논문 표현 정리

• 인트로

  • 관련 분야 소개
    • In order to better understand classifier decisions and to gain insight into how these mod-els operate, a variety techniques have been proposed
  • 중요성 설정
    • A major current focus in population management is how to ensure sustainability of …
    • Deep learning has become a dominant machine learning tool in visual recognition and image processing

• 방법

  • 재료/방법에 대한 일반적 소개나 전반적 설명을 제공하거나 사용된 재료/장비의 출처를 밝혀라

  • 핵심적 배경지식을 제공하라

    • Recall from the introduction and Figure 2 that semantically similar inputs tend to elicit similar activation …
    • Activations are also used to define the losses in FitNets, … , and attention transfer. However the key difference is …

  • 재료/방법에 대한 구체적이고 상세한 정보를 제공하라

    • The details of the problem formulation are described in the following subsections.
    • The encoder is composed of four convolution layers. The resultant activation maps are then passed to … the activation map size is then expanded again .
    • We conducted experiments using wide residual networks (WideResNets) following …
    • We adopted the standard protocol for training wide residual networks on CIFAR-10 (SGD with Nesterov momentum; 200 epochs;BS of 128; and an initial LR of 0.1, decayed by a factor …)
    • ... Image sizes range from 300x300 to 640x640 ...

  • 당신의 선택을 정당화 시켜라

  • 적절한 주의력을 기울였음을 강조하라

  • 다른 연구의 재료/방법과 연관시켜 언급하라

  • 어디에서 문제점이 발생하였는지 설명하여라

• 결과

  • 결과에 대한 가능성 있는 제안
    • This validates our intuition that the activation similarities across images encode useful semantics learned by the teacher network
    • These curves indicate that the effective breadth is a minimum at the point of application of the load.
    • it is also possible to view the results from the perspective of the teacher network.
    • Figure 5 illustrates how the ...
    • We observed robust performance over  ...
    •  

• 결론

• 기타

  • 수식에 대한 의미 설명
    • On a linear model, the optimal attribution would be obtained by element-wise multiplying the signal with the weight vector: input = w (elementwise multiply) y
    • These material properties for the lining were chosen as follows: Eh = 11.5 kN mm−2, E l = 5.5 kN mm−2, vh = 0.29 and vl = 0.14, where, in order to be consistent with Betti’s condition, vl∙Eh = vh∙El, it was decided to adjust the value of Poisson’s ratio by taking the lower range of v l (v l being considered, in general, a more reliable test value than vh) at 0.14, and then working out vh at 0.29, a sensible approximation, as argued elsewhere [8].

참고문헌

Science Research Writing for Non-Native Speakers of English