On InstaHide, Phase Retrieval, and Sparse Matrix Factorization

Authors: Sitan Chen, Xiaoxiao Li, Zhao Song, Danyang Zhuo

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We describe an experiment demonstrating the utility of Insta Hide for Gaussian images and comparing to the utility of another data augmentation scheme, Mix Up Zhang et al. (2018). We also informally report on our implementation of LEARNPUBLIC and its empirical efficacy.
Researcher Affiliation Academia Sitan Chen MIT sitanc@mit.edu. Xiaoxiao Li Princeton Universtiy xiaoxiao.li@aya.yale.edu Zhao Song Columbia University, Princeton University / IAS magic.linuxkde@gmail.com. Danyang Zhuo Duke University danyang@cs.duke.edu.
Pseudocode Yes Algorithm 1: LEARNPUBLIC({([pj]S, yj)}j [d]), Algorithm 2: GRAMEXTRACT({y X,wi}i [m], η), Algorithm 3: FINDFLORALSUBMATRIX(M, k, r), Algorithm 4: LEARNPRIVATEIMAGE({y X,wi}i [m])
Open Source Code No The paper mentions 'Our implementation of LEARNPUBLIC and its empirical efficacy' but does not provide any link or explicit statement about the code being open-source.
Open Datasets No We generated random images x1, . . . , x1000 R10 from N(0, Id)...
Dataset Splits Yes In all of our experiments, we separate our original image data (before generating synthetic data) into two categories: 80% training data and 20% test data.
Hardware Specification No The paper does not provide any specific hardware details such as CPU/GPU models or memory for the experimental setup.
Software Dependencies No The paper mentions software components like 'Res Net-18', 'NASNet', 'SGD optimizer', 'Mix Up', but does not provide specific version numbers for any of them.
Experiment Setup Yes We use a 4-layer neural network as a binary classifier, y = arg max(softmax(W4σ(W3σ(W2(σ(W1x + b1)) + b2) + b3) + b4)), where x R10, W1 R100 10, W2 R100 100, W3 R100 100, W4 R2 100, b1 R100, b2 R100, b3 R100, b4 R100. We initialize the entries of each Wl and bl to be i.i.d. draws from N(ul, 1), where ul is sampled from N(0, α) at the outset. We train the neural network for 100 epochs with cross-entropy loss and SGD optimizer with a learning rate of 0.01.