EP4662597A1 - Method to train a decision module model in a split point range architecture - Google Patents

Abstract

A decision module determines a split point of a deep neural network model. One device implements the DNN model up to the split point. The first device sends intermediate data to a second device which implements the remainder of the DNN model after the split point. The second device sends further data back to the first device. In an embodiment, the first and second devices are user equipment and an edge server. The DNN model can be learned before execution or learned through execution. Asynchronous or synchronous metrics collection is implemented.

Description

METHOD TO TRAIN A DECISION MODULE MODEL IN A SPLIT POINT

RANGE ARCHITECTURE

CROSS REFERENCE TO RELATED APPLICATION

This application claims the benefit of European Serial No. 23315025.9, filed February 10, 2023, which is incorporated by reference herein in its entirety.

TECHNICAL FIELD

At least one of the present embodiments generally relates to a method or an apparatus to implement an efficient decision module in a split point range architecture.

BACKGROUND

Split computing or distributed inference aims at achieving several goals:

The computing load is distributed across the mobile device and edge/cloud server. The energy consumption is shared between the mobile device and the remote server. The privacy is preserved by transmitting partially processed data rather than transmitting raw data. A user-task specific can be handled on the mobile side, while the generic part can be done on the server side.

The DNN model is first partitioned into multiple parts according to the current system environment factors such as network bandwidth, device resources (memory, power, processing units) and edge server workload.

The UE device executes the DNN model up to a specific layer and sends the intermediate data to the edge server. The edge server will execute the remaining layers and sends the prediction results back to the device. Conversely, the server can execute the first layers and then the intermediate data to the mobile for processing the remaining layers.

SUMMARY

At least one of the present embodiments generally relates to a method or an apparatus in the context of split point range architecture. In particular, one objective of the described embodiments is enabling a smart and automatic change of the split point, and thus a smart and automatic change of the subpart running on UE and network part.

According to a first aspect, there is provided a method. The method comprises steps for executing at least a portion of a deep neural network (DNN) up to a specific layer on a first device; sending intermediate data from the first device to a second device; executing remaining layers of the DNN on the second device; and, sending information from the second device to the first device.

According to another aspect, there is provided an apparatus. The apparatus comprises a processor. The processor can be configured to implement the general aspects by executing any of the described methods.

According to another aspect, there is provided an apparatus, configured to perform executing at least a portion of a deep neural network (DNN) up to a specific layer on a first device; sending intermediate data from the first device to a second device; executing remaining layers of the DNN on the second device; and, sending information from the second device to the first device.

According to another aspect, there is provided a decision module for determining at least one split point indicative of a specific layer of a deep neural network model for execution among a plurality of devices.

According to another aspect, there is provided a device configured to receive intermediate data from another device and execute a portion of a deep neural network model, said portion indicated by at least one split point input.

According to another general aspect of at least one embodiment, there is provided a device comprising an apparatus according to any of the embodiments; and at least one of (i) an antenna configured to receive a signal, the signal including the video block, (ii) a band limiter configured to limit the received signal to a band of frequencies that includes the video block, or (iii) a display configured to display an output representative of a video block.

According to another general aspect of at least one embodiment, there is provided a non-transitory computer readable medium containing data content generated according to any of the described encoding embodiments or variants.

According to another general aspect of at least one embodiment, there is provided a signal comprising video data generated according to any of the described encoding embodiments or variants.

According to another general aspect of at least one embodiment, a bitstream is formatted to include data content generated according to any of the described encoding embodiments or variants.

According to another general aspect of at least one embodiment, there is provided a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out any of the described decoding embodiments or variants. These and other aspects, features and advantages of the general aspects will become apparent from the following detailed description of exemplary embodiments, which is to be read in connection with the accompanying drawings.

According to another general aspect of at least one embodiment, there is provided a non-transitory computer readable medium containing data content comprising instructions to perform any of the encoding or decoding methods.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 illustrates AI/ML model splitting where a UE runs the first part of the model (left) and the edge/cloud server runs the first part of the model (right).

Figure 2 illustrates a block diagram of a local decision module.

Figure 3 illustrates an example of decision module I/O.

Figure 4 illustrates an overview of reinforcement learning.

Figure 5 illustrates an example of decision module based on reinforcement learning - evaluation on UE side.

Figure 6 illustrates an example of a variant of a decision module based on reinforcement learning - evaluation on UE side.

Figure 7 illustrates an example of a decision module based on reinforcement learning - evaluation on server side.

Figure 8 illustrates an example of decision module model training.

Figure 9 illustrates one embodiment of contextual metrics transmission: UE data source.

Figure 10 illustrates one embodiment of contextual metrics transmission: network data source.

Figure 11 illustrates one embodiment of contextual metrics transmission: UE data source with first processing in network.

Figure 12 illustrates one embodiment of 3GPP sequence flow diagram - synchronous exchange.

Figure 13 illustrates one embodiment of a 3GPP sequence flow diagram - asynchronous exchange, machine learning solution.

Figure 14 illustrates one embodiment of 3GPP sequence flow diagram - asynchronous exchange, reinforcement learning, reward on UE side solution.

Figure 15 illustrates one embodiment of 3GPP sequence flow diagram - asynchronous exchange, reinforcement learning, reward on network side solution. Figure 16 illustrates one embodiment of a method for executing a DNN model using the described embodiments.

Figure 17 illustrates another embodiment of a method for executing at least a portion of a DNN model at a device using the described embodiments.

Figure 18 illustrates one embodiment of an apparatus for executing at least a portion of a DNN model using the described embodiments.

Figure 19 illustrates a standard, generic video compression scheme.

Figure 20 illustrates a standard, generic video decompression scheme.

Figure 21 illustrates a processor-based system for encoding/decoding under the general described aspects.

DETAILED DESCRIPTION

Distributed inference consists in splitting a DNN model in at least two sub-parts. Sub-parts are distributed towards the network infrastructure nodes being the UEs, the edge servers and the cloud servers, this is illustrated in Figure 1.

If a system architecture is defined that enables a smart and automatic change of the split point, and thus a smart and automatic change of the subpart running on UE and network part.

Messages between UE and the network side enabling this change have also been defined.

On the defined architecture the decision to change the split point relies on a module called “local decision module”. However, this “local decision module” is described as a black-box: input and outputs are described, but no details on the implementation are provided here as there are several methods possible to communicate the change of a split point between the UE and the network side.

The following embodiments address the implementation of a local decision module to the architecture of, for example, Figure 2.

Technical problems to solve to get a smart decision by the local decision module are the following.

