Microsoft Research Open Data Project lays down evolving standards for data access and reproducible research

Microsoft Research has announced that it will be adopting the Open Use of Data Agreement (O-UDA) data for several datasets that the company offers. 

The Open Use of Data Agreement (O-UDA) is intended to make it easier for individuals and organizations that want to share data to do so, with minimal requirements for users and no restrictions on use. The O-UDA is complemented by the Computational Use of Data Agreement (C-UDA), an agreement intended for situations where a specific data use scenario is desirable or required.

It is not appropriate for datasets that include any data that might include materials subject to privacy laws (such as the GDPR or HIPAA) or other unlicensed third-party materials. 

The O-UDA meets the open definition: it does not impose any restriction with respect to the use or modification of data other than ensuring that attribution and limitation of liability information is passed downstream. In the research context, this implies that users of the data need to cite the corresponding publication with which the data is associated. This aids in findability and reusability of data, an important tenet in the FAIR guiding principles for scientific data management and stewardship.

Microsoft recognizes that in certain cases, datasets useful for AI and research analysis may not be able to be fully “open” under the O-UDA. For example, they may contain third-party copyrighted materials, such as text snippets or images, from publicly available sources. 

The law permits their use for research, so following the principle that research data should be “as open as possible, as closed as necessary,” Microsoft developed the Computational Use of Data Agreement (C-UDA) to make data available for research while respecting other interests. The software giant will prefer the O-UDA where possible, but perceives the C-UDA as a useful tool for ensuring that researchers continue to have access to important and relevant datasets.

Microsoft researcher John Krumm and collaborators collected GPS data from 21 people who carried a GPS receiver in the Seattle area. Users who provided their data agreed to it being shared as long as certain geographic regions were deleted. 

This work covers key research on privacy preservation of GPS data as evidenced in the corresponding paper, “Exploring End User Preferences for Location Obfuscation, Location-Based Services, and the Value of Location,” which was accepted at the Twelfth ACM International Conference on Ubiquitous Computing (UbiComp 2010). The paper has been cited 147 times, including for research that builds upon this work to further the field of preservation of geo-privacy for location-based services providers.

Another example dataset is that of labeled hand images and video clips collected by researchers Eyal Krupka, Kfir Karmon, and others. The research addresses an important computer vision and machine learning problem that deals with developing a hand-gesture-based interface language. 

The data was recorded using depth cameras and has labels that cover joints and fingertips. The two datasets included are FingersData, which contains 3,500 labeled depth frames of various hand poses, and GestureClips, which contains 140 gesture clips (100 of these contain labeled hand gestures and 40 contain non-gesture activity). 

The research associated with this dataset is available in the paper “Toward Realistic Hands Gesture Interface: Keeping it Simple for Developers and Machines,” which was published in Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems.

Finally, the FigureQA dataset generated by researchers Samira Ebrahimi Kahou, Adam Atkinson, Adam Trischler, Yoshua Bengio and collaborators, introduces a visual reasoning task for research that is specific to graphical plots and figures. 

Microsoft Research Open Data project was conceived from the start to reflect Microsoft Research’s commitment to fostering open science and research and to achieve this without compromising the ethics of collecting and sharing data. The company aims to make it easier for researchers to maintain provenance of data while having the ability to reference and build upon it.

Microsoft Research shows that logarithmic mapping allows for low discount factors by creating action gaps similar in size

While reinforcement learning (RL) has seen significant successes over the past few years, modern deep RL methods are often criticized for how sensitive they are with respect to their hyper-parameters. One such hyper-parameter is the discount factor, which controls how future rewards are weighted compared to immediate rewards.

The objective that one wants to optimize in RL is often best described as an undiscounted sum of rewards (for example, maximizing the total score in a game). 

In practice, however, a discount factor is introduced to avoid some of the optimization challenges that can occur when directly optimizing on an undiscounted objective. And while in theory a discount factor can take on any value between 0 and 1, in reality good performance is only obtained for a small subset of values close to 1.

