KR20210081518A - Method and Apparatus for Mobile Sensing for Few-Shot Adapting to Untrained Conditions - Google Patents

Abstract

Disclosed are a mobile sensing device and a method which adapts to an untrained situation using the small number of data. The embodiment of the present invention proposes a method for generating a conditioned task for efficiently using limited mobile sensing data. Provided are the mobile sensing device and the method, which train a deep learning-based sensing model in a meta-learning method using the conditioned task generated according to the proposed method, and thus it is possible to adapt to untrained conditions using few data (few-shots).

Description

A mobile sensing device and method for adapting to an untrained situation using a small number of data {Method and Apparatus for Mobile Sensing for Few-Shot Adapting to Untrained Conditions}

The present invention relates to a mobile sensing device and method for adapting to an untrained situation using a small number of data.

The content described below merely provides background information related to the present invention and does not constitute the prior art.

In smart devices such as smartphones or wearables, mobile sensing is applied to various fields based on the development of deep learning and mobile AI (Artificial Intelligence) processors. Here, some examples of application fields of mobile sensing include human activity recognition (HAR), acoustic context recognition, complex physical exercise recognition, device-free authentication, and sign language (HAR). sign language recognition, prediction of emotional status, etc. Therefore, when a user enjoys various services provided by a smart device, the application of mobile sensing may be a very important factor.

Although the application of mobile sensing on a smart device can be deployed in a wide variety of applications, in actual situations, mobile sensing faces a problem of performance degradation. The degradation of mobile sensing is mainly due to a wide range of individual conditions. Individual situations arise from various devices and different users using different devices, and may be different from the situation in which the deep learning-based sensing model was initially trained. For example, there are sensors that perform heterogeneous operations according to software and hardware specifications, and various smart devices equipped with such sensors. Also, even in the same smart device, the sensor may generate different results depending on the user's patterns of activities and how the smart device is worn or stored. Countless combinations of these individual situations can severely degrade the performance of mobile sensing.

One of the naive solutions for coping with the performance degradation is to collect sufficient data for a myriad of situations, and then use the collected data to train a sensing model on the device. However, in consideration of the time and cost required for data collection and labeling, and the complexity of training reflecting the avoidance of overfitting, this method is very impractical. Another solution is to calibrate the sensor or estimate condition-independent features to minimize dependence on the sensor. Although this method can be used to resolve local dependence on a specific sensor, there is a limit to being applied to all of various sensors and sensor applications.

Therefore, in mobile sensing, there is a need for a mobile sensing method capable of efficiently coping with each individual situation and minimizing the consumed resources.

Non-Patent Document 1: Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic metalearning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70. JMLR. org, 1126-1135. Non-Patent Document 2: Seyed Ali Rokni, Marjan Nourollahi, and Hassan Ghasemzadeh. 2018. Personalized Human Activity Recognition Using Convolutional Neural Networks. In Thirty-Second AAAI Conference on Artificial Intelligence. Non-Patent Document 3: Jake Snell, Kevin Swersky, and Richard Zemel. 2017. Prototypical networks for few-shot learning. In Advances in Neural Information Processing Systems. 4077-4087.

The present disclosure proposes a method for generating a conditioned task for efficiently using limited mobile sensing data, and deep learning-based sensing using the situation task generated according to the proposed method. The main objective is to provide a mobile sensing device and method capable of adapting to untrained conditions using a small number of data (few-shots) by training a model in a meta-learning method. .

According to an embodiment of the present invention, there is provided a learning method for a mobile sensing device, the method comprising: generating a plurality of conditioned tasks for learning using a source dataset; a process of executing basic training for a deep learning-based sensing model using the plurality of situational tasks; obtaining a small number of data (few shots) and generating a target dataset by using the small number of data; and using the target dataset to provide a learning method implemented on a computer, characterized in that it comprises the step of executing adaptive training for the sensing model on which the basic training has been performed.