The number of parameters influencing the decision are potentially very high and fluctuating in time: a solution based on static rules established by an expert will be complicated to set-up and give inefficient results as soon as the system encounters new environment related to either the UE resources, the input data source, or the network resources. The number of models to manage may be very high making the possibility of defining the rules by experts too heavy a task.

Optimization choices are multiple: goals may be to optimize the latency, the energy consumption on the UE side, the energy consumption on the network side, energy consumption on both sides.

The local decision module must consider the effective quality of its responses and be able to improve them over time.

The local decision module may encounter a new model or new parameters values never seen before.

The solution relies on following ideas:

To use live learning (reinforcement learning) or offline learning (machine learning) to implement an efficient decision module in a split point range architecture.

To propose the creation of metrics enabling the training of the above machine learning and reinforcement solution.

To propose exchange of those metrics by two ways: synchronously and asynchronously.

The decision module takes in input:

• The AI/ML model to be used by the application.

• The set of envisaged/possible split points for this AI/ML model.

• UE Local information on which the first part of the model will be inferred (device

1 eg UE): o Available Memory o CPU/GPU/TPU used (Processing Unit)

• Edge/network Local information on which the second part of the module will be inferred (device 2 - e.g. Edge server): o Available memory o CPU/GPU/TPU used (Processing Unit)

• Input data characteristics o Resolution (width, height) o Dimension (8bits, 16bits, 32bits, ...) o Frame rate (25FPS, 30FPS, ...)

• Optimization choices: o Minimize UE energy cost o Minimize network energy cost o Minimize total energy cost o Minimize latency o Maximize accuracy o Best compromise combining energy cost I latency I prediction accuracy Output is Best Split point and/or expected performance per split point:

• Expected latency

• UE Expected energy cost

• Network expected energy cost

• Total expected energy cost

• AI/ML Model Prediction result accuracy

This is a problem that can be solved by a Machine Learning algorithm. There are two important characteristics to consider in this machine learning algorithm for this problem: 1 ) when learning takes place, live or offline and 2) what type of learning is used, supervised or reinforcement learning.

Live or offline learning

Offline learning or live Learning cannot be opposed, they may depend on the application provider or on the network operator.

In the first case, an AI/ML model that has been beforehand learned and then deployed, in the other an algorithm that learns by experiencing an environment through trial and error.

In the first case, an algorithm that gives relevant results immediately, in the other an algorithm that may needs time to learn.

In the first case, an algorithm that can’t adapt without being updated, in the other an algorithm that continues learning. Even for live learning, an initial model is often first trained offline, deployed and finishes training live.

Reinforcement or supervised learning

Reinforcement Learning and Supervised Learning are the two machine learning techniques that are envisioned for this disclosure. Both can be trained live or offline.

Thus, unsupervised learning is dismissed or out of scope this disclosure, indeed the goal of the Decision Module is not to cluster data or to find any category of data, but rather to render a decision from a set of input variables that will be described below. In both cases the learning algorithm can be deployed on the UE or on the Network.

The supervised learning is based on a Ground Truth. The learning phase consists in a mapping between the input samples data and this Ground Truth. The outcome of this learning phase is a model. This model may be based on a neural network architecture, or on more classical machine learning algorithm like decision tree, random forest. Once deployed on a device, e.g., a UE, the model can be operated via an application. The application feeds this AI/ML model with sample data which are inferred, and the model delivers a prediction result. Thus, the decision module that embeds such a model could deliver a predicted Split Point, i.e. the Split Point that matches the best the input sample data.

In addition, Deep Reinforcement Learning suits well for a personalization of the Split Point selection. Indeed, from one UE to the other, from one user to the other, from one location to the other, from one period to the other and etc ... the choice of the Split Point may differ. If for a given State, the Action of Split Point leads to a good latency (example) then a Reward may be applied. This mechanism can be reproducible for energy.

The reward will depend on the application:

• Online gaming will require very low latency.

• Applications relying on file transfer will require high bandwidth.

• Some services will require constant bandwidth even if it is low, i.e., with no jitter.

Live reinforcement learning variant

A reinforcement learning principle can be seen as in Figure 4.

One of the differences compare to the other system is that the reinforcement learning evaluates the status of the system at each action and computes a reward that reflects the quality of the current situation regarding the final objective. This reward is used by the training algorithm of the agent: the agent will attempt to maximize future rewards. Thus, a quantified evaluation of the performance of the system based on the choice of the agent is needed to compute this reward. How this reward is computed from the performance metrics is a design decision. A few possibilities are described in the described aspects. In the context of this disclosure, the agent corresponds to the main function of the local decision module. Furthermore, this reward must also be based on user preferences to reflect his/her intent.

The reward is linked to the objective of the system as the reward is the function the agent strives to maximize. It is therefore logical to build this reward function from a set of metrics quantifying the performance of the system. These metrics can be a combination of III contextual metrics and network contextual metrics and must evaluate the optimization goals for the system. Based on the optimization goals listed above the contextual metrics could be the followings: UE energy cost, Network energy cost, Latency, and Accuracy.

To ponder the different optimization goals the reward can balance different metrics. For example, the reward could be:

R = - ai UE_energy_cost - 02 Network_energy_cost - 03 latency + cu accuracy where each a_t is 0 or a positive real number and quantifies importance of each optimization goal.

The training algorithm of the agent trains a decision function. The input of this decision function is called the state and it is computed from the set of inputs of the local decision module defined above. For example, the state could be the concatenation of all inputs, a subset of those inputs, any transformation of those inputs or subset of inputs which could include normalization, scaling, removal of outliers and/or processing by neural networks. The state could also consist of a concatenation or a summary of the past inputs of the decision module.

The output of the decision function is the output of the agent: the action, which is the split point of the AI/ML model that is used for split inference.

The agent can use any of the shelf reinforcement learning algorithm such as Q learning, SARSA, UCB, Deep Q Learning, PPO or RAINBOW.

The evaluation of the system may be host on the UE side or in the server side.

With an evaluation of the system on the UE side the decision module can look like Figure 5. In this embodiment the reward is computed by a separate module called “evaluation of the system”. It takes as input the optimization choice and the values necessary to compute the reward. Those values come either from the UE device or from the network and are denoted by the term “contextual metrics” in the figure. Its output is the reward. This reward is an input to the local decision module where the agent resides, in addition to all the inputs listed above. The output of the decision module is the split point (=the action), which is used by the AI/ML model partition module.

