# US 2019/0318270 A1 · Memory-based data selection scheme for machine learning training on limited memory resources

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

- Number: US 2019/0318270 A1
- Document ID: US20190318270A1
- Status: Published application
- Kind: A1
- Application number: 15/953,440
- Filed: 2018-04-14
- Published: 2019-10-17
- Earliest claimed: 2018-04-14
- Assignees: International Business Machines Corporation
- Applicants: International Business Machines Corporation
- Inventors: Celestine Duenner; Thomas P. Parnell; Charalampos Pozidis
- Web page: https://patentgrep.com/patent/US20190318270A1

## CPC

- G06N 20/00: Machine learning
- G06F 9/5016: {the resource being the memory}
- G06N 7/01: Probabilistic graphical models, e.g. probabilistic networks
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A method for a machine learning model training is provided which operates in a mixed CPU/GPU environment. The amount of general processing unit memory is larger than the amount of special processing unit memory. The method includes loading a complete training data set into the memory of the general processing unit, determining importance values relating to training data vectors in the provided training data set of the training data vectors, dynamically transferring training data vectors of the training data set from the general processing unit memory to a special processing unit memory using as decision criteria the importance value of the training data vector, wherein the importance value used is taken from an earlier training round of the machine learning model, and executing a training algorithm on the special processing unit with the training data vectors having the highest available importance values of one of the earlier training rounds.

## Claims

### Claim 1

A method for training of a machine learning model in a mixed environment comprising a general processing unit and a special processing unit, wherein an amount of general processing unit memory directly accessible by said general processing unit is larger, than an amount of special processing unit memory directly accessible by said special processing unit, said method comprising:
loading a complete training data set into said general processing unit memory;
determining importance values relating to training data vectors in said provided training data set of said training data vectors;
dynamically transferring training data vectors of said training data set from said general processing unit memory to a special processing unit memory using as decision criteria said importance value of said training data vector, wherein said importance value used is taken from an earlier training round of said machine learning model; and
executing a training algorithm on said special processing unit with said training data vectors having highest available importance values of one of said earlier training rounds.

### Claim 2 (depends on 1)

The method according to claim 1, wherein said importance value of a training data vector is a result of a duality gap determination of said training data vector in light of training data set for a machine learning model.

### Claim 3 (depends on 1)

The method according to claim 1, wherein said importance value of training data vectors is determined for randomly identified training data vectors of said training data set.

### Claim 4 (depends on 1)

The method according to claim 1, wherein said special processing unit has a higher processing throughput than said general processing unit.

### Claim 5 (depends on 1)

The method according to claim 1, wherein said special processing unit is a graphic processing unit or a field programmable gate array.

### Claim 6 (depends on 1)

The method according to claim 1, wherein said amount of said special processing unit memory is smaller than required for said training data set.

### Claim 7 (depends on 1)

The method according to claim 1, wherein only those training data vectors are transferred from said general processing unit memory to said special processing unit memory only if they are not available in said special processing unit memory.

### Claim 8 (depends on 1)

The method according to claim 1, further comprising transferring back results of said executed training algorithm from said special processing unit to said general processing unit.

### Claim 9 (depends on 1)

The method according to claim 1, wherein said graphic processing unit finishes a training algorithm using said dynamically transferred training data vectors before said general processing unit has finished said determination of said importance values of all training data vectors.

### Claim 10 (depends on 1)

The method according to claim 1, wherein a random set of training data vectors is transferred from said general processing unit to said special processing unit if no importance values are available for said training data vectors.

### Claim 11

A system for training of a machine learning model, said system comprising a general processing unit and a special processing unit, wherein an amount of general processing unit memory directly accessible by said general processing unit is larger than an amount of special processing unit memory directly accessible by said special processing unit, said system comprising:
a loading unit adapted for loading a complete training data set into said general processing unit memory;
an importance determination unit adapted for determining importance values relating to training data vectors in said provided training data set of said training data vectors;
a transfer unit adapted for dynamically transferring training data vectors of said training data set from said general processing unit memory to a special processing unit memory using as decision criteria said importance value of said training data vector, wherein said importance value used is taken from an earlier training round of said machine learning model; and
wherein said special processing unit is adapted for executing a training algorithm on said special processing unit with said training data vectors having said highest available importance values of one of said earlier training round.

### Claim 12 (depends on 11)

The system according to claim 11, wherein said importance determination unit is adapted for determining said importance value of a training data vector as a result of a duality gap determination of said training data vector in light of training data set for a machine learning model.

### Claim 13 (depends on 11)

The system according to claim 11, wherein said importance determination unit is adapted for randomly selecting training data vectors for determining a related importance value.

### Claim 14 (depends on 11)

The system according to claim 11, wherein said special processing unit has a higher processing throughput than said general processing unit.

### Claim 15 (depends on 11)

The system according to claim 11, wherein said special processing unit is a graphic processing unit or a field programmable gate array.

### Claim 16 (depends on 11)

The system according to claim 11, wherein a storage capacity of said special processing unit memory is smaller than required for the training data set.

### Claim 17 (depends on 11)

The system according to claim 11, wherein said transfer unit is adapted for only transferring those training data vectors from said general processing unit memory to said special processing unit memory only if they are not available in said special processing unit memory.

### Claim 18 (depends on 11)

The system according to claim 11, further comprising:
a transferring back unit adapted for transferring back results of said executed training algorithm from said special processing unit to said general processing unit.

### Claim 19 (depends on 11)

The system according to claim 11, wherein said special processing unit is adapted to finish a training algorithm using said dynamically transferred training data vectors before said determination of said importance values of all training data vectors is performed by said general processing unit.

### Claim 20

A computer program product for training of a machine learning model in a mixed environment comprising a general processing unit and a special processing unit, wherein an amount of general processing unit memory, directly accessible by said general processing unit, is larger than an amount of special processing unit memory directly accessible by said special processing unit, said computer program product comprising a computer readable storage medium having program instructions embodied therewith, said program instructions being executable by one or more computing systems to cause said one or more computing systems:
to load a complete training data set into said general processing unit memory;
to determine importance values relating to training data vectors in said provided training data set of said training data vectors;
to transfer dynamically training data vectors of said training data set from said general processing unit memory to a special processing unit memory using as decision criteria said importance value of said training data vector, wherein said importance value used is taken from an earlier training round of said machine learning model; and
to execute a training algorithm on said special processing unit with said training data vectors having highest available importance values of one of the earlier training rounds.