According to another embodiment of the present invention, there is provided a mobile sensing device mounted on a mobile device, comprising: an input unit configured to obtain sensing data from a mobile sensor; a deep learning-based sensing model trained in advance using basic training using a source dataset and adaptive training using a small number of data (few-shots); a microprocessor for generating a sensing result by inputting the sensing data into the sensing model; and a program implementing the sensing model and a memory for storing the sensing data.

According to another embodiment of the present invention, there is provided a computer program stored in a computer-readable recording medium in order to execute each step included in the learning method of a mobile sensing device.

As described above, according to the present embodiment, a method for generating a conditioned task for efficiently using limited mobile sensing data is proposed, and a deep dive using the situation task generated according to the proposed method is proposed. By providing a mobile sensing device and method for training a learning-based sensing model in a meta-learning method, adapting to untrained conditions using a small number of data (few-shots) It has the effect of making it possible.

In addition, according to the present embodiment, by providing a mobile sensing apparatus and method for training a deep learning-based sensing model in a meta-learning manner, individual conditions that can be derived from various devices and users ) has the effect of making it possible to effectively deal with

1 is a block diagram of a mobile sensing device according to an embodiment of the present invention.
2 is a conceptual diagram for learning of a sensing model according to an embodiment of the present invention.
3 is a flowchart of a learning method of a sensing model according to an embodiment of the present invention.
4 is a conceptual diagram of a situational task creation process according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to exemplary drawings. In adding reference numerals to the components of each drawing, it should be noted that the same components are given the same reference numerals as much as possible even though they are indicated on different drawings. In addition, in describing the present embodiments, if it is determined that a detailed description of a related well-known configuration or function may obscure the gist of the present embodiments, the detailed description thereof will be omitted.

Also, in describing the components of the present embodiments, terms such as first, second, A, B, (a), (b), etc. may be used. These terms are only for distinguishing the elements from other elements, and the essence, order, or order of the elements are not limited by the terms. Throughout the specification, when a part 'includes' or 'includes' a certain component, it means that other components may be further included, rather than excluding other components, unless otherwise stated. . In addition, the '... Terms such as 'unit' and 'module' mean a unit that processes at least one function or operation, which may be implemented as hardware or software or a combination of hardware and software.

DETAILED DESCRIPTION The detailed description set forth below in conjunction with the appended drawings is intended to describe exemplary embodiments of the present invention and is not intended to represent the only embodiments in which the present invention may be practiced.

This embodiment discloses a mobile sensing device and method for adapting to an untrained situation using a small amount of data. In more detail, a method for generating a conditioned task for efficiently using limited mobile sensing data is proposed, and deep learning-based sensing using the situation task generated according to the proposed method. Provided are a mobile sensing device and method for training a model in a meta-learning method.

Meta-learning (also known as 'learning to learn') is one of the methods of training a deep learning model so that it can quickly adapt to a situation not encountered during the training process (see Non-Patent Document 1). Based on the meta-learning rapid adaptation method to an untrained situation, the present embodiment provides a mobile sensing device and method capable of coping with various individual conditions encountered by a mobile sensing model.

It is assumed that the mobile sensing apparatus according to the present embodiment is implemented on a mobile device.

A shot defining the size of the training data means the number of given training data per one label. Hereinafter, the expression “few-shot” or a small number of data indicates a case in which the number of given training data per label is very small (eg, 1 or 2).

1 is a block diagram of a mobile sensing device according to an embodiment of the present invention.

In an embodiment of the present invention, the mobile sensing device 100 acquires sensing data from mobile sensors, and inputs the acquired sensing data to a deep learning-based sensing model to sense produce results. The mobile sensing device 100 includes all or part of H (H is a natural number) input units and H sensing models. Here, the components included in the mobile sensing device 100 according to the present embodiment are not necessarily limited thereto. For example, a training unit (not shown) for training each sensing model may be additionally provided on the mobile sensing device 100 , or may be implemented in the form of interworking with an external training unit.

