Deep neural networks such as GoogLeNet, ResNet, and BERT have achieved impressive performance in tasks such as image and text classification. To understand how such performance is achieved, we probe a trained deep neural network by studying neuron activations, i.e.combinations of neuron firings, at various layers of the network in response to a particular input. With a large number of inputs, we aim to obtain a global view of what neurons detect by studying their activations. In particular, we develop visualizations that show the shape of the activation space, the organizational principle behind neuron activations, and the relationships of these activations within a layer. Applying tools from topological data analysis, we present TopoAct, a visual exploration system to study topological summaries of activation vectors. We present exploration scenarios using TopoAct that provide valuable insights into learned representations of neural networks. We expect TopoAct to give a topological perspective that enriches the current toolbox of neural network analysis, and to provide a basis for network architecture diagnosis and data anomaly detection.

No abstract

Background: To evaluate the clinical value of foetal intelligent navigation echocardiography (5D Heart) for the display of key diagnostic elements in basic sections. Methods: 3D volume datasets of 182 normal singleton foetuses were acquired with a four chamber view by using a volume probe. After processing the datasets by using 5D Heart, eight cardiac diagnostic planes were demonstrated, and the image qualities of the key diagnostic elements were graded by 3 doctors with different experiences in performing foetal echocardiography. Results: A total of 231 volume datasets acquired from the 182 normal foetuses were used for 5D Heart analysis and display. The success rate of 8 standard diagnostic views was 88.2%, and the success rate of each diagnostic view was 55.8-99.2% and 70.7-99.0% for the random four chamber view as the initial section and for the apical four chamber view as the initial section, respectively. The success rate of each diagnostic element in the 8 diagnostic sections obtained by 5D Heart was 58.9%~100%. Excellent agreement was found between experienced sonographers and less-experienced sonographers (kappa> 0.769). Inter-and intra-observer agreement were substantial to near-perfect, kappa values ranging from 0.612 to 1.000 (Cohen's kappa). Conclusions: 5D Heart can significantly improve the image quality of key diagnostic elements in foetal echocardiography with low operator dependency and good reproducibility.

No abstract

Variational quantum circuits are one of the promising ways to exploit the advantages of quantum computing in the noisy intermediate-scale quantum technology era. The design of the quantum circuit architecture might greatly affect the performance capability of the quantum algorithms. The quantum architecture search is the process of automatically designing quantum circuit architecture, aiming at finding the optimal quantum circuit composition architecture by the algorithm for a given task, so that the algorithm can learn to design the circuit architecture. Compared to manual design, quantum architecture search algorithms are more effective in finding quantum circuits with better performance capabilities. In this paper, based on the deep reinforcement learning, we propose an approach for quantum circuit architecture search. The sampling of the circuit architecture is learnt through reinforcement learning based controller. Layer-based search is also used to accelerate the computational efficiency of the search algorithm. Applying to data classification tasks we show that the method can search for quantum circuit architectures with better accuracies. Moreover, the circuit has a smaller number of quantum gates and parameters.