Figure 6 shows a different embodiment. With an evaluation of the system on the UE side, the evaluation of the system (= reward computation) may also be located within the Local Decision module itself. Therefore, the inputs of the module evaluating the system are now inputs of the decision module, the reward does not leave the decision module. The two embodiments have different advantages. When the reward module is located outside the decision module, those two modules can be developed independently by different companies so the system is more modular. When the evaluation of the system is located within the agent, the architecture of the system can also be used for supervised learning so the architecture of the system is more generic.

The evaluation of the system may also be located on the network side. Figure 7 illustrates such a possible embodiment.

Interests of such a solution may be to benefit from large resources available at network side compared to the UE, to generalize an evaluation based on information from several UEs. Besides moving the evaluation of the system module from the device to the network, there is one difference in the information sent by the UE to the network: the optimization choice must be included so that it can be used by the environment evaluation module. Sending the network contextual metrics to the device is optional but can be useful for training the decision module. With an evaluation of the system on the network side the decision module looks like Figure 7.

Offline supervised Training - Best split

The input and outputs for the training are the same as defined above.

The target for the loss function during the training is to: minimize the expected latency, minimize the energy cost, and maximize the prediction result accuracy.

A formula to express the best compromise combining energy cost I latency I prediction accuracy could be as:

Let’s name BC, the “Best Compromise”:

BC = ai UE_energy_cost + 02 Network_energy_cost + 03 latency - cu accuracy With H₌₁ a_k = 1

The parameters ai may be either fixed or learned.

The training is made on the server side by a specific module “Model training”.

This module can realize the training by using data collected by the module “Global Metrics Manager”, (see next clause).

The training of the model may be triggered by different thresholds such as time (he model training may be activated at regular intervals), metrics size/number (size or number of new collected contextual metrics not yet integrated in the decision module model), and change of format (update on the input data format or on the output data format).

Once the model is trained it is communicated to the UE by appropriate means.

The choice of the moment to send to communicate the model can be a network operator decision, triggered by some thresholds, such as the percentage of improvement of the new model or the number of updates since the last communication. The model can also be communicated on UE request.

Data collection

To train efficiently the model, latency and energy costs are collected from both sides, UE and network, they can be used to refine the model once it is deployed.

Collection of these information may be done synchronously or asynchronously.

For synchronously collection, on the exchanges between UE and network, two messages need to be modified to also carry “contextual metrics”. First, theone from the UE to the network already carrying the used split point and intermediate data and second, the one from the network to the UE carrying the inference results.

On the first message, UE to the network, these contextual metrics are the latency of the processing on the split model M1 , the energy used for the processing on the split model M1 the battery level of the UE, the memory used at the UE level, the CPU load at the start of the processing and regularly sampled during all of the processing (i.e. , every ms).

For the second message, network to UE, these contextual metrics are thelatency of the processing on the split model M2, the energy used for the processing on the split model M2, the memory used at the network level, or the CPU load at the start of the processing and regularly sampled during all the processing (i.e. every ms). For asynchronously collection, contextual metrics are the same than those described in these two above messages, but this collection relies on a different mechanism.

At each message exchanged between UE and network (the message “split point + intermediate data”) contextual metrics are stored in a local database, host by the module “Local metrics Manager”.

At regular intervals, a part or all of the metrics are sent to the module “Global metrics manager” host in the network side.

At regular intervals or triggered by some threshold (see above clause) the “Global metrics manager” is in capacity to train the model “local decision model”.

The two approaches listed above are among the best solutions to build a split point decision module. Nevertheless, other options are possible. Some examples are as follows.

An offline reinforcement learning approach is possible. Such a model can be trained from a database of observations (decision module inputs, split point and UE/System contextual metrics) and/or using a simulatorortested, using approaches such as bandit algorithms or off-policy or batch reinforcement learning.

An online supervised learning approach that predicts the performance of a given split is possible. Such a training algorithm would observe data points that are generated live when using the system and would train or finetune an initial model.

A supervised learning model predicting the split point directly cannot be trained live because typically the optimal split point is not known. Training such a model would only be possible by either sending models used to our server that would generate training data by extensively testing all split points or by first training a supervised model described in the previous bullet point and using this model to estimate the best split point and using this estimate as a label.

Federated learning (distributed learning over multiple devices) can also be used with any live approaches including reinforcement learning.

Messages

Two kinds of messages can be used to collect and exchange metrics: Synchronous messages based on existing messages between UE and server in charge of the communication of the split point and intermediate data, (“piggybaking messages”. To these existing messages is added a part including the metrics.

Asynchronous messages: on a regular basis of following a specific rule, messages carrying metrics may be exchanged between UEs and server.

Those two kinds of messages relies on a new structure that encompasses metrics.

Metrics wrapper

For example, of a struct for all combination: sender may be UE or Server, first part of the model M1 may have run either on client or on server:

Struct ContextualMetrics {

Int Model_ReflD; # Model Identifier

Int SplitPointID; # Split point

Bit Source; # 0= inference on UE ; 1= inference on server

Int SourcelD; # ID of UE or ID of Server

Bit SplitPart; # 0= inference on first subset M1 ;

# 1 inference on second subset M2

Int InferenceLatencyForSubsetModel; # Inference time of the “splitpart” of the model (ie M1 or M2) on the source (ie UE or server)

Int EstimatedEnergyCostsForThisInference; # Estimated or measured energy consumption for the inference (UE if uplink, Server if downlink)

Int Memory A vailableBeforelnference; # UE or server side

Int MemoryAvailableDuringlnference; # UE or server side

Int CPULoadBeforelnference; # UE or server side

Int CPULoadDuringlnference; # UE or server side

Int GPULoadBeforelnference; # UE or server side

Int GPULoadDuringlnference; # UE or server side

Int NumberOfApplicationsRunningBeforelnference; # UE specific I nt NumberOfApplicationsRunningAfterlnference; # UE specific

Array input_data_characteristics[]; # width, height, nb_bits

(8/10/12), frame rate

Int optimization choice; # optimization choice (0: latency, 1 :minimize

UE energy, 2:minimize network consumption, 3:minimize both UE and network energy consumption, 4: best compromise

}

Synchronous messages

Existing messages from previous approaches can be used with added metrics information.

Intermediate data transport

Intermediate data are wrapped in a structure indicating the origin of the data.

Structure is like:

Struct lntermediateDataWrapper {

Int Model_ReflD; # Model Identifier bit ChangeOfSplitPoint; # 1 if the splitpoint is changed from the previous message

Int SplitPointID; # Split point used by the UE

Int NextSplitpointID; # Next split point used by the UE (optional)

Int IntermediateDataLength; # Length of Intermediate Data

Dim IntermediateDataDim; # Array dimension of intermediate data byte EncodingMethod; # Indicate if data are compressed and by which algorithm int SequenceNumber; # Sequential identifier of input data double TimeStamp; # Timestamp of the intermediate data

Byte lntermediateData[] # Intermediate Data int UEContextualMetricsLength[] # Length of UE contextual

Metrics

ContextualMetrics UEContextualMetricsData[] # UE contextual Metrics

Asynchronous messages

These messages are exchanged between the UE module “Local Metrics manager” and the Server module “Global Metrics Manager” modules.

