To Transfer or Not to Transfer: Suppressing Concepts from Source Representations
Authors: Vijay Sadashivaiah, Keerthiram Murugesan, Ronny Luss, Pin-Yu Chen, Chris Sims, James Hendler, Amit Dhurandhar
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our approach qualitatively and quantitatively in the visual domain showcasing its efficacy for classification and generative source models. |
| Researcher Affiliation | Collaboration | Vijay Sadashivaiah EMAIL Department of Computer Science, Rensselaer Polytechnic Institute Keerthiram Murugesan EMAIL IBM Research, Yorktown Heights |
| Pseudocode | Yes | Algorithm 1 Controllable Concept Transfer Concatenate with Concept Search method (CCT-cat(cs)) Algorithm 2 TRAIN-CDN: Train Concept Disentangling Network Algorithm 3 CONCEPT-SEARCH: Search Concepts to Suppress |
| Open Source Code | No | The paper does not provide an explicit statement about the release of source code for the methodology, nor does it include a link to a code repository. |
| Open Datasets | Yes | For the bulk of our experiments in Sections 4 and 5 we use the MNIST (Le Cun et al., 1998), EMNIST (Cohen et al., 2017) and Celeb Faces Attributes (Celeb A) (Liu et al., 2015) datasets. ... Specifically, it has been shown that object recognition models can spuriously rely on the image background instead of the objects themselves (Ribeiro et al., 2016). We study this phenomenon using the Water Birds dataset (Sagawa et al., 2019) |
| Dataset Splits | Yes | EMNIST. EMNIST is a set of handwritten characters derived from NIST Special Database 19. ... There are 88,800 training examples and 14,800 testing examples. ... Celeb Faces Attributes. ... There are 162,770 training examples, 19962 test examples and 19867 validation examples. ... The Diabetes Health Indicators Dataset contains healthcare statistics and lifestyle survey information about people along with their diagnosis of diabetes. ... We first divide the dataset into source and target tasks, where the source task has 69,692 samples leaving the target task with 1000 samples. |
| Hardware Specification | Yes | The models were trained in parallel with the specifications shown in Table 8. Resource Setting CPU IBM Power 9 CPU @ 3.15GHz Memory 512GB GPUs 1 x NVIDIA Tesla V100 16 GB Disk 1.2 TB OS Red Hat8 |
| Software Dependencies | No | The paper mentions the use of the Adam optimizer (Kingma & Ba, 2014) but does not provide specific version numbers for any software libraries or dependencies used in the implementation. |
| Experiment Setup | Yes | For our experimental analysis in the main paper, we set the number of epochs for training to E = 50 for all models. We train all models using a batch size of 25 and a learning rate of 10 4 for the Adam optimizer (Kingma & Ba, 2014). All models were randomly initialized before training. |