README

This repository contains the summaries of projects, and code implementation from NIPS 2017. It also has a few tips to use to make an awesome summary and implmentation.

Tip - Structuring your projects

There is an awesome skeleton project structure for Machine Learning projects. You can find it here : https://drivendata.github.io/cookiecutter-data-science/ . It has an easy to understand strcture, and uses click library for command line arguments, which greatly simplifies the code , allowing readers to focus on what matters most.

To create an empty structure, all you have to do is run

cookiecutter https://github.com/drivendata/cookiecutter-data-science

How to add your own Summary

1. Fork this repo
2. Add a folder with your Paper Name
3. Create a Readme.md file in that folder
4. Write your summary there. You can include additional files (E.g. reference images) under your folder
5. Update the Readme.md at the root of this project (This File), to include a short summary of your paper
6. Submit a pull request on github

This is done to prevent merge conflicts

---

NIPS - 2017 Paper Summaries

Towards Accurate Binary Convolutional Neural Network

Link: https://papers.nips.cc/paper/6638-towards-accurate-binary-convolutional-neural-network.pdf

Summary : This paper talks about using only -1 or +1 as weights and activations while building CNNs. Paper claims that this tweak to the traditional CNNs uses significantly lesser amount of memory, faster inference in runtime on test data and comparable performance on IMAGENET.

To-do to implement the paper:

• Constrain the weights to {-1, +1} - This ensures convolutions are only additions or subtractions
• Use five binary activations instead of two.
• ...(Work in progress)

Implementations and Results:

• Detailed Summary and Results: Link-to-Your-Folder-Name-In-This-Repo. Code: Your-Repo-Here.

---

Selective Classification for Deep Neural Networks

Link: https://papers.nips.cc/paper/7073-selective-classification-for-deep-neural-networks.pdf

Summary : This paper proposes a new method to construct a selective classifier for a given trained DNN. At test time the classifier rejects instances as needed to grant the desired risk. Performance is measured on ImageNet, CIFAR-10, CIFAR-100 datasets.

To-do to implement the paper:

• Implement SR and MC-dropout confidence-rate functions, κf , and the induced rejection function, gθ(x|κf ) on trained models.

---

Learning multiple visual domains with residual adapters

Link: https://nurture.ai/p/10625bb6-46aa-4b03-aff4-c9b558852de0

Summary : This paper proposes a residual neural net based network that can work well on multiple image domains all at the same time, i.e. without forgetting learning from previous domains (unlike just finetuning). Specifically, the paper discusses the architecture and performance for a new 'decathlon challenge' that contains images from (a) Aircraft, (b) CIFAR-100, (c) Daimler Pedestrians, (d) Describable Textures, (e) German Traffic Signs, (f) ILSVRC (ImageNet) 2012, (g) VGG-Flowers, (h) OmniGlot, (i) SVHN, (j) UCF101 Dynamic Images. The proposed architecture is a modification of Wide Residual Network Resnet28 - additional simplified residual blocks added to the network. The idea is that 90% of the learnt weights are shared for all the datasets and 10% parameters are specifc to each domain.

To-do to implement the paper:

• Train WRN Resnet28 on the imagenet dataset to get baseline weights, then add the additional residual blocks to the network.
• Train the new network on each of the 10 datasets and carry out evaluation for the decathlon challenge as explained in the paper.