MathNAS: If Blocks Have a Role in Mathematical Architecture Design

NeurIPS2023   Paper   Code

Wang Qinsi*, Jinhan Ke*, Zhi Liang, Sihai Zhang

Task : Image classification/text translation.
What : Automatic design of neural network architecture.
Motivation : The design of network architecture for specific tasks and specific resource constraints requires considerable time and computing power costs.
Insights : We propose a block performance evaluation formula and decompose the network design into the block designs.
Advantages : Reduces the cost of network architecture design from exponential level to polynomial level.

DGL: Device Generic Latency Model for Neural Architecture Search on Mobile Devices

IEEE TMC 2023   Paper

Wang Qinsi, Sihai Zhang

Task : Image classification.
What : Neural network architecture design on mobile devices.
Motivation : Designing network architecture for specific devices and latency constraints is difficult.
Insights : We propose a network latency prediction formula by considering specific hardware configuration parameters.
Advantages : For specific mobile devices, a network that meets latency requirements can be quickly designed without prior deployment.

Diff-Transfer: Model-based Robotic Manipulation Skill Transfer via Differentiable Physics Simulation

Submitted to ICLR 2023   Paper   Vedio

Yuqi Xiang, Feitong Chen, Qinsi Wang, Gang Yang, Xiang Zhang, Xinghao Zhu, Xingyu Liu, Lin Shao

Task : Control the robotic arm to close the grill/change the clock/open the door/open the drawer.
What : Transfer learning in the field of robotics.
Motivation : For intelligent robots, the ability to transfer mastered skills to complete a range of similar but novel tasks is crucial.
Insights : We introduce Diff-Transfer, a novel framework that leverages differentiable physics simulations to efficiently transfer robotic skills.
Advantages : The robot can quickly learn challenging transfer tasks.

Differentiable Actor-Critic: Efficient Use of Gradients in Model-free RL

The project is ongoing.

Wang Qinsi, Lin Shao

Task : Reinforcement learning control task.
What : Gradient-based model-free reinforcement learning algorithms.
Motivation : Model-free reinforcement learning algorithms are severely limited in their applicability to complex real-world domains due to the large amounts of training data they require.
Insights : We blend gradient-based methods with model-free methods, gaining the benefits of both.
Advantages : Fast training of reinforcement learning algorithms.

Research on Invasive Insect Image Recognition Based on Artificial Intelligence

ICBAIE 2021   Paper

Wang Qinsi, Jiang Xueyi, Xiao Xiaolong, Wang Xueshun

Task : Image recognition.
What : Optimization of recognition algorithms for real-world problems.
Motivation : There are problems with poor quality and a small proportion of correct samples when performing image recognition of insect pictures.
Insights : We combine image recognition with species distribution and propose a new framework to improve accuracy in practical applications.
Advantages : Significantly improves sample recognition accuracy on real-world data sets.

Transportation Planning Method Based on Improved Kruskal Algorithm

Computer Applications 2021   Paper

Wang Qinsi, He Yuhui

Task : Transportation planning.
What : An effective transportation planning method for transportation planning problems with transfer center constraints.
Motivation : Transportation problem with links is an NP problem.
Insights : We build a minimum spanning tree model with constraints from the perspective of graph theory and improve Kruskal algorithm.
Advantages : Provides relatively accurate and complete solutions to this type of transportation planning problem.

Image Segmentation Based on Level Set

Task :Use level set segmentation algorithm to complete image segmentation with python. Excellent segmentation is achieved on multiple classic images such as cups, twocells and vessels.

Image Retrieval Based on Inverted Index

Task: The SIFT features of 1000 images are randomly sampled. And the hierarchical k-means algorithm is used to train the visual codebook (codebook size is 10000). L2 normalization is used to process the visual histogram. The inverted index is used for the final output.

Handwritten Word Recognition

Task: A multinomial logistic regression model is used for image classification to recognize each letter separately. The conditional probability formula for words is constructed by considering the letter classification and the probability of consecutive occurrence of two adjacent letters. The Viterbi algorithm is used to find the sequence of letters that maximizes the conditional probability of a word to achieve word recognition.

Contour Extraction

Task: Laplacian and Sobel operators are used to extract contours. The program is optimized to be more suitable for GPU parallel computing, and obtains a speedup ratio of 30 compared to CPU.

Lyric Prediction

Task: 1000 lyrics (from Jay Chou's songs) are used as training samples. Character indices are built and onehot vectors are used as input to the network. A recurrent neural network is trained and used for lyrics prediction. This project is my work when I participated in the 2020 Public Welfare AI Learning Together Training Camp.