In this embodiment, it is assumed that one sensing model exists for one application area. Here, the application area may be one of several fields including human behavior recognition, voice recognition, complex motion recognition, device-independent authentication, sign language recognition, emotional state prediction, and the like. Accordingly, when the mobile device on which the mobile sensing apparatus 100 is mounted supports a plurality of application areas, as shown in FIG. 1 , the mobile sensing apparatus 100 may include a plurality of sensing models. Since the concept of each input unit providing sensing data and the operation concept of each sensing model are similar, hereinafter, one input unit 101 and one sensing model 102 will be described as an example.

The input unit 101 obtains sensing data from mobile sensors. The mobile sensor generates sensing data corresponding to an application area from a stimulus external to the mobile device. For example, the behavior recognition sensor may generate sensing data in the form of acceleration in the x, y, and z axis directions from a human action, or the voice recognition sensor may recognize a voice to generate the sensed data on the time axis.

The sensing model 102 generates a sensing result using sensing data. For example, with respect to the sensing data obtained from the behavior recognition sensor, the sensing result may be differentiated motions, and for the sensing data obtained by the voice recognition sensor, the sensing result may be recognized keywords.

The sensing model 102 according to this embodiment is implemented as a deep learning-based neural network, and it is based on meta-learning in advance so that it is possible to cope with individual situations caused by the combination of various mobile devices and various users. are trained on

As shown in FIG. 2 , the training of the sensing model 102 consists of two stages: basic training and adapting training.

Basic training for the sensing model 102 uses meta-learning concepts. The training unit generates a plurality of conditioned tasks from N (N is a natural number) sub-dataset included in a source dataset that is sensing data for learning. As described above, the sensing data for learning for one sensing model 102 is data for one application area. The source dataset includes multiple subsets in order to mimic various individual situations.

The training unit updates parameters of the sensing model 102 using meta-learning based on the generated task. The training unit updates the parameters of the sensing model 102 based on a distance metric between the label and the value inferred by the sensing model 102 using a plurality of context tasks. Here, as the distance metric, any one capable of expressing a metric difference between two comparison objects, such as cross entropy, L1 or L2 metric, may be used.

In order to finally adapt to a user using a mobile device, a target dataset is generated by acquiring a small number of data (few-shots) specialized for the user, and then the training unit is basically trained using the small number of data. The sensing model 102 is adaptively trained. The training unit updates the parameters of the sensing model 102 based on a distance metric between the label and the value inferred by the sensing model 102 using a small number of data.

The mobile sensing device 100 may generate a target dataset by attaching a label after acquiring a small number of data for use in adaptive training from the mobile sensor using a short time interval.

Details of basic training and adaptive training for the sensing model 102 will be described later.

1 is an exemplary configuration according to the present embodiment, and implementation including other components or other connections between components is possible depending on the shape of the mobile sensor and the sensing model 102 .

The mobile device on which the mobile sensing apparatus 100 according to the present embodiment is mounted may be any device capable of acquiring sensing data using various mobile sensors. The mobile device may be a programmable computer, and includes a program implementing the sensing model 102 and a memory storing sensing data, a microprocessor executing the program, and at least one communication interface capable of connecting to a server. .

The adaptive training for the sensing model 102 as described above may be performed in the mobile device using the computing power of the mobile device on which the mobile sensing apparatus 100 is mounted.

Basic training based on meta-learning for the sensing model 102 as described above may be performed in the server. For a deep learning model having the same structure as the sensing model 102 , which is a component of the mobile sensing device 100 mounted on the mobile device, the training unit of the server may perform basic training based on meta-learning. Using a communication interface connected to the mobile device, the server transmits the parameters of the trained deep learning model to the mobile device, and the mobile sensing device 100 can update the parameters of the sensing model 102 using the received parameters. have. In addition, the parameter of the sensing model 102 may be set at the time of shipment of the mobile device or the time when the mobile sensing apparatus 100 is mounted on the mobile device.