Microsoft Research will present a paper at the thirty-third Conference on Neural Information Processing Systems (NeurIPS 2019), called “Using a Logarithmic Mapping to Enable Lower Discount Factors in Reinforcement Learning.” 

The company will present a novel hypothesis as to why the effective discount-factor range is so small. Also, based on its hypothesis, Microsoft outlined a technique to avoid this discount-factor sensitivity and introduce a method, which is called logarithmic Q-learning, based on this technique. 

Logarithmic Q-learning is the first method able to achieve good performance for low discount factors on sparse-reward tasks with function approximation. In addition, the method not only reduces discount-factor sensitivity, but can also improve performance altogether.

While a logarithmic mapping function makes action gaps more homogenous, there are a number of challenges to overcome to build a robust algorithm, based on this outcome, that can be applied to stochastic domains with both positive and negative rewards. 

Microsoft managed to overcome these challenges and develop an algorithm, logarithmic Q-learning, that can be applied to general tasks and has convergence guarantees under standard conditions. These results do not only show great performance for low discount factors. Early performance of high discount factors is also better than it was without the logarithmic mapping.

Microsoft Research releases PipeDream system to train deep neural networks using pipeline parallelism

Microsoft researchers in the Systems Research Group, along with students and colleagues from Carnegie Mellon University and Stanford University, have proposed a new way to parallelize DNN training. The system, called PipeDream, achieves up to 5.3 times faster training time than traditional approaches across a range of models.

DNN training happens in iterations of forward and backward pass computations. In each iteration, the training loop processes a minibatch of input data and performs an update to the model parameters. The most common approach to parallelize DNN training is a method called data parallelism, which partitions input data across workers (accelerators).

Unfortunately, despite advances in performance optimizations to speed up data parallelism, it can suffer from high communication costs at scale when training on cloud infrastructure. Also, rapid increases in GPU compute speed over time will further shift the bottleneck of training towards communication for all models.

Deep Neural Networks (DNNs) have facilitated tremendous progress across a range of applications, including image classification, translation, language modeling, and video captioning. DNN training is extremely time-consuming, needing efficient multi-accelerator parallelization.

PipeDream, a system developed as part of Microsoft Research’s Project Fiddle, introduces pipeline parallelism, a new way to parallelize DNN training by combining traditional intra-batch parallelism (model and data parallelism) with inter-batch parallelism (pipelining).

PipeDream revisits using model parallelism for performance, as opposed to the traditional motivation of working set size limitations for training large models. It uses pipelining of multiple inputs to overcome the hardware efficiency limitations of model-parallel training. A general pipeline parallel setup involves layers split across stages, with each stage potentially replicated and running data parallel. 

Multiple batches are injected into the pipeline to keep it full in steady state. Pipeline-parallel training, in most cases, communicates far lesser data than data-parallel training as it needs to communicate only the activations and gradients at the boundary of two stages. In steady state all workers are busy doing work with no pipeline stalls as in model-parallel training.

As DNNs do not always divide evenly among available workers, PipeDream may decide to use data parallelism for some stages—multiple workers can be assigned to a given stage, processing different minibatches in parallel. PipeDream uses a scheduling algorithm called 1F1B to keep hardware fully utilized, while achieving semantics similar to data parallelism.

PipeDream has been built to use PyTorch (an earlier version of PipeDream uses Caffe). Its evaluation, encompassing many combinations of DNN models, datasets, and hardware configurations, confirms the training time benefits of PipeDream’s pipeline parallelism. 

Compared to data-parallel training, PipeDream reaches a high target accuracy on multi-GPU machines up to 5.3 times faster for image classification tasks, up to 3.1 times faster for machine translation tasks, 4.3 times faster for language modeling tasks, and 3 times faster for video captioning models. PipeDream is also 2.6 to 15 times faster than model parallelism and up to 1.9 times faster than hybrid parallelism.