Scalable and Practical Natural Gradient for Large-Scale Deep Learning.

Large-scale distributed training of deep neural networks results in models with worse generalization performance as a result of the increase in the effective mini-batch size. Previous approaches attempt to address this problem by varying the learning rate and batch size over epochs and layers, or ad hoc modifications of batch normalization. We propose scalable and practical natural gradient descent (SP-NGD), a principled approach for training models that allows them to attain similar generalization performance to models trained with first-order optimization methods, but with accelerated convergence. Furthermore, SP-NGD scales to large mini-batch sizes with a negligible computational overhead as compared to first-order methods. We evaluated SP-NGD on a benchmark task where highly optimized first-order methods are available as references: training a ResNet-50 model for image classification on ImageNet. We demonstrate convergence to a top-1 validation accuracy of 75.4 percent in 5.5 minutes using a mini-batch size of 32,768 with 1,024 GPUs, as well as an accuracy of 74.9 percent with an extremely large mini-batch size of 131,072 in 873 steps of SP-NGD.

Massively parallel algorithm and implementation of RI-MP2 energy calculation for peta-scale many-core supercomputers.

A new parallel algorithm and its implementation for the RI-MP2 energy calculation utilizing peta-flop-class many-core supercomputers are presented. Some improvements from the previous algorithm (J. Chem. Theory Comput. 2013, 9, 5373) have been performed: (1) a dual-level hierarchical parallelization scheme that enables the use of more than 10,000 Message Passing Interface (MPI) processes and (2) a new data communication scheme that reduces network communication overhead. A multi-node and multi-GPU implementation of the present algorithm is presented for calculations on a central processing unit (CPU)/graphics processing unit (GPU) hybrid supercomputer. Benchmark results of the new algorithm and its implementation using the K computer (CPU clustering system) and TSUBAME 2.5 (CPU/GPU hybrid system) demonstrate high efficiency. The peak performance of 3.1 PFLOPS is attained using 80,199 nodes of the K computer. The peak performance of the multi-node and multi-GPU implementation is 514 TFLOPS using 1349 nodes and 4047 GPUs of TSUBAME 2.5. © 2016 Wiley Periodicals, Inc.

Comparison of tunnel orientation between transtibial and anteromedial portal techniques for anatomic double-bundle anterior cruciate ligament reconstruction using 3-dimensional computed tomography.

PURPOSE: The aim of this study was to compare femoral and tibial tunnel placement, angle, and length between transtibial and anteromedial portal techniques for anatomic double-bundle anterior cruciate ligament (ACL) reconstruction. METHODS: Fifty patients were randomized to the 2 groups, and a femoral tunnel was created through the tibial tunnel (transtibial) and the far anteromedial portal (AMP) in 25 patients each. Both groups underwent anatomic double-bundle ACL reconstruction with hamstring tendons. Volume-rendering computed tomography (CT) was used to evaluate femoral and tibial tunnel placement, and transparent 3-dimensional CT image reconstruction, to evaluate tunnel angles, on the seventh postoperative day. Femoral tunnel length was measured intraoperatively. RESULTS: Anteromedial bundle (AMB) and posterolateral bundle (PLB) femoral tunnels were placed significantly lower and deeper with the AMP technique (shallow/deep direction: 21% and 30%, high/low direction: 18% and 48%) than with the transtibial technique (25% and 34%, 12% and 43%). Except for the tibial tunnel angle in the axial plane, AMB and PLB femoral and tibial tunnel angles differed significantly in 3 dimensions. AMB and PLB femoral tunnel lengths were significantly shorter with the AMP technique (AMB: 33 mm, PLB: 32 mm) than with the transtibial technique (AMB: 49 mm, PLB: 37 mm) (P < .001 and P = .001). Both femoral tunnel lengths in the AMP group correlated significantly with the tunnel angle in the sagittal (AMB: r = 0.69, PLB: r = 0.51) and axial (AMB: r = 0.58, PLB: r = 0.75) planes. CONCLUSIONS: AMB and PLB femoral tunnels were placed significantly deeper, lower, and closer to the femoral footprint reported in previous cadaveric studies in the anteromedial portal technique than in the transtibial technique. Femoral tunnel length was significantly shorter in the anteromedial portal group than in the transtibial group. LEVEL OF EVIDENCE: Level II, prospective comparative study.

