-
It has been brought to our attention that Metamask uses `m/44'/60'/x'` paths even for testnet version of the networks (e.g. Rinkeby) - https://github.com/trezor/trezor-firmware/issues/1474
For test…
-
Hello Professor:
For an experiment I tried to make changes to the c file, do I need to recompile him to a pyd file every time I change the c file? Can you please tell me how this step is achieved.
-
Description:
The `claim` function in the ClaimContract is vulnerable to signature replay attacks. The function uses a `_postfix` parameter in the signature verification process, but there's no mechan…
-
## 一言でいうと
予測の不確実性を推定する手法の提案。既存研究ではDropoutを使った不確実推定があったが、Dropoutよりもよく使われているBatch Normalizationを使い推定する。正規化に用いる平均/分散を学習データからサンプリングすることでモデル出力の分布を得る。
![image](https://user-images.githubusercontent.co…
-
## 一言でいうと
精度/実行速度のトレードオフを上手くチューニングする手法と、チューニング元のネットワーク(EfficientNet)の提案(これはAutoMLで構築)。ネットワークの幅・深さ・入力画像の解像度(H x M)が調整対象で、これらの相関関係から係数を算出することで単一パラメーターによる同時調整を可能にしている
![image](https://user-images.g…
-
The current documentation for nw:betweenness-centrality states that "As of now, link weights are not taken into account."
If support for weighted networks isn't too onerous to implement it would be p…
-
https://arxiv.org/pdf/1607.03516.pdf
In this paper, we propose a novel unsupervised domain adaptation algorithm based on deep learning for visual object recognition. Specifically, we design a new m…
leo-p updated
7 years ago
-
## 一言でいうと
入力画像に細工を施すことで、誤識別を誘発できるという検証。ネットワークの構成がわかっているホワイトボックス型と不明なブラックボックス型があるが、本研究では後者の手法で検証。ランダムに選択された画像の各所から最も識別に影響を与える箇所を探索する形(greedy local-search)で実装を行い、1~3%程度のピクセルを操作するだけでエラーレートを跳ね上げられることを確…
-
## 一言でいうと
数ピクセルの編集で画像分類のモデルをだますという研究。1ピクセルでも70%程度の確率で誤識別させることができ、さらに20%程度で特定のクラスに識別させることも可能(5ピクセルならもっと上がる)。元の画像に対するピクセルの変動量は、GE(差分進化法)により最小の値を求めている。
![image](https://user-images.githubuserconten…
-
## 一言でいうと
DNNの判断を理解するための研究のまとめ。ネットワークが反応する入力を見つける方法(Activation Maximizationなど)、判断根拠となった特徴を入力にマップする方法(Relevance Propagationなど)などを紹介、説明力の比較方法についても記載している
### 論文リンク
https://arxiv.org/abs/1706.079…