Khalid Ashraf

FireCaffe: Near-Linear Acceleration of Deep Neural Network Training on Compute Clusters

Long training times for high-accuracy deep neural networks (DNNs) impede research into new DNN architectures and slow the development of high-accuracy DNNs. In this paper we present FireCaffe, which successfully scales deep neural network training across a cluster of GPUs. We also present a number of best practices to aid in comparing advancements in methods for scaling and accelerating the training of deep neural networks. The speed and scalability of distributed algorithms is almost always limited by the overhead of communicating between servers, DNN training is not an exception to this rule. Therefore, the key consideration here is to reduce communication overhead wherever possible, while not degrading the accuracy of the DNN models that we train. Our approach has three key pillars. First, we select network hardware that achieves high bandwidth between GPU servers - Infiniband or Cray interconnects are ideal for this. Second, we consider a number of communication algorithms, and we find that reduction trees are more efficient and scalable than the traditional parameter server approach. Third, we optionally increase the batch size to reduce the total quantity of communication during DNN training, and we identify hyperparameters that allow us to reproduce the small-batch accuracy while training with large batch sizes. When training GoogLeNet and Network-in-Network on ImageNet, we achieve a 47x and 39x speedup, respectively, when training on a cluster of 128 GPUs.

Probing electric field control of magnetism using ferromagnetic resonance

Exchange coupled CoFe/BiFeO3 thin-film heterostructures show great promise for power-efficient electric field-induced 180° magnetization switching. However, the coupling mechanism and precise qualification of the exchange coupling in CoFe/BiFeO3 heterostructures have been elusive. Here we show direct evidence for electric field control of the magnetic state in exchange coupled CoFe/BiFeO3 through electric field-dependent ferromagnetic resonance spectroscopy and nanoscale spatially resolved magnetic imaging. Scanning electron microscopy with polarization analysis images reveal the coupling of the magnetization in the CoFe layer to the canted moment in the BiFeO3 layer. Electric field-dependent ferromagnetic resonance measurements quantify the exchange coupling strength and reveal that the CoFe magnetization is directly and reversibly modulated by the applied electric field through a ~180° switching of the canted moment in BiFeO3. This constitutes an important step towards robust repeatable and non-volatile voltage-induced 180° magnetization switching in thin-film multiferroic heterostructures and tunable RF/microwave devices.

Nature of magnetic domains in an exchange coupled BiFeO3/CoFe heterostructure

We use the micromagnetic model to simulate magnetization in a multidomain BiFeO3(BFO)/CoFe bilayer. We show that CoFe couples to weak ferromagnetism on the BFO surface and breaks into domains that correspond exactly with BFO domains. Switching the direction of the BFO spins switches the corresponding CoFe domains and reverses the net magnetization. Since magnetization is coupled to polarization in BFO, this demonstrates a mechanism for controlling magnetization reversal with an electric field. Comparison with experimental values of BFO surface moment and hysteresis allows us to extract BFO/CoFe exchange and spin canting energies and predict behavior for domains of varying sizes.

Audio-Based Multimedia Event Detection with DNNs and Sparse Sampling

This paper presents advances in analyzing audio content information to detect events in videos, such as a parade or a birthday party. We developed a set of tools for audio processing within the predominantly vision-focused deep neural network (DNN) framework Caffe. Using these tools, we show, for the first time, the potential of using only a DNN for audio-based multimedia event detection. Training DNNs for event detection using the entire audio track from each video causes a computational bottleneck. Here, we address this problem by developing a sparse audio frame-sampling method that improves event-detection speed and accuracy. We achieved a 10 percentage-point improvement in event classification accuracy, with a 200x reduction in the number of training input examples as compared to using the entire track. This reduction in input feature volume led to a 16x reduction in the size of the DNN architecture and a 300x reduction in training time. We applied our method using the recently released YLI-MED dataset and compared our results with a state-of-the-art system and with results reported in the literature for TRECVIDMED. Our results show much higher MAP scores compared to a baseline i-vector system - at a significantly reduced computational cost. The speed improvement is relevant for processing videos on a large scale, and could enable more effective deployment in mobile systems.