Hereinafter, a training method for a sensing model based on meta-learning will be described with reference to FIGS. 3 and 4 .

3 is a flowchart of a learning method of a sensing model according to an embodiment of the present invention. Figure 3 (a) is an overall process for the learning method of the sensing model, Figure 3 (b) is a detailed process for the situation task generation (step S301), Figure 3 (c) is basic training (step S302) ) is a detailed process for 4 is a conceptual diagram of a situational task creation process according to an embodiment of the present invention.

The training unit according to the present embodiment generates a situational task for learning by using the source dataset (S301).

를 포함하는데, 여기서

는 사전에 수집된 센싱 데이터이고,

는

에 대응되는 레이블이다. 소스 테이터세트 D는 M(M은 자연수) 개의 레이블을 포함하는 것으로 가정한다. 소스 테이터세트 D에 포함된 데이터 쌍의 수는 트레이닝을 실행하기에 충분할 정도로 많다고 가정한다. 센싱 모델(102)의 파라미터를

라 할 때, 센싱 데이터

에 대하여 센싱 결과는

로 표현한다.The source dataset D is the data pair

includes, where

is the pre-collected sensing data,

is

is the label corresponding to . It is assumed that the source data set D includes M (M is a natural number) labels. It is assumed that the number of data pairs included in the source dataset D is large enough to run training. parameters of the sensing model 102.

, the sensing data

The sensing result for

expressed as

를 ICD(individual condition dataset)라 하고, 수학식 1에 표현된 바와 같은 관계를 만족하는 것으로 가정한다. It is assumed that the source data set D includes sensing data for C (C is a natural number) individual situations. A subset containing sensing data for each individual situation

is called an individual condition dataset (ICD), and it is assumed that the relationship expressed in Equation 1 is satisfied.

를 만족한다. 소스 테이터세트 D가 M 개의 레이블을 보유하는 경우, 각 ICD도 M 개의 레이블에 대한 센싱 데이터를 포함하는 것으로 가정한다.where i and j are different natural numbers,

is satisfied with If the source data set D has M labels, it is assumed that each ICD also includes sensing data for M labels.

의 집합 T

를 생성한다.The training unit uses the source dataset D to perform a learning situational task.

set of T

create

별로 단일 상황 과제

를 생성하되, 각 상황 과제는 M 개의 레이블을 보유하도록 한다. 생성된 단일 상황 과제는 C 개이다. 단일 상황 과제는, 모바일 디바이스가 사용자에게 전달되었을 때, 센싱 모델(102)이 당면하게 되는 미출현 상황을 반영하기 위하여 생성된다.First, the training unit generates a single per-condition task using the ICD included in the source dataset (S311). As shown in Fig. 4, the training unit is each ICD

Not a single situational task

, but each situational task has M labels. There are C single situational tasks generated. The single context task is created to reflect the out-of-the-box context the sensing model 102 will encounter when the mobile device is delivered to the user.

In addition, the training unit generates multi-conditioned tasks by using the ICD (S312). There are C single situational tasks generated. As shown in FIG. 4 , the training unit extracts sensing data corresponding to each of M labels from an arbitrary ICD for one multi-context task. Multi-situational tasks are created to reflect more diverse situations compared to single-situational tasks and to avoid overfitting that single-situational tasks can result in.

The training unit according to the present embodiment maintains consistency with respect to the label for each of the single situational task and the multi-contextual task. That is, each task has M labels to maintain homogeneity between tasks, which is differentiated from the conventional meta-learning method (see Non-Patent Documents 1 and 3). The prior art has aimed at adapting to any task that may not possess the same label. On the other hand, in this embodiment, even if the data distribution of the task is different, since the label is meta-learning for the source and target datasets that may be the same, a consistent label can be maintained when creating a situational task.

In the process of collecting the source dataset, an imbalance may exist between the data distributions of the ICD. By generating a situational task by sampling data from ICD uniformly, there is an effect that it becomes possible to generate unbiased data in which imbalance between distributions of ICD data is reduced.

