已被顶级Q1期刊IEEE Transactions on Cybernetics接收
IEEE: Decoder Choice Network for Meta-Learning
arXiv: Decoder Choice Network for Meta-Learning

Abstract

We design a new optimization-based model with a new task embedding in order to get better performance on Omniglot and miniImageNet dataset. As a result, We achieved a lot of state-of-the-art in experiments. Among them, a structure similar to searching rule is combined in the model. The paper is still in the submission, so the details of the model cannot be given here.

针对元学习的数据处理方法

arXiv: Task Augmentation by Rotating for Meta-Learning

设计了针对彩色图片的元学习任务的数据处理方式，实验验证了方法的有效性，并以较大的间隔在CIFAR-FS, FC100和miniImageNet数据集上超过了当前的state-of-art结果。

深度网络中卷积操作的改进方法

This work was accepted by Neurocomputing.
Neurocomputing: DK-CNNs: Dynamic kernel convolutional neural networks

设计了新的卷积运算方式在多个图片识别任务上获得了比常规的卷积运算获得了更高的准确度。

arXiv: Stock Prices Prediction using Deep Learning Models

Examples of stock forecast curves

This work was accepted by UKCI.
arXiv: Gradient Boost with Convolution Neural Network for Stock Forecast

Abstract

This paper proposes a model combining the ensemble techniques of extreme gradient boost (XGBoost) [1] and 1D convolution neural network (CNN), called as CGBoost, to obtain a better performance. This work also tried to fit 6 market curves simultaneously using CGBoost and got better results, it is call as CGBoost6.

Face detection: MTCNN; Face recognition: Facenet.

Video of face attendance app.

Reference:
Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks.
FaceNet: A Unified Embedding for Face Recognition and Clustering

Deep Q-Learning on Cartpole

Reference: Human-level control through deep reinforcement learning.

Hybrid Reward Architecture on Cartpole

Reference: Hybrid Reward Architecture for Reinforcement Learning.

The images generated by ACGAN. Dataset is crifar-10. The images of different classes in different columns.

Reference: Conditional Image Synthesis With Auxiliary Classifier GANs.

Mathematical process about Deterministic Policy Gradient Algorithms.

目录

1. Policy Gradient
2. Stochastic Actor-Critic
3. Gradients of Deterministic Policies
4. Limit of the Stochastic Policy Gradient
5. Compatible Function Approximation
6. DDPG

An example of DDPG on pendulum