Google to Improve Mobile Machine Learning Models

The team used multi-criteria optimization to achieve high speed and accuracy in the neural network MnasNet
09 August 2018   363

The neural network MnasNet uses a training method with reinforcement to select the appropriate mobile device architecture.

The MnasNet system contains:

  • object management based on a recurrent neural network for learning and selecting the architecture of the model;
  • he coach who builds the models and teaches their accuracy;
  • an output machine that calculates the speed of the model on the phones using TensorFlow.

The team used multi-criteria optimization to achieve high speed and accuracy. In addition, scientists used the learning algorithm with reinforcement and remuneration function. Thus, MnasNet finds Pareto optimality for each platform. For each mobile device, individual features are created that require a specific architecture.

Automated Neural Design Approach
Automated Neural Design Approach

To achieve the optimal balance between the flexibility of search and the scope of feasible solutions, a hierarchical approach has been applied. It represents a convolutional neural network in the form of a series of blocks. After, sequential search to assign layers of architecture to each block is being used. Due to this, each layer uses different operations and connections. In this case, the layers of blocks are copied by force.

MnasNet blocks
MnasNet blocks

Scientists have included the search speed of architecture in the function of rewarding the search algorithm.

MnasNet Blocks
MnasNet Blocks

According to the developers, the system picks up the model 1.5 times faster than MobileNetV2 and 2.4 times faster than NASNet, while taking into account the optimal accuracy and speed. As a proof, the result was demonstrated by the example of working with ImageNet image database.

AI to be Used to Create 3D Motion Sculptures

The system developed by the MIT and Berkeley scientists is called MoSculp and is based on artificial inteligence
21 September 2018   110

MoSculp, the joint work of MIT scientists and the University of California at Berkeley, is built on the basis of a neural network. The development analyzes the video recording of a moving person and generates what the creators called "interactive visualization of form and time." According to the lead specialist of the project Xiuming Zhang, software will be useful for athletes for detailed analysis of movements.

At the first stage, the system scans the video frame-by-frame and determines the position of key points of the object's body, such as elbows, knees, ankles. For this, scientists decided to resort to the OpenPose library, developed by the Carnegie Mellon University. Based on the received data, the neural network compiles a 3D model of the person in each frame, and calculates the trajectory of the motion, obtaining a "motion sculpture".

At this stage, the image, according to the developers, suffers from a lack of textures and details, so the application integrates the "sculpture" in the original video. To avoid overlapping, MoSculp calculates a depth map for the original object and the 3D model.

MoSculp 3D Model
MoSculp 3D Model

The operator can adjust the image during the processing, select the "sculpture" material, color, lighting, and also what parts of the body will be tracked. The system is able to print the result using a 3D printer.

The team of researchers announced plans to further develop the MoSculp technology. Developers want to achieve from the processing system more than one object on the video, which is currently impossible. The creators of the technology believe that the program will be used to study group dynamics, social disorders and interpersonal interactions.

The principle of creating a 3D model based on human movements has been used before. For example, in August 2018, scientists at the same University of California at Berkeley demonstrated an algorithm that transfers the movements of one person to another.