Structure is like:

Struct AsynchronousContextualMetricsMessages {

Int UEJd ; # UE identifier

Int Server d; # Server identifier

Int metrics_ id ; # Metrics identifier

TimeStamp timestamp ; # Timestamp of the metrics (date and time)

Contextual Metrics metrics ; # metrics

}

Figure 12 shows a call flow based on a prior work for the synchronous exchange of the metrics:

Figure 12 presents a sequence diagram according to current 3GPP SA4 AI/ML components.

The steps under the general aspects described here are the step 6 that is completed to integrate the exchange of the UE contextual metrics, the step 6 that is completed to integrate the exchange of the optimization choice when reinforcement Learning is used, the step 7 that is completed to integrate the exchange of the network contextual metrics, and the step 7 that is completed to integrate the exchange of the reward when reinforcement Learning is used.

Steps 1 and 2 are the AI/ML application request for a dynamic split configuration range to the network service. The range as proposed in a previous section, considers that an AI/ML model (e.g.. “M”) may be dynamically split (e.g. in MO and M1) according to different split points (e.g. A, B, C, D, E), thus part running in the UE inference or in the network inference subset. For example, here, split points B and C can be considered on both sides. Following steps 1 ,2, a UE inference can dynamically start processing the capture sequence Frame 1 up to layer A, B or C.

For Step 3, the AI/ML application computes its internal resources and application requirements and, initialize the AI/ML UE subset including split points A, B.

For Steps 4 and 5, the inference engine processes the first sequence Frame 1 up to split point B.

In Step 6, when Frame 1 is performed, the UE delivers the output data to the network inference thanks to intermediate data transfer function. Transferred data includes data payload and metadata information useful to identify how to process the received data, for example at least an indication of the current split point.

For these embodiments, metadata information will also include contextual metrics information as described in the clause concerning the Metrics wrapper.

For these embodiments, when reinforcement learning is used, metadata information will also include optimization choice as described in the clause concerning Metrics wrapper.

Step 7: Upon continuous result, for example streaming overlay of the current input video, the inference engine sends back the final result of the AI/ML model. For these embodiments, metadata information will also include network contextual metrics information as described in clause 4.6.1 Metrics wrapper.

For these embodiments, when reinforcement learning is used and when the evaluation of the system is on the network side, metadata information will also include contextual metrics information as described in clause 4.6.1 Metrics wrapper.

In Step 8, depending on different internal triggers (e.g., processing power decreasing), the UE makes use of the flexibility to process different split points and, locally update the new split point from B -> A starting from sequence Frame 2 to the UE inference.

Steps 9,10, 11 follow a similar process to steps 4,5,6, except that UE delivers data at the split point A instead of B.

In Step 12, the network inference engine computes the split point information and change input to split point B as such. On another embodiment, the UE may trigger the network inference engine when receiving the AI/ML application update. Step 13 is similar to step 7. This shows a seamless split model processing.

Asynchronous exchange

For asynchronously collection, contextual metrics are the same than those described in these two above messages, but this collection relies on a different mechanism.

At each message exchanged between UE and network (the message “split point + intermediate data”) contextual metrics are stored in a local database, host by the module “Local metrics Manager”. In case of Reinforcement Learning, optimization choice is also stored.

At regular intervals, a part or all of the metrics are sent to the module “Global metrics manager” host in the network side.

In case of a machine learning based solution, at regular intervals or triggered by some threshold (see above clause) the “Global metrics manager” is in capacity to train the model “local decision model”.

In case of a Reinforcement learning based solution, at regular intervals or triggered by some threshold (see above clause) the “RL Evaluation of the system” module at network communicates the reward to the UE agent.One embodiment of a method 1600 for executing a learned DNN model is shown in Figure 16. The method commences at Start bock 1601 and proceeds to block 1610 for executing at least a portion of a deep neural network (DNN) up to a specific layer on a first device. Control proceeds from block 1610 to block 1620 for sending intermediate data from the first device to a second device. Control proceeds from block 1620 to block 1630 for executing remaining layers of the DNN on the second device. Control proceeds from block 1630 to block 1640 for sending information from the second device to the first device.

One embodiment of a method 1700 for executing a portion of a DNN model is shown in Figure 17. The method commences at Start block 1701 and proceeds to block 1710 for receiving intermediate data and split point information. Control proceeds from block 1710 to block 1720 for executing a portion of a DNN model based on the split point information. Control proceeds from block 1720 to block 1730 for sending results of the execution to another device.

Figure 18 shows one embodiment of an apparatus 1000 for performing any of the aforementioned methods. The apparatus comprises Processor 1010 and can be interconnected to a memory 1020 through at least one port. Both Processor 1010 and memory 1020 can also have one or more additional interconnections to external connections.

Processor 1810 is also configured to either insert or receive information in a bitstream or in data and, executing at least a portion of a DNN model using the aforementioned methods.

The embodiments described here include a variety of aspects, including tools, features, embodiments, models, approaches, etc. Many of these aspects are described with specificity and, at least to show the individual characteristics, are often described in a manner that may sound limiting. However, this is for purposes of clarity in description, and does not limit the application or scope of those aspects. Indeed, all of the different aspects can be combined and interchanged to provide further aspects. Moreover, the aspects can be combined and interchanged with aspects described in earlier filings as well.

The aspects described and contemplated in this application can be implemented in many different forms. Figures 19, 20, and 21 provide some embodiments, but other embodiments are contemplated and the discussion of Figures 19, 20, and 21 does not limit the breadth of the implementations. At least one of the aspects generally relates to video encoding and decoding, and at least one other aspect generally relates to transmitting a bitstream generated or encoded. These and other aspects can be implemented as a method, an apparatus, a computer readable storage medium having stored thereon instructions for encoding or decoding video data according to any of the methods described, and/or a computer readable storage medium having stored thereon a bitstream generated according to any of the methods described.

In the present application, the terms “reconstructed” and “decoded” may be used interchangeably, the terms “pixel” and “sample” may be used interchangeably, the terms “image,” “picture” and “frame” may be used interchangeably. Usually, but not necessarily, the term “reconstructed” is used at the encoder side while “decoded” or “reconstructed” is used at the decoder side.

