forecasting to allocate resources;
where in building the augmented machine learning model and the variant machine learning model, the edge network machines in the first and second subsets operate on an ad hoc basis and without central control.
2 . The method as described in claim 1 wherein the overlay network optimization is a content pre-fetching operation carried out at the first edge network machine.
3 . The method as described in claim 1 wherein the information is received from the second edge machine via a gossip protocol.
4 . The method as described in claim 1 wherein the first and second edge network machines are co-located.
5 . The method as described in claim 1 wherein the first and second edge network machines are not physically co-located but are close to one another in a network sense.
6 . The method as described in claim 1 wherein the overlay network is a content delivery network (CDN) managed by a service provider.
7 . The method as described in claim 1 wherein the information is received from the second edge machine periodically.
8 . The method as described in claim 1 wherein the machine learning model is one of: a K-nearest neighbor model, and a neural network.
9 . An overlay network, comprising:
a set of machine processes associated together, each of the machine processes building a knowledge model local to that machine process, the set of machine processes participating in selective transfer learning among each other on an ad hoc basis and without central control, wherein individual knowledge models local to a particular process operating in a first subset of machine processes converge to a first emergent knowledge solution, and wherein individual knowledge models local to the particular process operating in a second subset of machine processes converge to a second emergent knowledge solution, the second emergent knowledge solution being a variant of the first emergent knowledge solution; and
performing an overlay network function or optimization by applying the first or second emergent knowledge solution;
wherein the set of machine processes are software executing on hardware.
10 . The overlay network as described in claim 9 wherein the overlay network function or optimization is one of: predictive pre-fetching, anomaly detection, image management, forecasting to allocate resources, mapping control and others.
11 . The overlay network as described in claim 9 wherein the transfer learning is facilitated using a gossip protocol or other group communication mechanism.
12 . The overlay network as described in claim 9 wherein a knowledge model is a machine learning classification algorithm.
13 . The overlay network as described in claim 9 wherein first and second of the set of machine processes are co-located.
14 . The overlay network as described in claim 9 wherein first and second of the set of machine processes are not necessarily physically co-located but are close to one another in a network sense.
15 . A distributed computing system, comprising:
machines and associated processes that collaborate amongst each other on an ad hoc basis and without central control with respect to a behavior of interest to converge local knowledge at each machine toward a model of emergent behavior of interest across first and second subsets of the machines, wherein the model of emergent behavior of interest across the first subset of the machines is a variant of the model of emergent behavior of interest across the second subset of the machines; and
a particular machine that applies the model of emergent behavior of interest to control a networking function or optimization local to the particular machine.
16 . The distributed computing system as described in claim 15 wherein the machines are edge servers in a content delivery network.
17 . The distributed computing system as described in claim 16 wherein the networking function or optimization local to the particular machine is a content pre-positioning operation.