The total number of single-context tasks and multi-context tasks generated by the training unit is 2C. 4 shows a process in which 5 single-context tasks and 5 multi-context tasks are generated for a source dataset including 5 ICDs.

As the number of situational tasks increases, the sensing model 102 can reflect more diverse situations in the basic training process, thereby helping to improve the performance of the mobile sensing device 100 . However, when increasing the number of situational tasks, the time and cost of collecting the source dataset may increase, so it is necessary to maintain a balance.

The training unit executes basic training on the sensing model using the situational task (S302). Based on meta-learning using the situational task, the training unit updates the parameters of the sensing model 102 .

In this embodiment, a meta-learning method based on MAML (Model-Agnostic Meta Learning, see Non-Patent Document 1) is used. A useful assumption according to MAML is that there are initial parameters transferable to the target task with only a small amount of data. Therefore, even in this embodiment, the parameters of the sensing model essential for adaptive training are generated using MAML-based basic training.

에 대하여, 트레이닝부는 지원 집합(support sets) 및 문의 집합(query sets)을 추출한다(S321). 지원 집합

는 센싱 모델(102)을 트레이닝하여, 과제 특화된(task specific) 파라미터를 생성하기 위하여 사용되고, 문의 집합

는 과제 특화된 파라미터가 설정된 센싱 모델(102)을 트레이닝하기 위하여 사용된다. 지원 집합과 문의 집합은 서로 겹치지지 않도록 추출되는 것으로 가정한다.each situational task

For , the training unit extracts support sets and query sets (S321). support set

is used to train the sensing model 102 to generate task specific parameters, and a set of queries

is used to train the sensing model 102 in which task-specific parameters are set. It is assumed that the support set and the statement set are extracted so as not to overlap each other.

Each of the support set and the query set includes K shots, where K may be equal to the number of shots used for subsequent adaptive training. In this embodiment, by keeping K to a very small value, it is possible to reduce the time and cost consumed for collecting the source dataset.

에 대하여, 트레이닝부는 과제 특화된 손실 함수(loss function)를 생성한다(S322). 과제 특화된 손실 함수는 수학식 2에 나타낸 바와 같다.Since mobile sensing is a classification for multiple labels, each support set

For , the training unit generates a task-specific loss function (S322). The task-specific loss function is as shown in Equation (2).

Here, the loss function uses cross-entropy as a distance metric, but any L1 or L2 metric, such as an L1 or L2 metric, can be used as long as it can express a metric difference between two comparison objects.

The training unit generates task-specific parameters by updating the parameters of the sensing model based on the task-specific loss function (S323). The update of the parameters of the sensing model 102 is as shown in Equation (3).

는 하이퍼파라미터(hyper-parameter)로서, 큰 값으로 설정되어(예컨대, 0.1), 수 차례의 파라미터 업데이트만으로

가 생성되도록 한다.here,

is a hyper-parameter, which is set to a large value (eg, 0.1), and requires only several parameter updates.

to be created.

을 이용하여 메타 손실 함수(meta loss function)를 생성한다(S324). 메타 손실 함수는 수학식 4에 나타낸 바와 같다. The training department sets task-specific parameters and queries

A meta loss function is generated by using (S324). The meta loss function is as shown in Equation (4) .

는 수학식 2에 나타낸 바와 동일한 손실 함수이며, 인자로서

를 사용한다. here,

is the same loss function as shown in Equation 2, and as a factor

use

The training unit updates the parameters of the sensing model in a direction to minimize the meta loss function (S325). The update of the parameters of the sensing model 102 is as shown in Equation (5).