Various methods are described herein, and each of the methods comprises one or more steps or actions for achieving the described method. Unless a specific order of steps or actions is required for proper operation of the method, the order and/or use of specific steps and/or actions may be modified or combined. Additionally, terms such as “first”, “second”, etc. may be used in various embodiments to modify an element, component, step, operation, etc., such as, for example, a “first decoding” and a “second decoding”. Use of such terms does not imply an ordering to the modified operations unless specifically required. So, in this example, the first decoding need not be performed before the second decoding, and may occur, for example, before, during, or in an overlapping time period with the second decoding.

Various methods and other aspects described in this application can be used to modify modules, for example, the intra prediction, entropy coding, and/or decoding modules (160, 360, 145, 330), of a video encoder 100 and decoder 200 as shown in Figure 19 and Figure 20. Moreover, the present aspects are not limited to WC or HEVC, and can be applied, for example, to other standards and recommendations, whether preexisting or future-developed, and extensions of any such standards and recommendations (including WC and HEVC). Unless indicated otherwise, or technically precluded, the aspects described in this application can be used individually or in combination.

Various numeric values are used in the present application. The specific values are for example purposes and the aspects described are not limited to these specific values.

Figure 19 illustrates an encoder 100. Variations of this encoder 100 are contemplated, but the encoder 100 is described below for purposes of clarity without describing all expected variations.

Before being encoded, the video sequence may go through pre-encoding processing (101 ), for example, applying a color transform to the input color picture (e.g., conversion from RGB 4:4:4 to YCbCr 4:2:0), or performing a remapping of the input picture components in order to get a signal distribution more resilient to compression (for instance using a histogram equalization of one of the color components). Metadata can be associated with the pre-processing and attached to the bitstream.

In the encoder 100, a picture is encoded by the encoder elements as described below. The picture to be encoded is partitioned (102) and processed in units of, for example, CUs. Each unit is encoded using, for example, either an intra or inter mode. When a unit is encoded in an intra mode, it performs intra prediction (160). In an inter mode, motion estimation (175) and compensation (170) are performed. The encoder decides (105) which one of the intra mode or inter mode to use for encoding the unit, and indicates the intra/inter decision by, for example, a prediction mode flag. Prediction residuals are calculated, for example, by subtracting (110) the predicted block from the original image block.

The prediction residuals are then transformed (125) and quantized (130). The quantized transform coefficients, as well as motion vectors and other syntax elements, are entropy coded (145) to output a bitstream. The encoder can skip the transform and apply quantization directly to the non-transformed residual signal. The encoder can bypass both transform and quantization, i.e. , the residual is coded directly without the application of the transform or quantization processes.

The encoder decodes an encoded block to provide a reference for further predictions. The quantized transform coefficients are de-quantized (140) and inverse transformed (150) to decode prediction residuals. Combining (155) the decoded prediction residuals and the predicted block, an image block is reconstructed. In-loop filters (165) are applied to the reconstructed picture to perform, for example, deblocking/SAO (Sample Adaptive Offset) filtering to reduce encoding artifacts. The filtered image is stored at a reference picture buffer (180).

Figure 20 illustrates a block diagram of a video decoder 200. In the decoder 200, a bitstream is decoded by the decoder elements as described below. Video decoder 200 generally performs a decoding pass reciprocal to the encoding pass as described in Figure 19. The encoder 100 also generally performs video decoding as part of encoding video data.

In particular, the input of the decoder includes a video bitstream, which can be generated by video encoder 100. The bitstream is first entropy decoded (230) to obtain transform coefficients, motion vectors, and other coded information. The picture partition information indicates how the picture is partitioned. The decoder may therefore divide (235) the picture according to the decoded picture partitioning information. The transform coefficients are de-quantized (240) and inverse transformed (250) to decode the prediction residuals. Combining (255) the decoded prediction residuals and the predicted block, an image block is reconstructed. The predicted block can be obtained (270) from intra prediction (260) or motion-compensated prediction (i.e., inter prediction) (275). Inloop filters (265) are applied to the reconstructed image. The filtered image is stored at a reference picture buffer (280).

The decoded picture can further go through post-decoding processing (285), for example, an inverse color transform (e.g., conversion from YcbCr 4:2:0 to RGB 4:4:4) or an inverse remapping performing the inverse of the remapping process performed in the pre-encoding processing (101). The post-decoding processing can use metadata derived in the pre-encoding processing and signaled in the bitstream.

Figure 21 illustrates a block diagram of an example of a system in which various aspects and embodiments are implemented. System 1000 can be embodied as a device including the various components described below and is configured to perform one or more of the aspects described in this document. Examples of such devices include, but are not limited to, various electronic devices such as personal computers, laptop computers, smartphones, tablet computers, digital multimedia set top boxes, digital television receivers, personal video recording systems, connected home appliances, and servers. Elements of system 1000, singly or in combination, can be embodied in a single integrated circuit (IC), multiple ICs, and/or discrete components. For example, in at least one embodiment, the processing and encoder/decoder elements of system 1000 are distributed across multiple ICs and/or discrete components. In various embodiments, the system 1000 is communicatively coupled to one or more other systems, or other electronic devices, via, for example, a communications bus or through dedicated input and/or output ports. In various embodiments, the system 1000 is configured to implement one or more of the aspects described in this document.

The system 1000 includes at least one processor 1010 configured to execute instructions loaded therein for implementing, for example, the various aspects described in this document. Processor 1010 can include embedded memory, input output interface, and various other circuitries as known in the art. The system 1000 includes at least one memory 1020 (e.g., a volatile memory device, and/or a non-volatile memory device). System 1000 includes a storage device 1040, which can include non-volatile memory and/or volatile memory, including, but not limited to, Electrically Erasable Programmable Read-Only Memory (EEPROM), Read-Only Memory (ROM), Programmable Read-Only Memory (PROM), Random Access Memory (RAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), flash, magnetic disk drive, and/or optical disk drive. The storage device 1040 can include an internal storage device, an attached storage device (including detachable and non-detachable storage devices), and/or a network accessible storage device, as non-limiting examples.

System 1000 includes an encoder/decoder module 1030 configured, for example, to process data to provide an encoded video or decoded video, and the encoder/decoder module 1030 can include its own processor and memory. The encoder/decoder module 1030 represents module(s) that can be included in a device to perform the encoding and/or decoding functions. As is known, a device can include one or both of the encoding and decoding modules. Additionally, encoder/decoder module 1030 can be implemented as a separate element of system 1000 or can be incorporated within processor 1010 as a combination of hardware and software as known to those skilled in the art.

