Measuring the Interpretability of Unsupervised Representations via Quantized Reversed Probing

Authors: Iro Laina, Yuki M Asano, Andrea Vedaldi

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We use our method to evaluate a large number of self-supervised representations, ranking them by interpretability, highlight the differences that emerge compared to the standard evaluation with linear probes and discuss several qualitative insights. We use our method to evaluate a wide range of recent self-supervised representation learning and clustering techniques.
Researcher Affiliation Academia Iro Laina University of Oxford iro.laina@eng.ox.ac.uk Yuki M. Asano University of Amsterdam y.m.asano@uva.nl Andrea Vedaldi University of Oxford vedaldi@robots.ox.ac.uk
Pseudocode No The paper describes the method using text and diagrams (Figure 1), but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Code at: https://github.com/iro-cp/ssl-qrp.
Open Datasets Yes in our experiments we focus on Image Net (IN-1k) (Deng et al., 2009) as the target data due to the large availability of pre-trained self-supervised models on this dataset. We apply the model to our target datasets and keep only object categories predicted with a confidence higher than 0.5. We use a Deep Lab-v2 model (Chen et al., 2017) trained for segmentation on MS COCO 2017 (Lin et al., 2014; Caesar et al., 2018).
Dataset Splits Yes We divide the data into train and test sets by splitting all cluster assignments with a 80/20 ratio and stratified sampling; from the training split we also reserve 20% of the data for validation.
Hardware Specification Yes Our method computes cluster assignments using the efficient K-means implementation of faiss (which takes less than 5min for 256k 2048-d vectors on 4 NVIDIA RTX A4000). Training of the linear model converges in less than 100 epochs in a matter of minutes on a single GPU (1 epoch takes 2sec).
Software Dependencies No The paper mentions using 'faiss' for K-means but does not specify its version number or any other software dependencies with explicit version details.
Experiment Setup Yes We train for up to 100 epochs with batch size 512 and optimize using SGD with a momentum of 0.9 and initial learning rate of 3.5 which is further reduced by a factor of 10 at epochs 60 and 80. We also add L2-regularization with weight 3 10 6.