Score based generative models are a new class of generative models that have been shown to accurately generate high dimensional calorimeter datasets. Recent advances in generative models have used images with 3D voxels to represent and model complex calorimeter showers. Point clouds, however, are likely a more natural representation of calorimeter showers, particularly in calorimeters with high granularity. Point clouds preserve all of the information of the original simulation, more naturally deal with sparse datasets, and can be implemented with more compact models and data files. In this work, two state-of-the-art score based models are trained on the same set of calorimeter simulation and directly compared.
A prototype bi-directional neural interface system is presented with closed-loop and embedded DSP capabilities. The system includes 32-electrode stimulation capability, eight multiplexed low-noise low-power bio-potential sensing channels with on-chip digital FFT, and a Cortex M3-based microcontroller for implementing closed-loop algorithms. The stimulation subsystem can provide a maximum stimulation current of 12mA with approximately 3% accuracy per channel. The sensing subsystem uses a fully differential chopping amplifier achieving a noise floor <;100nV/rtHz with approximately 1μA current from 2V supply. A 13bit continuous-time sigma-delta ADC is used to sample the amplifier output at the rate of 33kHz. The ADC consumes approximately 100nA while achieving an ENOB of 12b with a 250Hz date rate. A hardware implementation of a cached-FFT is included in the sensing IC to perform spectral analysis of the bio-potential signal. The resulting multiple spectral signatures can be selectively sent to the microcontroller allowing for algorithmic control of the stimulator system. The system partition was designed to minimize overall computational power while providing user flexibility. The sensing performance of the prototype has been demonstrated with 2-D cursor control in a non-human primate brain machine interface (BMI) model using 20 features simultaneously.
We develop a method for emotion recognition from facial imagery. This problem is challenging in part because of the subjectivity of ground truth labels and in part because of the relatively small size of existing labeled datasets. We use the FER+ dataset [8], a dataset with multiple emotion labels per image, in order to build an emotion recognition model that encompasses a full range of emotions. Since the amount of data in the FER+ dataset is limited, we explore the use of a much larger face dataset, MS-Celeb-1M [41], in conjunction with the FER+ dataset. Specific layers within an Inception-ResNet-v1 [13, 38] model trained for facial recognition are used for the emotion recognition problem. Thus, we leverage the MS-Celeb-1M dataset in addition to the FER+ dataset and experiment with different architectures to assess the overall performance of neural networks to recognize emotion using facial imagery.