[Effect of S-1 in a patient with metastatic eyelid sebaceous carcinoma].

We report a 68-year-old female with a history of repeated right cervical lymph node metastasis secondary to eyelid sebaceous carcinoma. She had undergone right neck dissection due to lymph node metastasis at the department of surgery in an other local hospital. After the orbital exenteration, she underwent multiple operations, including 2 right partial parotidectomy and then total parotidectomy, 4 right cervical lymph node excisions and 1 left upper neck dissection due to lymph node metastasis. Adjuvant chemotherapy with S-1(at a dose of 80 mg per day)alone was applied as tumor dormancy therapy on an outpatient basis. Any adverse events during S-1 medication were observed. New metastatic lesions were not found until 29 months after the beginning of chemotherapy with S-1 alone. Although the effectiveness of S-1 for eyelid sebaceous carcinoma has not been demonstrated, S-1 might be useful in patients with recurrent eyelid sebaceous carcinoma based on our patient's recurrence-free survival.

Correlation between canal paresis and spontaneous nystagmus during early stage of acute peripheral vestibular disorders.

CONCLUSIONS: This study demonstrates that the resolution period of spontaneous nystagmus (SN) may provide an indication of vestibular dysfunction on a particular day in the primary care setting. OBJECTIVE: We aimed to predict canal paresis using fundamental observations of SN during the early stage of acute peripheral vestibular disorders. METHODS: The study involved 87 patients who had recently experienced their first episode of acute spontaneous vertigo and direction-fixed horizontal nystagmus. Although they did not exhibit any other neurological deficits, they had been hospitalized with severe acute symptoms between 2004 and 2007. A correlation between the resolution period of SN and the results of laboratory caloric testing was reviewed. RESULTS: The receiver operating characteristic analysis showed that the resolution period of SN may be a predictive indicator of unilateral vestibular hypofunction in the acute stage. In about half of the patients, SN disappeared on the third day after their initial visit. However, in 20% of the patients SN still persisted on the eighth day. Among the patients with SN, the prevalence of canal paresis increased with the increase in the resolution period of SN. When SN was observed on the fifth day, the prevalence was approximately 70%.

In vivo magnetic resonance imaging measurement of tibiofemoral relation with different knee flexion angles after single- and double-bundle anterior cruciate ligament reconstructions.

PURPOSE: Double-bundle anterior cruciate ligament (ACL) reconstruction has been shown to restore better kinematics in vitro, but it is uncertain whether this technique can achieve this in vivo. We investigated whether anatomic double-bundle ACL reconstruction can restore a better tibiofemoral relation in the sagittal plane under static unloading conditions. METHODS: The tibiofemoral relation was assessed with an open magnetic resonance imaging scanner (0.5 T) in 15 patients with anatomic double-bundle reconstruction (double-bundle group) and 14 patients with single-bundle reconstruction (single-bundle group) by use of hamstring tendons. T1-weighted magnetic resonance imaging for both knees was obtained at 0 degrees, 45 degrees, 90 degrees, and 120 degrees of flexion without external force and muscle contraction 6 months after surgery. The position of the posterior femoral condyles relative to the tibia was measured in the midmedial and midlateral sagittal sections of the knee. Clinical evaluation (range of motion, KT-2000 measurement [MEDmetric, San Diego, CA], and pivot-shift test) was performed at the same time. RESULTS: KT-2000 testing showed that the mean side-to-side difference in the double-bundle group (0.7 mm) was significantly smaller than that in the single-bundle group (1.7 mm). In the double-bundle group, the tibiofemoral relation in operated knees was not significantly different from that in the contralateral knees at 0 degrees, 45 degrees, 90 degrees, and 120 degrees of flexion, although at 0 degrees of flexion, the femoral condyles were positioned anteriorly relative to the tibia compared with that in the contralateral knees. Similar results were obtained in the single-bundle group. The difference in the tibiofemoral relation between reconstructed and contralateral knees was not significantly different between the 2 groups. CONCLUSIONS: According to KT-2000 measurement, this study found that anteroposterior stability was better with anatomic double-bundle ACL reconstruction than with single-bundle ACL reconstruction. However, under static conditions without an anterior drawer force, anatomic double-bundle ACL reconstruction did not show superiority in terms of restoring a better tibiofemoral relation compared with single-bundle ACL reconstruction. LEVEL OF EVIDENCE: Level III, comparative study.