Daily Reading is a task I always want to challenge for years. ~~Although I also summarized papers on a specific topic, I failed to track those miscellaneous papers I read to broaden my horizon.~~ Writing summaries is an efficient way to document my thoughts. Hope that I can keep doing it and share more papers with you.

Notice that most papers express their ideas in an extremely succinct way. Therefore when I am summarizing, many details might be left out. If you are interested in any of those, **please read the original paper on your own**. My summary serves more as a memory-enhancement and a discussion on possible future works. Talks are now also included in the Reading List.

Daily Reading List

- [DR023]
**Deep Variational Information Bottleneck**- Restrict the embedding by limiting the mutual information between the input and the encoding (don’t remember the details).

- [DR022]
**Machine Theory of Mind**- An observer is asked to predict how another agent will behave in tasks. They design some supervised tasks to train that observer.

- [DR021]
**SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient Daily Reading**- Uses LSTM as the generator as well as the policy and CNN as the discriminator as well as the reward function. Train in a policy gradient way.

- [DR020]
**Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization**- ICCV 2017> A generalized improvement of CAM (Class Activation Mapping), which can localize the most/least contributing area to a specific class in the image. Interesting insights of various tasks and networks are discovered using Grad-CAM.

- [DR019]
**Efficient Estimation of Word Representations in Vector Space Daily Reading**- arXiv 2013> Word2Vec: Use hierarchical softmax to encode the output and design a new dataset to test the embedding.

- [DR018]
**iLab-20M: A large-scale controlled object dataset to investigate deep learning**- CVPR 2016> They introduced a dataset where every object is captured under different settings and studied several questions related to the invariability of CNN feature.

- [DR017]
**Understanding Neural Networks Through Deep Visualization**- ICML workshop 2015> Visualization of activation and features.

- [DR016]
**What is wrong with convolutional neural nets?**- MIT Talk 2014> Several arguments against CNN/pooling are given and a new model with concept “capsule” is introduced, where agreement/clustering is an important measurement.

- [DR015]
**Pointer Networks**- NIPS 2015> Apply the attention mechanism on the output of RNN so that it can generate outputs consisting of indices (positions) of the input sequence.

- [DR014]
**Dual Path Networks**- arXiv 2017> Combine ResNet with DenseNet to balance between feature re-usage and exploration.

- [DR013]
**Distral: Robust Multitask Reinforcement Learning**- arXiv 2017> Use Multi-Task Learning to improve RL by regularizing task policy over the policy space with a shared/average policy.

- [DR012]
**Attention is all you need**- arXiv 2017> Use attention mechanism to replace RNN and CNN in machine translation.

- [DR011]
**Neural Turing Machine**- arXiv 2014> Attach neural network with an external memory module.

- [DR010]
**Relative Entropy Policy Search**- AAAI 2010> Limit the divergence between the updated policy and previous policy to prevent forgetting the learned experience in RL.

- [DR009]
**Learning to learn by gradient descent by gradient descent**- NIPS 2016> Use LSTM to learn the learning rate along with the major network.

- [DR008]
**Self-Normalizing Neural Networks**- arXiv 2017> Design an activation function so that the mean and variance of the activation will converge to a fixed point.

- [DR007]
**Unsupervised Pixel–Level Domain Adaptation with Generative Adversarial Networks**- CVPR 2017> Adapt by translating images across domains.

- [DR006]
**Learning Step Size Controllers for Robust Neural Network Training**- AAAI 2016> Hand-crafted some features based on the distribution of neuron activation to adjust the learning rate.

- [DR005]
**Depth Map Prediction from a Single Image using a Multi-Scale Deep Network**- NIPS 2014> Use a scale-invariant loss to reconstruct depth map.

- [DR004]
**Asynchronous Methods for Deep Reinforcement Learning**- ICML 2016> Have multiple agents to interact with the environment independently.

- [DR003]
**Domain Separation Networks**- NIPS 2016> Use two networks to learn shared and domain-specific feature respectively.

- [DR002]
**Scalable and Sustainable Deep Learning via Randomized Hashing**- arXiv 2016> Accelerate the max-dropout with hashing method.

- [DR001]
**Hierarchical Deep Reinforcement Learning: Integrating Temporal Abstraction and Intrinsic Motivation**- NIPS 2016> Split RL into two components: the meta-controller is to choose intrinsic goals and the controller is to accomplish them.

Pending LIst:

- Emphatic Temporal-Difference Learning
- One model to learn all
- Deep Neural Networks Rival the Representation of Primate IT Cortex for Core Visual Object Recognition
- Explaining and Harnessing Adversarial Example