Amazon Technologies, Inc.

Xiong Zhou

Gurumurthy Swaminathan

Vineet Khare

Ragav Venkatesan

Runfei Luo

Leo Parker Dirac

Fedor Zhdanov

Li Erran Li

Alex Williams

Weifeng Chen

Jonathan Buck

Min Bai

This disclosure describes a system that models the complexity of image labeling tasks utilizing image data, labeling instructions, and label requirements. A labeling service of a service provider network includes an application that determines a task complexity value based on a data complexity value, a cognitive complexity value, and a product complexity value. The task complexity value is used to predict or estimate needed resources for an image labeling task for labeling image data, e.g., the time and effort needed to label (annotate) the image data. Once the needed resources are estimated, associated costs may also be estimated. The needed resources and associated costs may be provided to a user that submitted the image data, who may then provide an indication with respect to proceeding with the image labeling task.

Automated estimation of needed resources related to labeling a labeling service within a service provider network

Compression profiles may be searched for trained neural networks. An iterative compression profile search may be performed response to a search request. Different prospective compression profiles may be generated for trained neural networks according to a search policy. Performance of compressed versions of the trained neural networks according to the compression profiles may be tracked. The search policy may be updated according to an evaluation of the performance of the compression profiles for the compressed versions of the trained neural networks using compression performance criteria. When a search criteria is satisfied, a result for the compression profile search may be provided.

Searching compression profiles for trained neural networks

Techniques for zero-shot and few-shot transfer of domain-adapted base networks are described. Multiple machine learning task layers are trained using a shared base feature extractor network. At least one task layer is trained with samples and corresponding labels from a first domain as well as a second domain. At least one other task layer is trained with samples and corresponding labels from only the first domain. Ultimately, the other task layer(s) are adapted to generate labels for the first domain via the base network being weighted based on all trainings.

Zero-shot transfer of domain-adapted base networks

Neural networks with similar architectures may be compressed using shared compression profiles. A request to compress a trained neural network may be received and an architecture of the neural network identified. The identified architecture may be compared with the different network architectures mapped to compression profiles to select a compression profile for the neural network. The compression profile may be applied to remove features of the neural network to generate a compressed version of the neural network.

Applying compression profiles across similar neural network architectures

A compression policy to produce compression profiles for compressing trained machine learning models may be trained using reinforcement learning. An iterative reinforcement learning may be performed response to a search request. Different prospective compression profiles may be generated for received machine learning models according to a compression policy being trained. Performance of compressed versions of the trained neural networks according to the compression profiles may be caused using data sets used to train the machine learning models. The compression policy may be updated according to reward signal determined from an application of a reward function for performance criteria to performance results of the different versions of the machine learning models. When a search criteria is satisfied, the trained compression policy may be provided.

Reinforcement learning for training compression policies for machine learning models

Features related to systems and methods expediting generation of a machine learning model, such as an image recognition model, are described. Existing machine learning models are analyzed to identify a starting point for creating the new machine learning model. An existing machine learning model can suggest learning parameters (e.g., training parameters or structural features of the model) that can be used to expedite the generating and training process along with training data that can augment the training of the new machine learning model.

Generating artificial intelligence image processing services

Features related to systems and methods for generating a machine learning model that is a composite of at least two other models (e.g., crowd-sourced models contributed by users) are described. Each of the contributed models provide output values that may not be to scale. To account for these differences, a normalization factor for a first machine learning model is generated to adjust values produced by the first machine learning model to correspond with results from the second machine learning model. The crowd-sourced models along with the normalization factor are included in the new image model generated in the claims.

Growing community of inventors

Seattle, WA, United States of America

Xiong Zhou