Kickstarting the Commons: The YFCC100M and the YLI Corpora

The publication of the Yahoo Flickr Creative Commons 100 Million dataset (YFCC100M)--to date the largest open-access collection of photos and videos--has provided a unique opportunity to stimulate new research in multimedia analysis and retrieval. To make the YFCC100M even more valuable, we have started working towards supplementing it with a comprehensive set of precomputed features and high-quality ground truth annotations. As part of our efforts, we are releasing the YLI feature corpus, as well as the YLI-GEO and YLI-MED annotation subsets. Under the Multimedia Commons Project (MMCP), we are currently laying the groundwork for a common platform and framework around the YFCC100M that (i) facilitates researchers in contributing additional features and annotations, (ii) supports experimentation on the dataset, and (iii) enables sharing of obtained results. This paper describes the YLI features and annotations released thus far, and sketches our vision for the MMCP.

Shallow Networks for High-accuracy Road Object-detection.

The ability to automatically detect other vehicles on the road is vital to the safety of partially-autonomous and fully-autonomous vehicles. Most of the high-accuracy techniques for this task are based on R-CNN or one of its faster variants. In the research community, much emphasis has been applied to using 3D vision or complex R-CNN variants to achieve higher accuracy. However, are there more straightforward modifications that could deliver higher accuracy? Yes. We show that increasing input image resolution (i.e. upsampling) offers up to 12 percentage-points higher accuracy compared to an off-the-shelf baseline. We also find situations where earlier/shallower layers of CNN provide higher accuracy than later/deeper layers. We further show that shallow models and upsampled images yield competitive accuracy. Our findings contrast with the current trend towards deeper and larger models to achieve high accuracy in domain specific detection tasks.

Real-Time Observation of Local Strain Effects on Nonvolatile Ferroelectric Memory Storage Mechanisms

We use in situ transmission electron microscopy to directly observe, at high temporal and spatial resolution, the interaction of ferroelectric domains and dislocation networks within BiFeO3 thin films. The experimental observations are compared with a phase field model constructed to simulate the dynamics of domains in the presence of dislocations and their resulting strain fields. We demonstrate that a global network of misfit dislocations at the film-substrate interface can act as nucleation sites and slow down domain propagation in the vicinity of the dislocations. Networks of individual threading dislocations emanating from the film-electrode interface play a more dramatic role in pinning domain motion. These dislocations may be responsible for the domain behavior in ferroelectric thin-film devices deviating from conventional Kolmogorov-Avrami-Ishibashi dynamics toward a Nucleation Limited Switching model.

Phase field model of domain dynamics in micron scale, ultrathin ferroelectric films: Application for multiferroic bismuth ferrite

In this work, we report a massively parallel and time domain implementation of the 3D phase field model that can reach beyond micron scale and consider for arbitrary electrical and mechanical boundary conditions. The first part of the paper describes the theory and the numerical implementation of the model. A mixed-mode approach of finite difference and finite element grid has been used for calculating the nonlocal electrostatic and elastic interactions respectively. All the local and non-local interactions are shown to scale linearly up to thousands of processors. This massive paralleling allows to compare our results directly with experiments at the same length scales where the experiments themselves are performed. The second part of the paper presents results of ferroelectric domain switching in devices based on the multi-ferroic BiFeO3. We have particularly emphasized the importance of charge driven domain growth and the effect of electrical boundary conditions that explain the temporal evolution of f...

Effect of anti-ferromagnet surface moment density on the hysteresis properties of exchange coupled antiferromagnet-ferromagnet systems: The case of bismuth-ferrite

We propose a generic formalism to estimate the anisotropies, exchange energies, and the surface antiferromagnet (AFM) moment of AFM-ferromagnet (FM) heterostructure systems that show spin glass like behavior. This scheme provides quantitative agreement with multiple experiments on epitaxial bismuth ferrite (BFO)-FM system that have been reported recently. We find that a single value of the interface coupling energy can reproduce both the exchange bias and the coercivity enhancement observed in experiments. We also find a surprisingly high surface AFM moment density that agrees well with measured values. This high moment on the BFO surface is indicative of a significant modulation of magnetic properties at the BFO-FM interface.

Electric field induced magnetic switching at room temperature: Switching speed, device scaling and switching energy

The switching energy, speed and scaling behavior are calculated for a magneto-electric device incorporating the multiferroic material bismuth ferrite (BFO). For this purpose, a massively parallel phase field model is developed for simulation of the device. First, multidomain switching of thin film ferroelectric is shown to match with experimental measurements. We show that the switching energy in these devices can be an order of magnitude lower than other alternative approaches for magnetization reversal. We also show that the coercive voltage scales almost linearly in scaled BFO islands; however, the switching speed requires further study.