[last updated on Aug. 4, 2017]
Think about Deep Learning
I built this webpage to host my notes, thoughts and resources I collected during my study of deep learning (DL). Some of the tutorials were written by myself, while a large number of online rescoures are also included. Since my researches focus on medical image analysis, my DL study was aiming more at medical imaging applications.
Introduction to Convolutional Neural Network (CNN)
See my post here
Introduction to Recurrent Neural Network (RNN)
See my post here
Introduction to Deep Reinforcement Learning
I don’t have too much experience in deep reinforcement learning (DRL). To my understanding, this is a more general deep learning algorithm, but relatively less tailored for specific computer vision problems.
DRL is attracting more and more attentions nowadays. It is probably due to the success of DeepMind (AlphaGo). One key algorithm is the so-called Deep Q-Learning. Some implementations can be found at link1, link2, link3, and link4.
Jump Start:
- Essential Idea of DRL: (link). Several additional sources can be found in the external links in this awesome post.
- High-level description of each key component in DRL: Andrej Karpathy’s blog (link)
- Complete Introduction: David’s course (link) and John Schulman’s course (link)
More resources:
-
A very interesting work published recently is to use Deep Q-learning to accelerate the training of CNN (paper).
-
The most recent tutorial (link) on ICML 2016. (Note: Really inspiring!)
-
An example to practice (link)
Papers to read
- Must read: A paper from Arterys, a startup in the bay area. (Link) The method itself is not of great interest. But, I am eager to see how deep learning is used in startups and how they present the methods.
- Another paper by a startup, Imagia. (Link) Co-authored by Yoshua Bengio. The method is also very interesting.
- Dual Path Network Link
- Active learning and fine tuning for biomedical image analysis Link
- Quick Read:
- Dynamic steerable blocks Link
- steerable CNN Link
- Fine tuning for medical image analysis Link
- segAN Link
- Morphology error detection Link
- dilated residual network Link
- anatomically constrained 3D FCN Link
- for tracking Link
- 3D FCN and random walk Link
- Hierarchical 3D FCN Link
- SeGAN Link
- DNN for multi-task multi-modality medical image segmentation Link
- review on DNN for semantic segmentation Link
- Proximal Segmentation in medical images Link
- Wavelete Residual Network Link
- Multi-modale CNN for brain tumor segmentation Link
- Recurrent context learning Link
- 3D deep supervision Link, and another similar one Link
- Predicting the ambiguity of foreground object Link
Deformable Convolution Network Link- Guide to semantic segmentation with deep learning, by Qure.ai Link
- Xception, depthwise separable convolutions. Link
- CortexNet Link
- Rethinking Atrous Convolution Link
- Train longer, generalize better Link,Github
- Evaluation without ground truth:
- Train ImageNet in 1 Hour Link
- Apple’s ML journal: Link
- The future of deep learning, by the Francois Chollet. Link
- Pixel deconvolution network Link
Dense Transformer Network Link- An intereting application of adversarial network in segmentation Link
- Network Dissection Link
- Interpretable explanation of black box by meaningful pertubation Link,Github
- Incorporating Built-in Prior to deep learning for segmentation: Link
- Registration:
- Unsupervised Learning: This is the most urgent thing I want to learn.
- A general overview: link
Insight from Yann LeCunn and the Facebook AI group about adversial network, a new paradigm for unsuperised learning. linkand a key paper introducing DCGAN.- A tutorial on Variational Autoencoder, paper
- A comprehensive tutorial on the concept of “adversarial network”, a very promising solution to unsupervised learning. link
- Instance Segmentation using deep learning: performing segmentation on the instance level (maybe one by one in a recurrent framework or in one shot)
- Combinining with CRF link1 and link2
ENet: for real-time application link- Dealing with high resolution images, paper
- Image colorization and syhthesis using deep learning:
MUST READ: Online hard example mining (OHEM). I have thought about this idea for a long time. It is finally solved.- An important paper: Bayesian Deep Learning. link. The key idea is to elaborate a model to unify deep learning and graphical models.
- 3D segmentation:
- Weakly Supervised Segmentation
Point-level supervision for semantic segmentation: (paper). I have been seeking an elegant way to utilize point level supervision for a long time. This work provides some good experiments and thoughts. In this paper, however, the labelled points in each image still requires to cover all instance (one point per intance), which is not feasible in certain circumstances.- Seed, Expand and Constrain. link
- Expressive power of DNN: People may say DNN is able to express very complex non-linear function. But how to quantify such capability? I just found this paper to read. link
- New Schemes for parameter initialization in CNN, two papers in ICLR 2016: paper1 paper2
- Weight initialization for RNN. paper
- Weight Normalization: a new scheme for better training of CNN. link
- SqueezeNet: A CNN using much less memory without reducing accuracy. paper
- Style transfer paper and code
- Extreme Learning machine, another type of machine learning algorithms which is closers to human brain review paper
- Binary CNN: paper1 and paper2
Combining CNN with Active Contour: paper- CNN on graphs: Extend classic CNN which applies on grids (e.g., images) to graph data paper1 and paper2
- Inside-Outside Network: A method combining information insider region of interest and outside the region of interest. (link)
- A comprehensive comparison of different choices in each component in CNN, such as different activation function or different pooling. link
- CNN in convert sketch to image, link
- Piecewise Training link
- Object detection using R-FCN, link, code
- Pixel Recuurent Neural Network (best paper in ICML 2016), link
- Edge detection by DNN. github link
Densely connected CNN linkWide Residual Network Torch code,paper- Guide to Tensorflow, link1, link2
- Deep learning and human-in-the-loop. link
- Translation aware FCN link
- How to implement your own layers in Torch7. link
- Learning Recursive Filters for Low-Level Vision via a Hybrid Neural Network. link
- Open problems in AI. link
- Depthwise Separable Convolutions. link
- Bilateral Solver for edge aware smoothing in deep learing. link
- Multiple instance learning in neural network. link
- Deep vision tracking benchmark. link
- CliqueNet. link