Program code to be loaded onto processor 1010 or encoder/decoder 1030 to perform the various aspects described in this document can be stored in storage device 1040 and subsequently loaded onto memory 1020 for execution by processor 1010. In accordance with various embodiments, one or more of processor 1010, memory 1020, storage device 1040, and encoder/decoder module 1030 can store one or more of various items during the performance of the processes described in this document. Such stored items can include, but are not limited to, the input video, the decoded video or portions of the decoded video, the bitstream, matrices, variables, and intermediate or final results from the processing of equations, formulas, operations, and operational logic.

In some embodiments, memory inside of the processor 1010 and/or the encoder/decoder module 1030 is used to store instructions and to provide working memory for processing that is needed during encoding or decoding. In other embodiments, however, a memory external to the processing device (for example, the processing device can be either the processor 1010 or the encoder/decoder module 1030) is used for one or more of these functions. The external memory can be the memory 1020 and/or the storage device 1040, for example, a dynamic volatile memory and/or a non-volatile flash memory. In several embodiments, an external non-volatile flash memory is used to store the operating system of, for example, a television. In at least one embodiment, a fast external dynamic volatile memory such as a RAM is used as working memory for video coding and decoding operations, such as for MPEG-2 (MPEG refers to the Moving Picture Experts Group, MPEG-2 is also referred to as ISO/IEC 13818, and 13818-1 is also known as H.222, and 13818-2 is also known as H.262), HEVC (HEVC refers to High Efficiency Video Coding, also known as H.265 and MPEG-H Part 2), or VVC (Versatile Video Coding, a new standard being developed by JVET, the Joint Video Experts Team).

The input to the elements of system 1000 can be provided through various input devices as indicated in block 1130. Such input devices include, but are not limited to, (i) a radio frequency (RF) portion that receives an RF signal transmitted, for example, over the air by a broadcaster, (ii) a Component (COMP) input terminal (or a set of COMP input terminals), (iii) a Universal Serial Bus (USB) input terminal, and/or (iv) a High Definition Multimedia Interface (HDMI) input terminal. Other examples, not shown in Figure 21 , include composite video.

In various embodiments, the input devices of block 1130 have associated respective input processing elements as known in the art. For example, the RF portion can be associated with elements suitable for (i) selecting a desired frequency (also referred to as selecting a signal, or band-limiting a signal to a band of frequencies), (ii) downconverting the selected signal, (iii) band-limiting again to a narrower band of frequencies to select (for example) a signal frequency band which can be referred to as a channel in certain embodiments, (iv) demodulating the downconverted and bandlimited signal, (v) performing error correction, and (vi) demultiplexing to select the desired stream of data packets. The RF portion of various embodiments includes one or more elements to perform these functions, for example, frequency selectors, signal selectors, band-limiters, channel selectors, filters, downconverters, demodulators, error correctors, and demultiplexers. The RF portion can include a tuner that performs various of these functions, including, for example, downconverting the received signal to a lower frequency (for example, an intermediate frequency or a near-baseband frequency) or to baseband. In one set-top box embodiment, the RF portion and its associated input processing element receives an RF signal transmitted over a wired (for example, cable) medium, and performs frequency selection by filtering, downconverting, and filtering again to a desired frequency band. Various embodiments rearrange the order of the above-described (and other) elements, remove some of these elements, and/or add other elements performing similar or different functions. Adding elements can include inserting elements in between existing elements, such as, for example, inserting amplifiers and an analog-to-digital converter. In various embodiments, the RF portion includes an antenna.

Additionally, the USB and/or HDMI terminals can include respective interface processors for connecting system 1000 to other electronic devices across USB and/or HDMI connections. It is to be understood that various aspects of input processing, for example, Reed-Solomon error correction, can be implemented, for example, within a separate input processing IC or within processor 1010 as necessary. Similarly, aspects of USB or HDMI interface processing can be implemented within separate interface les or within processor 1010 as necessary. The demodulated, error corrected, and demultiplexed stream is provided to various processing elements, including, for example, processor 1010, and encoder/decoder 1030 operating in combination with the memory and storage elements to process the datastream as necessary for presentation on an output device.

Various elements of system 1000 can be provided within an integrated housing, Within the integrated housing, the various elements can be interconnected and transmit data therebetween using suitable connection arrangement, for example, an internal bus as known in the art, including the Inter-IC (I2C) bus, wiring, and printed circuit boards.

The system 1000 includes communication interface 1050 that enables communication with other devices via communication channel 1060. The communication interface 1050 can include, but is not limited to, a transceiver configured to transmit and to receive data over communication channel 1060. The communication interface 1050 can include, but is not limited to, a modem or network card and the communication channel 1060 can be implemented, for example, within a wired and/or a wireless medium.

Data is streamed, or otherwise provided, to the system 1000, in various embodiments, using a wireless network such as a Wi-Fi network, for example IEEE 802.11 (IEEE refers to the Institute of Electrical and Electronics Engineers). The Wi-Fi signal of these embodiments is received over the communications channel 1060 and the communications interface 1050 which are adapted for Wi-Fi communications. The communications channel 1060 of these embodiments is typically connected to an access point or router that provides access to external networks including the Internet for allowing streaming applications and other over-the-top communications. Other embodiments provide streamed data to the system 1000 using a set-top box that delivers the data over the HDMI connection of the input block 1130. Still other embodiments provide streamed data to the system 1000 using the RF connection of the input block 1130. As indicated above, various embodiments provide data in a non-streaming manner. Additionally, various embodiments use wireless networks other than Wi-Fi, for example a cellular network or a Bluetooth network.

The system 1000 can provide an output signal to various output devices, including a display 1100, speakers 1110, and other peripheral devices 1120. The display 1100 of various embodiments includes one or more of, for example, a touchscreen display, an organic light-emitting diode (OLED) display, a curved display, and/or a foldable display. The display 1100 can be for a television, a tablet, a laptop, a cell phone (mobile phone), or another device. The display 1100 can also be integrated with other components (for example, as in a smart phone), or separate (for example, an external monitorfor a laptop). The other peripheral devices 1120 include, in various examples of embodiments, one or more of a stand-alone digital video disc (or digital versatile disc) (DVR, for both terms), a disk player, a stereo system, and/or a lighting system. Various embodiments use one or more peripheral devices 1120 that provide a function based on the output of the system 1000. For example, a disk player performs the function of playing the output of the system 1000.

