# US 2021/0279585 A1 · Method, apparatus, and system for progressive training of evolving machine learning architectures

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0279585 A1
- Document ID: US20210279585A1
- Status: Published application
- Kind: A1
- Application number: 16/809,243
- Filed: 2020-03-04
- Published: 2021-09-09
- Earliest claimed: 2020-03-04
- Assignees: HERE Global B.V.
- Applicants: HERE Global B.V.
- Inventors: Tero Juhani KESKI-VALKAMA
- Web page: https://patentgrep.com/patent/US20210279585A1

## CPC

- G06N 3/082: modifying the architecture, e.g. adding, deleting or silencing nodes or connections
- G06F 18/217: {Validation; Performance evaluation; Active pattern learning techniques}
- G06F 18/24765: {Rule-based classification}
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- G06N 3/045: Combinations of networks
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

An approach is provided for progressive training of long-lived, evolving machine learning architectures. The approach involves, for example, determining alternative paths for the evolution of the machine learning model from a first architecture to a second architecture. The approach also involves determining one or more migration step alternatives in the alternative paths. The migration steps, for instance, include architecture options for the evolution of the machine learning model. The approach further involves processing data using the options to determine respective model performance data. The approach further involves selecting a migration step from the one or more migration step alternatives based on the respective model performance data to control a rate of migration steps over a rate of training in the evolution of the machine learning model. The approach further involves initiating a deployment the selected migration step to the machine learning model.

## Claims

### Claim 1

A computer-implemented method for an evolution of a long-lived machine learning model over multiple architectures comprising:
determining a plurality of alternative paths for the evolution of the machine learning model from a first architecture to a second architecture;
determining one or more migration step alternatives in the plurality of alternative paths, wherein the one or more migration step alternatives include one or more architecture options for the evolution of the machine learning model;
processing data using the one or more architecture options to determine respective model performance data;
selecting a migration step from the one or more migration step alternatives based on the respective model performance data; and
initiating a deployment the selected migration step to the machine learning model.

### Claim 2 (depends on 1)

The method claim 1, wherein the selecting of the migration step is further based on controlling a rate of migration steps over a rate of training in the evolution of the machine learning model.

### Claim 3 (depends on 1)

The method of claim 1, wherein the one or more migration step alternatives include adding or removing a neuron, adding or removing a neuronal layer, changing an activation function of the neuron, or a combination thereof of the machine learning model.

### Claim 4 (depends on 1)

The method of claim 1, wherein a training of the machine learning model continues as the first architecture is migrated to the second architecture.

### Claim 5 (depends on 1)

The method of claim 1, further comprising:
based on determining that the respective model performance data does not meet a model performance threshold, performing at least one of:
reducing the rate of migration steps;
reverting the deployment of the selected migration step;
redeploying the selected migration step with a reduced rate of change; and
applying the reduced rate of change for one or more subsequent migration steps.

### Claim 6 (depends on 1)

The method of claim 1, wherein the one or more migration step alternatives include adding an input to the machine learning model, and wherein the new input is added by adding the new input as a new connection in the machine learning model with a designated initial weight.

### Claim 7 (depends on 1)

The method of claim 1, wherein the one or more architecture options include removing an input of the machine learning model, and wherein the input is removed by incrementally adding noise to an associated input signal.

### Claim 8 (depends on 7)

The method of claim 7, wherein the input is removed by incrementally adding noise to an associated input signal.

### Claim 9 (depends on 7)

The method of claim 7, wherein the input is removed based on determining the noise is pure noise or has been incremented to greater than a threshold level.

### Claim 10 (depends on 1)

The method of claim 1, wherein the one or more architecture options include adding a new output of the machine learning model.

### Claim 11 (depends on 10)

The method of claim 10, wherein the new output is added next to an old output of the machine learning model.

### Claim 12 (depends on 11)

The method of claim 11, further comprising:
changing one or more loss functions, wherein the one or more loss functions are weighted so that a new weight of a new loss associated with the new output is at a first designated level and an old weight of an old loss associated with the old output is at a second designated level, wherein the new weight and the old weight are interpolated through time so that that the new weight reaches the second designated level and the old weight reaches the first designated level; and
removing the old output based on determining that the old weight reaches the first designated level.

### Claim 13 (depends on 1)

The method of claim 1, further comprising:
changing one or more activation functions in one or more neuronal units of the machine learning model by interpolating between an old activation function and a new activation function.

### Claim 14

An apparatus comprising:
at least one processor; and
at least one memory including computer program code for one or more programs,
the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus to perform at least the following,
identify one or more migration step alternatives configured to migrate a first architecture of a machine learning model to a second architecture of the machine learning model;
select a migration step from the one or more migration step alternatives; and
migrating the first architecture to the second architecture by applying the migration step to the first architecture based on a rate of change.

### Claim 15 (depends on 14)

The apparatus of claim 14, wherein the apparatus is further caused to:
process data using the one or more architecture options to determine respective model performance data.

### Claim 16 (depends on 15)

The apparatus of claim 15, wherein the apparatus is further caused to:
select the rate change based on the respective model performance data.

### Claim 17 (depends on 15)

The apparatus of claim 15, wherein the rate of change is based on a number of training steps between a migration step and a subsequent migration step.

### Claim 18

A non-transitory computer-readable storage medium, carrying one or more sequences of one or more instructions which, when executed by one or more processors, cause an apparatus to perform:
determining a plurality of alternative paths for the evolution of the machine learning model from a first architecture to a second architecture;
determining one or more migration step alternatives in the plurality of alternative paths, wherein the one or more migration step alternatives include one or more architecture options for the evolution of the machine learning model;
processing data using the one or more architecture options to determine respective model performance data;
selecting a migration step from the one or more migration step alternatives based on the respective model performance data; and
initiating a deployment the selected migration step to the machine learning model.

### Claim 19 (depends on 18)

The computer-readable storage medium of claim 18, wherein the selecting of the migration step is further based on controlling a rate of migration steps over a rate of training in the evolution of the machine learning model.

### Claim 20 (depends on 18)

The computer-readable storage medium of claim 18, wherein the one or more migration step alternatives include adding or removing a neuron, adding or removing a neuronal layer, changing an activation function of the neuron, or a combination thereof of the machine learning model.
