DISSECT: Disentangled Simultaneous Explanations via Concept Traversals
Authors: Asma Ghandeharioun, Been Kim, Chun-Liang Li, Brendan Jou, Brian Eoff, Rosalind Picard
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate DISSECT on several challenging synthetic and realistic datasets where previous methods fall short of satisfying desirable criteria for interpretability and show that it performs consistently well. Finally, we present experiments showing applications of DISSECT for detecting potential biases of a classifier and identifying spurious artifacts that impact predictions. |
| Researcher Affiliation | Collaboration | Asma Ghandeharioun Google Research / MIT aghandeharioun@google.com Been Kim Google Research beenkim@google.com Chun-Liang Li, Brendan Jou, Brian Eoff Google Research {chunliang,bjou,beoff}@google.com Rosalind W. Picard MIT picard@media.mit.edu |
| Pseudocode | No | The paper describes the model and optimization process but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code for all the models and metrics is publicly available at https://github.com/ asmadotgh/dissect. |
| Open Datasets | Yes | 3D Shapes. We first use 3D Shapes [7] purely for validation and demonstration purposes due to the controllability of all its factors. It is a synthetic dataset available under Apache License 2.0 composed of 480K 3D shapes procedurally generated from 6 ground-truth factors of variation. ... Additionally, we released the new dataset, which is publicly available at https://affect.media.mit.edu/dissect/synthderm. ... Celeb A. We also include the Celeb A dataset [44] that is available for non-commercial research purposes. |
| Dataset Splits | Yes | For evaluation, we used a hold-out set including 10K samples. For post hoc evaluation classifiers predicting Distinctness and Realism, 75% of the samples were used for training, and the results were reported on the remaining 25%. ... See Table 5 for the summary of the hyper-parameter values. (Table 5 shows 'hold-out test ratio 0.25') |
| Hardware Specification | Yes | Training and evaluation of all models across the three datasets have approximately taken 1000 hours on a combination of Nvidia GPUs including GTX TITAN X, GTX 1080 Ti, RTX 2080 Rev, and Quadro K5200. |
| Software Dependencies | No | The paper mentions software components like 'Adam optimizer' and 'Projection GAN' but does not specify version numbers for major libraries or frameworks like Python, PyTorch, TensorFlow, or CUDA. |
| Experiment Setup | Yes | See Table 5 for the summary of the hyper-parameter values. (Table 5 includes N, λc GAN, λrec, λf, D steps, G steps, batch size, epochs, K, λr) |