In various embodiments, control signals are communicated between the system 1000 and the display 1100, speakers 1110, or other peripheral devices 1120 using signaling such as AV.Link, Consumer Electronics Control (CEC), or other communications protocols that enable device-to-device control with or without user intervention. The output devices can be communicatively coupled to system 1000 via dedicated connections through respective interfaces 1070, 1080, and 1090. Alternatively, the output devices can be connected to system 1000 using the communications channel 1060 via the communications interface 1050. The display 1100 and speakers 1110 can be integrated in a single unit with the other components of system 1000 in an electronic device such as, for example, a television. In various embodiments, the display interface 1070 includes a display driver, such as, for example, a timing controller (T Con) chip.

The display 1100 and speaker 1110 can alternatively be separate from one or more of the other components, for example, if the RF portion of input 1130 is part of a separate set-top box. In various embodiments in which the display 1100 and speakers 1110 are external components, the output signal can be provided via dedicated output connections, including, for example, HDMI ports, USB ports, or COMP outputs.

The embodiments can be carried out by computer software implemented by the processor 1010 or by hardware, or by a combination of hardware and software. As a non-limiting example, the embodiments can be implemented by one or more integrated circuits. The memory 1020 can be of any type appropriate to the technical environment and can be implemented using any appropriate data storage technology, such as optical memory devices, magnetic memory devices, semiconductor-based memory devices, fixed memory, and removable memory, as non-limiting examples. The processor 1010 can be of any type appropriate to the technical environment, and can encompass one or more of microprocessors, general purpose computers, special purpose computers, and processors based on a multi-core architecture, as non-limiting examples.

Various implementations involve decoding. “Decoding”, as used in this application, can encompass all or part of the processes performed, for example, on a received encoded sequence to produce a final output suitable for display. In various embodiments, such processes include one or more of the processes typically performed by a decoder, for example, entropy decoding, inverse quantization, inverse transformation, and differential decoding. In various embodiments, such processes also, or alternatively, include processes performed by a decoder of various implementations described in this application.

As further examples, in one embodiment “decoding” refers only to entropy decoding, in another embodiment “decoding” refers only to differential decoding, and in another embodiment “decoding” refers to a combination of entropy decoding and differential decoding. Whether the phrase “decoding process” is intended to refer specifically to a subset of operations or generally to the broader decoding process will be clear based on the context of the specific descriptions and is believed to be well understood by those skilled in the art.

Various implementations involve encoding. In an analogous way to the above discussion about “decoding”, “encoding” as used in this application can encompass all or part of the processes performed, for example, on an input video sequence to produce an encoded bitstream. In various embodiments, such processes include one or more of the processes typically performed by an encoder, for example, partitioning, differential encoding, transformation, quantization, and entropy encoding. In various embodiments, such processes also, or alternatively, include processes performed by an encoder of various implementations described in this application.

As further examples, in one embodiment “encoding” refers only to entropy encoding, in another embodiment “encoding” refers only to differential encoding, and in another embodiment “encoding” refers to a combination of differential encoding and entropy encoding. Whether the phrase “encoding process” is intended to refer specifically to a subset of operations or generally to the broader encoding process will be clear based on the context of the specific descriptions and is believed to be well understood by those skilled in the art.

Note that the syntax elements used herein are descriptive terms. As such, they do not preclude the use of other syntax element names.

When a figure is presented as a flow diagram, it should be understood that it also provides a block diagram of a corresponding apparatus. Similarly, when a figure is presented as a block diagram, it should be understood that it also provides a flow diagram of a corresponding method/process.

Various embodiments may refer to parametric models or rate distortion optimization. In particular, during the encoding process, the balance or trade-off between the rate and distortion is usually considered, often given the constraints of computational complexity. It can be measured through a Rate Distortion Optimization (RDO) metric, or through Least Mean Square (LMS), Mean of Absolute Errors (MAE), or other such measurements. Rate distortion optimization is usually formulated as minimizing a rate distortion function, which is a weighted sum of the rate and of the distortion. There are different approaches to solve the rate distortion optimization problem. For example, the approaches may be based on an extensive testing of all encoding options, including all considered modes or coding parameters values, with a complete evaluation of their coding cost and related distortion of the reconstructed signal after coding and decoding. Faster approaches may also be used, to save encoding complexity, in particular with computation of an approximated distortion based on the prediction or the prediction residual signal, not the reconstructed one. Mix of these two approaches can also be used, such as by using an approximated distortion for only some of the possible encoding options, and a complete distortion for other encoding options. Other approaches only evaluate a subset of the possible encoding options. More generally, many approaches employ any of a variety of techniques to perform the optimization, but the optimization is not necessarily a complete evaluation of both the coding cost and related distortion.

The implementations and aspects described herein can be implemented in, for example, a method or a process, an apparatus, a software program, a data stream, or a signal. Even if only discussed in the context of a single form of implementation (for example, discussed only as a method), the implementation of features discussed can also be implemented in other forms (for example, an apparatus or program). An apparatus can be implemented in, for example, appropriate hardware, software, and firmware. The methods can be implemented in, for example, , a processor, which refers to processing devices in general, including, for example, a computer, a microprocessor, an integrated circuit, or a programmable logic device. Processors also include communication devices, such as, for example, computers, cell phones, portable/personal digital assistants (“PDAs”), and other devices that facilitate communication of information between end-users.

Reference to “one embodiment” or “an embodiment” or “one implementation” or “an implementation”, as well as other variations thereof, means that a particular feature, structure, characteristic, and so forth described in connection with the embodiment is included in at least one embodiment. Thus, the appearances of the phrase “in one embodiment” or “in an embodiment” or “in one implementation” or “in an implementation”, as well any other variations, appearing in various places throughout this application are not necessarily all referring to the same embodiment.

Additionally, this application may refer to “determining” various pieces of information. Determining the information can include one or more of, for example, estimating the information, calculating the information, predicting the information, or retrieving the information from memory.

Further, this application may refer to “accessing” various pieces of information. Accessing the information can include one or more of, for example, receiving the information, retrieving the information (for example, from memory), storing the information, moving the information, copying the information, calculating the information, determining the information, predicting the information, or estimating the information.

Additionally, this application may refer to “receiving” various pieces of information. Receiving is, as with “accessing”, intended to be a broad term. Receiving the information can include one or more of, for example, accessing the information, or retrieving the information (for example, from memory). Further, “receiving” is typically involved, in one way or another, during operations such as, for example, storing the information, processing the information, transmitting the information, moving the information, copying the information, erasing the information, calculating the information, determining the information, predicting the information, or estimating the information.

