arxivst stuff from arxiv that you should probably bookmark

March 23 2017

Deep MANTA Shows Fantastic Results with Vehicle Detection on KITTI Benchmark

A deep CNN is used for vehicle detection, orientation, and 3D location tasks, beating out current standards for these tasks. With self-driving cars around the corner, having models which can effectively identify other vehicles on the road is paramount.

deep learning state-of-the-art self-driving cars

Natural Language Object Retrieval Just Got a Lot Better

Identifying a specific object in an image has been a difficult task, with model accuracies being in the ~20% range. By using both spatial and temporal context, this new deep RL model is able to move a bounding-box around an image until it identifies the desired object with much higher accuracy on standard datasets such as RefCOCO (48.19%) and RefCOCOg (29.04%).

deep learning state-of-the-art object retreival

Utilizing Transfer Learning to Augment Small Medical Datasets

Medical datasets are notoriously small making it difficult to use modern deep learning techniques which rely on large amounts of data. This paper showcases that a deep network pre-trained on ImageNet, and then re-trained on a much smaller dataset provides strong results for melanoma screenings.

deep learning medical

March 22 2017

New Deep Network Techniques for Identifying Facial Expressions

Outperforming best results on the CK+, MMI, and Oulu-CASIA databases, this deep generative-contrastive network attempts to mimic the way human brains observe facial expressions.

state-of-the-art deep learning facial expressions

One-Shot Imitation Learning for Robots—Learn More with Less Data

Using a soft attention mechanism and a novel training method, you can create an RL agent that can learn from a few demonstartions of a task and then perform variants of that task!

reinforcement learning one-shot

Use GANs to Blend High-Resolution Images

New state of the art on the Transient Attributes Database. Integrate conditional GANs and gradient-based methods to generate high-resolution natural images.

state-of-the-art gan

March 21 2017

Boost Your Cross-Media Retrieval Process with Twitter100k

Training data for cross-media retrieval models is either lacking in diversity or written in formal language that does not match realistic applications. Twitter100k is a new large-scale dataset that addresses these issues and allows you to train your model on realistic data!


Counterfactual Fairness: Combat the Inherent Social Biases of Your Dataset

When developing ML models with real-world impacts (such as loan lending or predictive policing), it is important to take into account the different social biases that may arise towards individuals of a particular race, gender, or sexuality and compensate for these biases effectively. The Counterfactual Fairness model attempts to do just that.

dataset sociology

Mask R-CNN—new state of the art in image segmentation

Need a simple and flexible model that gives competitive results out-of-the-box? Try Mask R-CNN. An extension of Faster R-CNN, this network beats top results on all three tracks on the COCO suit of challenges.

state-of-the-art image segmentation instance segmentation cnn deep learning

March 20 2017

Diversifying Artificially Generated Image Descriptions

Combine reinforcement learning techniques with GANs to improve image captioning.

gan image caption rl

Getting Around Feature Engineering in Image-Based Anomaly Detection with GANs

Capturing markers for disease progression can be hindered by a limited vocabulary of known markers, and explicitly discovering/identifying these markers is time consuming. Instead, the authors propose a GAN to learn a manifold and teach it to score anomalies!

gan images unsupervised

Automatically Address Holes in Your Knowledge Base

Relational Graph Convolutional Networks (R-GCNs) can be used for link prediction and entry classification much more efficiently than walk-based models for statistical relational learning. Ensure that your large relational database isn’t missing any data!

big data gcns databases social networking

March 17 2017

SAENs Learn on Large Social Networks with Deep Learning

Most networks aimed towards studying graphs are limited by the size of the graph itself. SAENs (Shift Aggregate Extract Networks) are a novel technique that utilize a deep hierarchical network to break this barrier and allow learning on much larger graphs, especially those with high connectivity like social networks!

deep learning graphs large datasets

Security for your ML Model

Thinking of building a security critical app with Machine Learning at it’s core? This paper explores the overlap between current ML security problems and existing techniques used in the digital watermark space. Stay safe!


March 16 2017

CNNs with Sliding Windows Perform Well on EEGs and Other Timeseries

The authors use standard convolutional neural networks and a “cropped training strategy” (sliding input windows) to reach accuracies similar to state-of-the-art algorithms.

medicine eeg cnn time series

State Of The Art Results

New Datasets

Recent Abstracts