We introduce a data-free quantization method for deep neural networks that does not require fine-tuning or hyperparameter selection. It achieves near-original model performance on common computer vision architectures and tasks. 8-bit fixed-point quantization is essential for efficient inference in modern deep learning hardware architectures. However, quantizing models to run in 8-bit is a non-trivial task, frequently leading to either significant performance reduction or engineering time spent on training a network to be amenable to quantization. Our approach relies on equalizing the weight ranges in the network by making use of a scale-equivariance property of activation functions. In addition the method corrects biases in the error that are introduced during quantization. This improves quantization accuracy performance, and can be applied ubiquitously to almost any model with a straight-forward API call. For common architectures, such as the MobileNet family, we achieve state-of-the-art quantized model performance. We further show that the method also extends to other computer vision architectures and tasks such as semantic segmentation and object detection.

A glass with the composition 2BaO·TiO2·2.75SiO2 was annealed at 810 °C for 20 h. This led to surface crystallization. Immediately at the surface of the sample, a layer of Ba2TiSi2O8-type fresnoite crystals (layer I), with a thickness of approximately 7 μm, oriented with the crystallographic [101]-direction perpendicular to the surface, was formed. The pole of the (001)-plane rotates randomly around the [101]-direction. It is assumed that nucleation kinetics is decisive to the direction of growth. In the next layer (layer II), the crystals are oriented with the crystallographic [001]-direction (c-axis) perpendicular to the surface of the sample. This layer occurs at a distance of 7−60 μm from the surface. Here, crystals that are not oriented in that way hinder each other during crystal growth. At a distance > 60 μm, the orientation of the fresnoite crystals is random and is the result of volume crystallization. The main characterization method is electron backscatter diffraction/orientation imaging microscopy.

No abstract

A glass with the composition Ba 2 TiSi 2.75 O 9.5 was crystallized by electrochemically induced nucleation. A dc-voltage (4.8 V) was attached between a platinum crucible containing the melt and a platinum wire inserted into the melt. The platinum wire was the cathode. After 2 min, crystals were formed at the cathode, which grew toward the crucible. These crystals consisted of fresnoite Ba 2 TiSi 2 O 8 , were highly oriented, and showed dendritic growth. These structures were characterized in detail by using electron backscatter diffraction/orientation imaging microscopy. Only a few crystals deviate from the main orientation of the respective dendrite. They were located and quantified. Intermediate areas between dendrites were analyzed, and the true boundaries of areas with homogeneous orientation are presented. The orientation is not as pronounced in the first 600 μm next to the platinum wire.

While neural networks have advanced the frontiers in many applications, they often come at a high computational cost. Reducing the power and latency of neural network inference is key if we want to integrate modern networks into edge devices with strict power and compute requirements. Neural network quantization is one of the most effective ways of achieving these savings but the additional noise it induces can lead to accuracy degradation. In this white paper, we introduce state-of-the-art algorithms for mitigating the impact of quantization noise on the network's performance while maintaining low-bit weights and activations. We start with a hardware motivated introduction to quantization and then consider two main classes of algorithms: Post-Training Quantization (PTQ) and Quantization-Aware-Training (QAT). PTQ requires no re-training or labelled data and is thus a lightweight push-button approach to quantization. In most cases, PTQ is sufficient for achieving 8-bit quantization with close to floating-point accuracy. QAT requires fine-tuning and access to labeled training data but enables lower bit quantization with competitive results. For both solutions, we provide tested pipelines based on existing literature and extensive experimentation that lead to state-of-the-art performance for common deep learning models and tasks.

No abstract

An electromagnetic-based tracking and navigation system was evaluated for interventional radiology. The electromagnetic tracking system (CAPPA IRAD EMT, CASinnovations, Erlangen, Germany) was used for real-time monitoring of punctures of the lumbar facet joints and intervertebral disks in a spine phantom, three pig cadavers and three anaesthesized pigs. Therefore, pre-interventional computed tomography (CT) datasets were transferred to the navigation system and puncture trajectories were planned. A coaxial needle was advanced along the trajectories while the position of the needle tip was monitored in real time. After puncture tracts were marked with pieces of wire another CT examination was performed and distances between wires and anatomical targets were measured. Performing punctures of the facet joints mean needle positioning errors were 0.4 +/- 0.8 mm in the spine phantom, 2.8 +/- 2.1 mm ex vivo and 3.0 +/- 2.0 mm in vivo with mean length of the puncture tract of 54.0 +/- 10.4 mm (phantom), 51.6 +/- 12.6 mm (ex vivo) and 50.9 +/- 17.6 mm (in vivo). At first attempt, intervertebral discs were successfully punctured in 15/15 in the phantom study, in 12/15 in the ex-vivo study and 14/15 in the in-vivo study, respectively. Immobilization of the patient and optimal positioning of the field generator are essential to achieve a high accuracy of needle placement in a clinical CT setting.

The aim of this study was to prospectively evaluate the needle visualization and placement error and use of an electromagnetic field-based tracking navigation device for puncture procedures based on C-arm CT (CACT) images. A commercially available navigation device was mounted on an angiographic X-ray system setup for CACT. After the target was defined, needle placement was performed under real-time visualization of the virtual needle in CACT images. The final, real needle position was assessed by CACT. Punctures were performed in phantoms (n = 76) and in twelve patients (eight biopsies, three drainages, one injection). Procedure times, system error, user error and total error were assessed. In phantoms, mean total error was 2.3 +/- 0.9 mm, user error was 1.4 +/- 0.8 mm and system error was 1.7 +/- 0.8 mm. In the patient study, the targeted puncture was successful in all twelve cases. The mean total error was 5.4 mm +/- 1.9 mm (maximum 8.1 mm), user error was 3.7 +/- 1.7 mm, system error was 3.2 +/- 1.4 mm and mean skin-to-target time was less than 1 min. The navigation device relying on CACT was accurate in terms of needle visualization and useful for needle placement under both experimental and clinical conditions. For more complex procedures, electromagnetic field-based tracking guidance might be of help in facilitating the puncture and reducing both the puncture risk and procedure time.