arxivst stuff from arxiv that you should probably bookmark

April 25 2017

Automatic Evaluation of Generated Summaries

Looking for an automated way to assess the quality of your generated summaries? The authors of this paper generate a bunch of questions based off of a source text, and then ask those questions using the generated text as the database. If the information is found in both, then the generated text can be said to be a good representation of the source text. Seems reasonable to me. I might try it for this site.

nlp duc

Recursive Models Write Programs

A program that writes programs needs recursion is the conclusion of this new paper. The authors take several standard programming tasks and propose a model architecture that proves it has perfect generalizability with small amounts of training data.

recursion program-semantics

End to End Module Networks

Visual Q&A models are incredibly useful, but many of them are still fragile. This paper proposes an end-to-end solution that does away with parsers while reducing errors by almost 50%. Holy cow, batman.

n2nmns vqa end-to-end state-of-the-art clevr

April 13 2017

General Approach to Real World Text Extraction

Breakthrough on the French Street Names dataset. This paper boasts a huge improvement over previous state-of-the-art models including Inception V3 and Inception Resnet V2. And since that wasn’t hard enough, they also tested it against Google Street View, cause it’s all like nbd.

french-street-names state-of-the-art google-street-view

MAGAN, Better than BEGAN

What’s better than BEGAN? MAGAN! Or at least that’s what the authors of this paper are saying. They are probably right because a) they have a simpler setup and b) they use public datasets so their work is reproducible. Plus they are promising to release the code on Github in the near future.

magan mnist cifar-10 celeba state-of-the-art began

3d Point Cloud Dataset and Benchmark

New 3d point cloud dataset and benchmark with semantic labels. Should be a lot of use to people doing AR. They did a nice job with the benchmark too, and have an automated submission process with a public leaderboard.

dataset 3d point-cloud semantic3d

April 12 2017

New SOTA for VQA 1.0

There’s a lot of research around VQA right now. This new paper gets a new state-of-the-art on VQA 1, an improvement of 0.4%. It’s by a small margin, but it’s also using a simpler architecture.

vqa natural language state-of-the-art

Autoencode Your Networks

Autoencoding for networks. Struc2Vec gives you a new way to view and understand your networks.

node embeddings feature learning struc2vec

Adversarial Attack Prevention

New paper out of Stanford, Pennsylvania State and Google Brain delves into what happens when a ML model is attacked, and how to incorporate robustness. It’s an interesting read in an interesting space.

transferability attacks adversarial subspaces mnist drebin

April 11 2017

Build a Pong Simulator

Not just for pong, this paper simulates environments from 2d Atari games to 3d racing sims. It has DQN Scores on a bunch of games (and goes into high detail on how they got those scores), but they don’t explicitly compare those score with current state-of-the-art.

environment simulator dqn

Predicting Recomendations with TransNets

Building a recommendation engine? This paper gets a new state-of-the-art on ratings predictions. Show your users what they want to see.

state-of-the-art yelp17 amazon-product-data

April 07 2017

Use Machine Learning to Write Your Code For You

I am looking forward to the day when I can write a model that writes models for me. Maybe someone will turn it into an API and I won’t have to write anything. This paper gets us another step closer. Working with Python and other high-level languages, they have a new state of the art in code generation.

code-generation python state-of-the-art

Weakly Supervised Fine-Grained Classifications

Telling the difference between various types of red sedans or crows is hard for humans, let alone neural nets. This paper does it better and cheaper.

image-segmmentation cub-200-2011 cars-196 oxford-iiit-pet

Automatic Question Generation

Looking for ways to generate a FAQ from your docs? This paper points to a potential direction.

squad sequence-to-sequence

April 06 2017

Better Faster Segmentation

The next time we do a Kaggle comptition we’ll try this technique. When working on an image segmentation problem, it’s those last couple pixels around the edges that make all the difference in your mIoU.

segmentation pascalvoc-2012 cityscapes-dataset

State Of The Art Results

New Datasets

Recent Abstracts