는 하이퍼파라미터이다. 수학식 4 및 수학식 5를 이용하여, 트레이닝부는 각 문의 과제에 대한 평균적인 손실을 최소화할 수 있는 파라미터를 생성한다. 수학식 2부터 수학식 5까지 나타낸 바와 같이, 두 단계를 이용하여 생성된 파라미터를 초기값으로 이용하는 센싱 모델(102)은 추후 빠르게 새로운 상황에 적응할 수 있게 된다.here,

is a hyperparameter. Using Equations 4 and 5, the training unit generates a parameter capable of minimizing an average loss for each inquiry task. As shown in Equations 2 to 5, the sensing model 102 using the parameters generated using the two steps as initial values can quickly adapt to a new situation later.

를 만족한다.The training unit generates a target dataset by acquiring a small number of data (S303). During a short time period using the mobile sensor, the training unit acquires a small amount of data for use in adaptive training from the user, and then attaches a label to generate the target dataset U. The target dataset contains M labels and may contain a small number of shots dependent on the time to acquire the data. Also for the source dataset D,

is satisfied with

The training unit performs adaptive training on the sensing model using the target dataset (S304).

로 표기하고, 센싱 모델(102)의 초기값으로 이용하면, 센싱 모델(102)은

로 표현될 수 있다. 이러한 과정은 기본 트레이닝되는 모델과 적응 트레이닝되는 모델이 서로 다른 모델일 경우, 기본 트레이닝된 모델의 파라미터

가 적응 트레이닝을 위한 모델로 전이(transfer)되는 것으로도 볼 수 있다. parameters after basic training.

When expressed as and used as an initial value of the sensing model 102, the sensing model 102 is

can be expressed as If the basic trained model and the adaptively trained model are different models, the parameters of the basic trained model are

It can also be seen as being transferred to a model for adaptive training.

The training unit updates the parameters of the sensing model 102 as shown in Equation 6 based on the distance metric between the label and the value inferred by the sensing model 102 using the target dataset U. Any distance metric, such as cross entropy, L1, or L2 metrics, can be used as long as it can express a metric difference between two comparison objects.

는 수학식 3에서와 동일한 값이 사용될 수 있다.

가 큰 값으로 설정되므로, 수학식 6에 따른 적응 트레이닝이 신속하게 진행될 수 있다. here

The same value as in Equation 3 may be used.

Since is set to a large value, adaptive training according to Equation (6) can be quickly performed.

Hereinafter, an experimental example of the performance of the mobile sensing device 100 according to the present embodiment will be described.

In order to reflect various individual situations, a dataset for the experiment was collected using smart devices used by 10 users. The 10 smart devices include 7 smart phones and 3 wearable devices, each having a different operating system (OS) version.

In the experiment, IMU (Inertial Measurement Unit) and voice recognition sensor were used as mobile sensors. The IMU recognizes the user's actions, and the voice recognition sensor recognizes the user's voice. The labels for the user's actions used in the experiment were 9 actions, and the labels for the voice were 14 words.

The following six baselines were used for comparison.

In the baseline Src (source only), only basic training using the source dataset was run, and no adaptive training was run.

At the baseline Tgt (target only), training was performed using a small number of data (few-shots) acquired with the target user.

At the baseline Src+Tgt (source plus target), training was performed using both the source dataset and a small amount of data acquired with the target user.

The baseline Trc (Transfer Convolutional, see Non-Patent Document 2) is a state-of-the-art (SOTA) transfer learning method that adapts to the behavioral recognition of a target user by a mobile sensor using a small number of data samples. . In Trc, convolutional layers included in a trained Convolutional Neural Network (CNN) are fixed, and only fully connected layers of the final stage are adaptively trained using a small number of data.

The baseline PN (Prototypical Network, see Non-Patent Document 3) is a learning method using a small number of data based on meta-learning. In PN, prototypes corresponding to a plurality of labels are generated during a training process. In the reasoning process of PN, the closest prototype is classified using the distance metric.

Baseline MAML (see Non-Patent Document 1) is a method used as a basis for basic training in this embodiment, and is a learning method using a small number of data based on meta-learning. Compared with MAML, an important feature of the mobile sensing device 100 according to the present embodiment is the generation and use of a situational task for basic training.

For fair comparison, both the baseline and the present embodiment were implemented using CNN. The implemented CNN includes five CNN layers and three full-connection layers. The size K of the shot used for basic training was set to 5.

In the experiment, after one target user was selected, data for another user was used as the source dataset. Since the dataset for 10 users was collected, after a total of 10 experiments were run, the average accuracy was calculated using the results of each experiment.

Tables 1 and 2 show the average accuracy for the six baselines and this example. Table 1 shows the average accuracy for 9 types of behavior recognition, and Table 2 shows the average accuracy for 14 types of speech recognition. As shown in Tables 1 and 2, the average accuracy was calculated for the small number of shot sizes K 1, 2, 5, and 10 used for adaptive training.

As shown in Tables 1 and 2, the mobile sensing device 100 according to the present embodiment showed superior accuracy compared to all baselines regardless of the number of shots. In particular, in the case of behavior recognition using only one shot, the present embodiment showed an improvement in accuracy of 39% compared to the baseline Src and 15% compared to the baseline Trc, and even compared with the baseline PN and MAML using a small number of data. showed high accuracy.

In addition, in the case of speech recognition using only one shot, the present embodiment showed an accuracy improvement of 44% compared to the baseline Tgt and 38% compared to the baseline Trc, and 33% or more compared to the baseline PN and MAML using a small number of data improved accuracy. For both behavior recognition and speech recognition, it was shown that as the number of shots increased, the performance gap between this example and six baselines decreased. That is, the present embodiment showed excellent performance compared to the baseline only with adaptive training using a small number of data. Accordingly, it was confirmed that the present embodiment can be efficiently used even when the time interval for acquiring a small number of data for use in adaptive training from the user is very short.

As described above, according to the present embodiment, a method for generating a conditioned task for efficiently using limited mobile sensing data is proposed, and a deep dive using the situation task generated according to the proposed method is proposed. By providing a mobile sensing device and method for training a learning-based sensing model in a meta-learning method, adapting to untrained conditions using a small number of data (few-shots) It has the effect of making it possible.

In addition, according to this embodiment, by providing a mobile sensing apparatus and method for training a deep learning-based sensing model in a meta-learning method, individual conditions that can be derived from various devices and users ) has the effect of making it possible to effectively deal with

Although it is described that each process is sequentially executed in each flowchart according to the present embodiment, the present invention is not limited thereto. In other words, since it may be applicable to change and execute the processes described in the flowchart or to execute one or more processes in parallel, the flowchart is not limited to a time-series order.

Various implementations of the systems and techniques described herein include digital electronic circuitry, integrated circuits, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), computer hardware, firmware, software, and/or combination can be realized. These various implementations may include being implemented in one or more computer programs executable on a programmable system. The programmable system includes at least one programmable processor (which may be a special purpose processor) coupled to receive data and instructions from, and transmit data and instructions to, a storage system, at least one input device, and at least one output device. or may be a general-purpose processor). Computer programs (also known as programs, software, software applications or code) contain instructions for a programmable processor and are stored on a "computer-readable recording medium".

The computer-readable recording medium includes all types of recording devices in which data readable by a computer system is stored. These computer-readable recording media are non-volatile or non-transitory, such as ROM, CD-ROM, magnetic tape, floppy disk, memory card, hard disk, magneto-optical disk, and storage device. media, and may further include transitory media such as carrier waves (eg, transmission over the Internet) and data transmission media. In addition, the computer-readable recording medium is distributed in network-connected computer systems, and computer-readable codes may be stored and executed in a distributed manner.

Various implementations of the systems and techniques described herein may be implemented by a programmable computer. Here, the computer includes a programmable processor, a data storage system (including volatile memory, non-volatile memory, or other types of storage systems or combinations thereof), and at least one communication interface. For example, a programmable computer may be one of a server, a network appliance, a set-top box, an embedded device, a computer expansion module, a personal computer, a laptop, a Personal Data Assistant (PDA), a cloud computing system, or a mobile device.

The above description is merely illustrative of the technical idea of this embodiment, and various modifications and variations will be possible by those skilled in the art to which this embodiment belongs without departing from the essential characteristics of the present embodiment. Accordingly, the present embodiments are intended to explain rather than limit the technical spirit of the present embodiment, and the scope of the technical spirit of the present embodiment is not limited by these embodiments. The protection scope of this embodiment should be interpreted by the following claims, and all technical ideas within the equivalent range should be interpreted as being included in the scope of the present embodiment.

100: mobile sensing device 101: input unit
102: sensing model

Claims (15)

In the learning method of a mobile (mobile) sensing device,
generating a plurality of conditioned tasks for learning using a source dataset;
a process of executing basic training for a deep learning-based sensing model using the plurality of situational tasks;
obtaining a small number of data (few shots) and generating a target dataset by using the small number of data; and
The process of executing adaptive training for the sensing model on which the basic training has been performed using the target dataset
A learning method implemented on a computer, characterized in that it comprises a. According to claim 1,
The source dataset is
A computer-implemented learning method, characterized in that it is collected in advance and comprises a plurality of individual condition datasets (ICDs). 3. The method of claim 2,
The process of creating the situation task is,
and generating at least one per-condition task and at least one multi-conditioned task from the plurality of ICDs. 3. The method of claim 2,
The process of creating the situation task is,
for each of the ICDs, randomly selecting data from the one ICD to generate the single-context task, and randomly selecting data from all of the plurality of ICDs to generate the multi-context task. The learning method implemented. 3. The method of claim 2,
The process of creating the situation task is,
By generating each of the single context task and the multiple context task to include the same number of labels as the number of labels included in the source dataset, homogeneity of the labels of the context task is maintained. , a learning method implemented on a computer. 6. The method of claim 5,
The process of executing the basic training is,
Implementation on a computer, characterized in that for each of the situational tasks, a parameter of the sensing model is updated based on a distance metric between the value inferred by the sensing model and the label to generate a task specific parameter learning method to be. 7. The method of claim 6,
The process of executing the basic training is,
Direction of minimizing a loss function value based on a distance metric between the value inferred by the sensing model and the label with respect to the sensing model in which the task specific parameter is set, using the entire situation task A learning method implemented on a computer, characterized in that for updating the parameters of the sensing model. According to claim 1,
The process of generating the target data set includes:
Using a short time interval, after acquiring the small number of data from mobile sensors, and attaching a label to the small number of data, implemented on a computer learning method. 9. The method of claim 8,
The process of executing the adaptive training is,
Setting the parameters according to the basic training as the initial parameters of the sensing model, and using the target dataset, updating the parameters of the sensing model based on the distance metric between the value inferred by the sensing model and the label Characterized, a learning method implemented on a computer. In the mobile sensing device mounted on a mobile device (mobile device),
an input unit for acquiring sensing data from a mobile sensor;
a deep learning-based sensing model trained in advance using basic training using a source dataset and adaptive training using a small number of data (few-shots);
a microprocessor for generating a sensing result by inputting the sensing data into the sensing model; and
A program implementing the sensing model and a memory for storing the sensing data
Mobile sensing device comprising a. 11. The method of claim 10,
Mobile sensing device, characterized in that it further comprises at least one communication interface connected to the server. 12. The method of claim 11,
The basic training is executed in the server using a deep learning model having the same structure as the sensing model, and the adaptive training is executed in the mobile device. 13. The method of claim 12
Mobile sensing device, characterized in that receiving the parameters of the deep learning model from the server using the communication interface. 11. The method of claim 10,
The microprocessor is
Mobile sensing device, characterized in that the sensing result is generated by executing the program based on the sensing data. A computer program stored in a computer-readable recording medium to execute each step included in the learning method of a mobile sensing device according to any one of claims 1 to 9.

Images