One article combs multi task learning (mmoe/ple/dupn/essm, etc.)

When you're making models , Often focus on the optimization of a specific index , Such as click through rate model , Just optimize AUC, Do a binary model , Just optimize f-score. However , This ignores the information gain and effect improvement that the model can bring by learning other tasks . By sharing vector representations in different tasks , We can greatly improve the generalization effect of the model in various tasks . This method is what we're going to talk about today - Multi task learning (MTL).

So how to determine whether it's multi task learning ？ You don't need to look at the whole structure of the model , Just look at loss Function , If loss Contains many items , Each is a different goal , This model is multi task learning . Sometimes , Although your model is only an optimization goal , It can also be done through multitasking , Enhance the generalization effect of the model . For example, click through rate model , We can do this by adding transformation samples , Building AIDS loss( Estimated conversion ), So as to improve the generalization of the click through rate model .

Why multitasking is effective ？ for instance , A model has learned to distinguish colors , If this model is directly applied to the classification task of vegetables and meat ？ It's easy to learn that the green ones are vegetables , The other, more likely, is meat . Is regularization multitasking ？ Regularized, optimized loss Not only does it have its own return / Classification produces loss, also l1/l2 Produced loss, Because we think " It's right and just fitting " The parameters of the model should be sparse , And it's not easy to be too big , To put this assumption into the model to learn , So we have the regularization term , Nature is also an extra task .

MTL Two methods

The first is hard parameter sharing, As shown in the figure below :

Relatively simple , The first few floors dnn Share... For each task , Then separate out the different tasks layers. This method effectively reduces the risk of over fitting : The more tasks the model learns at the same time , In the sharing layer, the model needs to learn a common embedded expression to make every task perform better , So as to reduce the risk of over fitting .

The second is soft parameter sharing, As shown in the figure below :

In this way , Each task has its own model , It has its own parameters , But there are restrictions on the parameters between different models , The parameters of different models must be similar , There will be a distance Describe the similarity between parameters , Will be added to the model learning as an additional task , Similar to the regularization term .

Multitasking can improve , Mainly due to the following reasons :

• Implicit data enhancement ： Each task has its own sample , With multitasking learning , The sample size of the model will increase a lot . And the data is noisy , If you study alone A Mission , The model will put A The noise of data is also learned , If it's multitasking , Because the model requires B The task also needs to learn well , It's going to ignore A The noise of the mission , Empathy , Modeling A I'll ignore it when I'm in the middle of it B The noise of the mission , So multi task learning can learn a more accurate embedded expression .
• Focus on ： If the task has a lot of data noise , Very little data and very high dimensions , The model can't distinguish the related features from the non related features . Multi task learning can help models focus on useful features , Because different tasks will reflect the correlation between characteristics and tasks .
• Feature information theft ： There are some features in the mission B It's easy to learn in English , On mission A It's hard to learn in English , The main reason is the mission A The interaction with these features is more complex , And for the task A For example, other features may hinder the learning of some features , So through multitasking learning , The model can learn every important feature efficiently .
• Expression bias ：MTL Make the model learn the vector representation that all tasks prefer . This will also help to extend the model to new tasks in the future , Because it's assumed that space performs well for enough training tasks , Also good at learning new tasks .
• Regularization ： For a task , The learning of other tasks will regularize the task .

Multi task deep learning model

Deep Relationship Networks：

From the chart , We can see that the first few convolution layers are pre trained , The latter layers share parameters , Used to learn the connections between different tasks , Finally, independent dnn Modules are used to learn about tasks .

Fully-Adaptive Feature Sharing：

From the other extreme , Here's a bottom-up approach , Start with a simple network , In the training process, the network is greedily and dynamically expanded by using the grouping criteria of similar tasks . Greedy methods may not be able to find a globally optimal model , And each branch is only assigned to one task, which makes the model unable to learn the complex interaction between tasks .

cross-stitch Networks:

As mentioned above soft parameter sharing, The model is two completely separate model structures , The structure uses cross-stitch Unit to let the separated model learn the relationship between different tasks , As shown in the figure below , By means of pooling After the layer and full connection layer are added respectively cross-stitch Linear fusion of the feature expression learned earlier , And then output to the following convolution / Fully connected module .

A Joint Many-Task Model：

As shown in the figure below , The predefined hierarchy is made up of NLP Task composition , Low level structures are learned through word level tasks , So do the analysis , Block labeling, etc . The structure of the intermediate level is learned through the task of parsing level , Such as syntactic dependency . High level structure is learned through semantic level tasks .

weighting losses with uncertainty：

Considering the uncertainty of correlation between different tasks , Multi task loss function based on Gaussian likelihood maximization , Adjust the relative weight of each task in the cost function . The structure is shown in the following figure , Regression of pixel depth 、 Semantic and instance segmentation .

sluice networks:

The following model summarizes the deep learning based MTL Method , Such as hard parameter sharing and cross-stitch The Internet 、 Block sparse regularization method , And recently created the task hierarchy NLP Method . The model can learn which layers and subspaces should be shared , And at which layers the network learns the best representation of the input sequence .

ESSM:

In the e-commerce scene , Transformation refers to the process from Click to purchase . stay CVR Estimate when , We often have two problems ： Sample bias and data coefficients . Sample bias refers to the difference between training and test samples , Take e-commerce for example , The model trains with click data , And the estimate is the entire sample space . The problem of data sparsity is even more serious , There are very few samples per click , Even less transformation , So we can learn from the idea of multi task learning , Introducing assisted learning tasks , fitting pCTR and pCTCVR（pCTCVR = pCTR * pCVR）, As shown in the figure below :

• about pCTR Come on , Exposure events with click behavior can be taken as positive samples , Exposure events without click behavior are taken as negative samples
• about pCTCVR Come on , We can take the exposure events with both click behavior and purchase behavior as positive samples , Others as negative samples
• about pCVR Come on , Only the gradients in the sample with no exposure click can be returned to main task In the network

The other two subnetworks are embedding Layers are shared , because CTR The training sample size of the task is much larger than CVR The training sample size of the task , So it can alleviate the sparsity problem of training data .

DUPN：

The model is divided into behavior sequence layer 、Embedding layer 、LSTM layer 、Attention layer 、 Downstream multitasking layer (CTR、LTR、 Fashionistas focus on Forecasting 、 The strength and quantity of users' purchase ). As shown in the figure below

MMOE:

As shown in the figure below , Model (a) Most common , Shared the underlying network , It is connected with the full connection layer of different tasks . Model (b) Different experts can extract different features from the same input , By a Gate( similar ) attention structure , The features extracted by experts are filtered out to each task The most relevant feature , Finally, the full connection layer of different tasks is connected .MMOE The idea is for different tasks , Need information extracted by different experts , So each task needs an independent gate.

PLE：

Even if passed MMoE In this way, the negative transfer phenomenon is alleviated , The seesaw phenomenon is still widespread ( Seesaw phenomenon refers to when the correlation between multiple tasks is not strong , Information sharing will affect the effect of the model , There will be a task that becomes more generalized , Another weakening phenomenon ）.PLE The essence is MMOE Improved version , There are some expert It's mission specific , There are some expert Is Shared , Here's the picture CGC framework , For the task A for , adopt A Of gate hold A Of expert And shared expert To merge , To learn A.

Final PLE The structure is as follows , Integrated with customized expert and MMOE, Stack multiple layers CGC framework , As shown below :

The ten party @ Alchemy notes