It is to be appreciated that the use of any of the following “/”, “and/or”, and “at least one of”, for example, in the cases of “A/B”, “A and/or B” and “at least one of A and B”, is intended to encompass the selection of the first listed option (A) only, or the selection of the second listed option (B) only, or the selection of both options (A and B). As a further example, in the cases of “A, B, and/or C” and “at least one of A, B, and C”, such phrasing is intended to encompass the selection of the first listed option (A) only, or the selection of the second listed option (B) only, or the selection of the third listed option (C) only, or the selection of the first and the second listed options (A and B) only, or the selection of the first and third listed options (A and C) only, or the selection of the second and third listed options (B and C) only, or the selection of all three options (A and B and C). This may be extended, as is clear to one of ordinary skill in this and related arts, for as many items as are listed.

Also, as used herein, the word “signal” refers to, among other things, indicating something to a corresponding decoder. For example, in certain embodiments the encoder signals a particular one of a plurality of transforms, coding modes or flags. In this way, in an embodiment the same transform, parameter, or mode is used at both the encoder side and the decoder side. Thus, for example, an encoder can transmit (explicit signaling) a particular parameter to the decoder so that the decoder can use the same particular parameter. Conversely, if the decoder already has the particular parameter as well as others, then signaling can be used without transmitting (implicit signaling) to simply allow the decoder to know and select the particular parameter. By avoiding transmission of any actual functions, a bit savings is realized in various embodiments. It is to be appreciated that signaling can be accomplished in a variety of ways. For example, one or more syntax elements, flags, and so forth are used to signal information to a corresponding decoder in various embodiments. While the preceding relates to the verb form of the word “signal”, the word “signal” can also be used herein as a noun. As will be evident to one of ordinary skilled in the art, implementations can produce a variety of signals formatted to carry information that can be, for example, stored or transmitted. The information can include, for example, instructions for performing a method, or data produced by one of the described implementations. For example, a signal can be formatted to carry the bitstream of a described embodiment. Such a signal can be formatted, for example, as an electromagnetic wave (for example, using a radio frequency portion of spectrum) or as a baseband signal. The formatting can include, for example, encoding a data stream and modulating a carrierwith the encoded data stream. The information that the signal carries can be, for example, analog or digital information. The signal can be transmitted over a variety of different wired or wireless links, as is known. The signal can be stored on a processor-readable medium.

The preceding sections describe a number of embodiments, across various claim categories and types. Features of these embodiments can be provided alone or in any combination. Further, embodiments can include one or more of the following features, devices, or aspects, alone or in any combination, across various claim categories and types:

At least one embodiment comprises execution of at least a portion of a deep neural network up to a specific split point indicative of a layer of the DNN.

At least one embodiment comprises sending intermediate data, the DNN model, and split point information from a first device to a second device.

At least one embodiment comprises receiving intermediate data, the DNN model, and split point information from a first device for execution of a portion of the DNN model in the second device.

At least one embodiment comprises a decision model for determining at least one split point that is indicative of the layers of a DNN model to be executed on various devices.

At least one embodiment comprises a bitstream or signal that includes one or more of the described syntax elements, or variations thereof.

At least one embodiment comprises a bitstream or signal that includes syntax conveying information generated according to any of the embodiments described.

At least one embodiment comprises creating and/or transmitting and/or receiving and/or decoding according to any of the embodiments described.

At least one embodiment comprises parsing video data or a bitstream to determine operating point of a codec. At least one embodiment comprises a method, process, apparatus, medium storing instructions, medium storing data, or signal according to any of the embodiments described.

At least one embodiment comprises inserting in the signaling syntax elements that enable the decoder to determine decoding information in a manner corresponding to that used by an encoder.

At least one embodiment comprises creating and/or transmitting and/or receiving and/or decoding a bitstream or signal that includes one or more of the described syntax elements, or variations thereof.

At least one embodiment comprises a TV, set-top box, cell phone, tablet, or other electronic device that performs transform method(s) according to any of the embodiments described.

At least one embodiment comprises a TV, set-top box, cell phone, tablet, or other electronic device that performs transform method(s) determination according to any of the embodiments described, and that displays (e.g., using a monitor, screen, or other type of display) a resulting image.

At least one embodiment comprises a TV, set-top box, cell phone, tablet, or other electronic device that selects, bandlimits, or tunes (e.g., using a tuner) a channel to receive a signal including an encoded image, and performs transform method(s) according to any of the embodiments described.

At least one embodiment comprises a TV, set-top box, cell phone, tablet, or other electronic device that receives (e.g., using an antenna) a signal over the air that includes an encoded image, and performs transform method(s).

Claims

1. A method, comprising: executing at least a portion of a deep neural network (DNN) up to a specific layer on a first device; sending intermediate data from the first device to a second device; executing remaining layers of the DNN on the second device; and, sending information from the second device to the first device.

2. An apparatus configured to perform: executing at least a portion of a deep neural network (DNN) up to a specific layer on a first device; sending intermediate data from the first device to a second device; executing remaining layers of the DNN on the second device; and, sending information from the second device to the first device.

3. A decision module for determining at least one split point indicative of a specific layer of a deep neural network model for execution among a plurality of devices.

4. A device configured to receive intermediate data from another device and execute a portion of a deep neural network model, said portion indicated by at least one split point input.

5. The method of Claim 1 or the apparatus of Claim 2, wherein the DNN model is learned before execution.

6. The method of any one of Claim 1 or 5 the apparatus of any one of Claim 2 or 5, wherein the DNN model is learned while executing.

7. The method of any one of Claim 1 , or 5 through 6, or the apparatus of any one of Claim 2 or 5 through 6, wherein learning is deployed in either the first device or the second device.

8. The method of any one of Claim 1 or 5 through 7, or the apparatus of any one of Claim 2 or 5 through 7, wherein supervised learning is based on a ground truth.

9. The method of any one of Claim 1 or 5 through 8, or the apparatus of any one of Claim 2 or 5 through 8, wherein the model is operated via an application that feeds the model with data.

10. The method of any one of Claim 1 or 5 through 9, or the apparatus of any one of Claim 2 or 5 through 9, wherein metrics are collected either synchronously or asynchronously.

11 . The method of any one of Claim 1 or 5 through 10, or the apparatus of any one of Claim 2 or 5 through 10, using reinforcement leaning for a decision module with a reward at one device.

12. A non-transitory computer readable medium containing data content generated according to the method of claim 1 , for playback using a processor.

13. A signal comprising video data generated according to the method of Claim 1 , for playback using a processor.

14. A computer program product comprising instructions which, when a program is executed by a computer, cause the computer to carry out the method of either Claim 1 or any one of Claims 3 through 11 .

15. A non-transitory computer readable medium containing data content comprising instructions to perform the method of any one of claims 1 or 3, and 5 through