Content for  TR 22.874  Word version:  18.2.0

A brief story of machine learning is nothing but a computer (that has no or limited imprinted programs for a certain task) exploiting its own capability ("performance") towards a certain task using data ("experience"). There are several criteria to classify the types of Machine Learning depending on the characteristics of the method used. This use case is intended to describe a case of multi-agent multi-device ML operations with heavy data (i.e., the data size is huge) when there is partial or total disturbance for data collection/transfer (e.g., privacy regulation or temporary technical limitation like shortage of network resources or temporary failure). As depicted in Figure 7.3.1-1, this use case, Part I, is specifically related to a scenario that there are multiple agents and multiple collecting devices where the devices can perform ML operations, not necessarily in full but as much as they can (i.e., functional splitting is possible between a device and one or more learning agents).

In the age that privacy was not affecting the flow of data from the source to the learning agent (e.g., a computer), the expected performance is the outcome of all possible computational considerations of the data collected (e.g., refer to the green solid curve of Figure 7.3.1-2). However, if there is a certain level of disturbance in data collection, the achievable performance toward the given task would not be as good as the one with no such disturbances (refer to the gap between the green solid curve and blue dotted curve in the same Figure). Some examples of such disturbances include:

1. privacy regulations, such as EU's General Data Protection Regulation (GDPR) or California Consumer Privacy Act (CCPA);
2. limited capability of transport-layers, such as lack of network resources (e.g., radio resources due to temporal degradation, higher noise/interference level, highly crowed situations, partial/total break-down, and so on) preventing input data from being delivered in time:
   1. case 1: a portion of input data delivered in time, if any, is still useful (i.e., entropy can get increased/ improved)
   2. case 2: a portion of input data delivered in time, if any, is not useful (i.e., it's not enough to get entropy increased/ improved)

It is commonly understood that the more data a learning model utilizes the better performance the learning model can achieve (assuming the data are reasonably independent from statistics perspectives or sufficiently correctly labelled (when supervised learning is concerned)), if not too large. However, if there is some disturbance in data collection/transfer, such as regional regulations or technical limitations (as described above), the expected performance would not be as good as the case without the disturbance (the vertical difference in Figure 7.3.1-2); in addition, it is expected that the learning model would need to take more time to accumulate "experience" with the reduced feeding-rate of learning data caused by such disturbance (the horizontal difference in Figure 7.3.1-2).

Figure 7.3.1-3 shows an example of possible preparation action that can be taken in UE side if some predictive information can be made available when disturbance is about to happen. In the Figure, the preferred deadline for input data transfer is 1 sec (t = t0+1) with the amount of useful "input data" 3 bits in which two different kinds of scheduling are given: (a) is imperfect scheduling whereas (b) is good scheduling as an example, respectively. In reality, the transfer payload type is not limited to "input data" for learning agent(s) and it can also be applied to "learning model" transfer. The transfer direction can be "uplink" (e.g., for input data transfer) or "downlink" (e.g., for model distribution/transfer). This simplified example is intended to explain the justification why new technical requirements would be needed especially when some disturbance exists (e.g., by regulatory or technical causes).

Given the regulatory disturbance, it is intended to minimize (or to prepare to minimize) the impact on transfer caused by technology (e.g., scheduling and/or information necessary for 3GPP entity to perform "good scheduling"). The expected requirements and service flow description are as follows.

There are three UEs M1, M2 and M3 (computers or learning machines in the form of a UE). There are two learning agents/servers in the cloud. UEs M1, M2 and M3 collect data and they process the data for learning they have collected or are collecting, if available, but they don't have to complete the processing due to limited computational capability. Each UE (M1, M2 and M3, respectively) has a functional splitting point negotiated with their agent(s) regarding data processing for learning. Agent A1 working with agent A2 for a task can share its outcome with agent A2 so that the outcome from agent A2 can jointly improve the outcome, if possible, which is possibly better than A1's individual outcome and A2's individual outcome. UE M1 and UE M2 are located in Area 1 of some jurisdiction that doesn't restrict collecting certain type of data. UE M1 is connected to agent A1 (via eNB, ng-eNB, or gNB) when data connection is necessary (e.g., when needed to upload some data or when needed to download some model).

• Agent A1 provides UE M1 with an alternative agent (i.e., Agent A2) for the use of disturbance, which is (one of) participating agent(s) that agent A1 shares data. [See Description clause for typical types of data]
• UE M1 is transferring learning data to agent 1.

UE M2 is connected to agent A1.

• Agent A1 provides UE M2 with an alternative agent (i.e., Agent A2) for the use of disturbance, which is (one of) participating agent(s) that agent A1 shares data.
• UE M2 is transferring learning data to agent 1.

UE M3 is connected to agent A2.

• Agent A2 provides UE M3 with an alternative agent (i.e., Agent A1) for the use of disturbance, which is (one of) participating agent(s) that agent A2 shares data.
• UE M3 is transferring learning data to agent 2.

While transferring learning data to agent A1, UEs M1 and M2 move into Area 2 of some jurisdiction that has restrictive regulations for agent A1 to collect data from UEs within a specific area (or outside a specific area). As a result, UEs M1 and M2 are restricted to transfer their data to agent 1. While transferring learning data to agent A2, UE M3 moves into a different area where 5G system provides prior notification on possible traffic congestion which might disturb UE M3 from keeping transferring learning data to agent A2. UE M3 makes a selection of an action policy: (action 1) to defer the transfer or (action 2) to request to speed up the transfer.

• If action 1 is selected, UE M3 will resume transferring when it becomes available (not the main focus of this use case);
• If action 2 is selected, UE M3 will be able to (2a) more urgent/useful segment of data (send priority one over the other) and/or (2b) request more network resources to use.

UE M1 attempts to get connected to agent A2 based on the information that agent A1 has provided when initially connected.

• UE M1 keeps transferring the data to agent A2
• Agent A1 can share the collected data or its processed form of data, vice versa

UE M2 is connected to agent A2

• UE M1 keeps transferring the data to agent A2
• Agent A1 can share the collected data or its processed form of data, vice versa

UE M3 is allowed to use more network resources to speed up the transfer to agent A2 (case 2b).

In Table 7.3.6-1, the service-level implications of AI/ML operation are summarized in the third column of Table 7.3.6-1, which are used to derive high-level potential service requirements in communication layer in order to support the AI/ML operation.

Table 7.3.6-2 provides a summary of performance requirements for different usage scenarios. The required KPIs are dependent upon usage scenarios, especially on the task splitting points of given tasks even under the same usage scenarios. The calculation procedure can be referred to in [36]-[39]. For an example of image size 32 x 32 x 3 (32 wide, 32 high, 3 depth/colour channels), the weight is 3072; for images with more respectable size 200 x 200 x 3 = 120,000 weights; For a simple ConvNet for CIFAR-10 classification, the regular Neural Network architecture is INPUT→CONV→RELU→POOL→FC (Input layer, convolutional layer, pooling layer and fully-connected layer). EXAMPLE 1: EXAMPLE 2 (Language understanding): EXAMPLE 3 (Language understanding): EXAMPLE 4: