Support-Set Context Matters for Bongard Problems
Authors: Nikhil Raghuraman, Adam W Harley, Leonidas Guibas
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our Transformer-based approach sets a new state-of-the-art on Bongard-LOGO (75.3%) and Bongard-HOI (76.4%) compared to approaches with equivalent backbone architectures, and also performs fairly well on the original Bongard problem set (60.8%). |
| Researcher Affiliation | Academia | Nikhil Raghuraman EMAIL Department of Computer Science Stanford University Adam W. Harley EMAIL Department of Computer Science Stanford University Leonidas Guibas EMAIL Department of Computer Science Stanford University |
| Pseudocode | No | The paper only describes steps in regular paragraph text without structured formatting or dedicated pseudocode blocks. |
| Open Source Code | Yes | Code is available at https://github.com/nraghuraman/bongard-context. |
| Open Datasets | Yes | We evaluate on the Bongard-LOGO dataset (Nie et al., 2021)... Bongard-HOI (Jiang et al., 2023)... Mikhail Bongard s original dataset (Foundalis, 2006; Yun et al., 2020)... We publicly release this cleaned training dataset to aid future works. |
| Dataset Splits | Yes | For Bongard-LOGO... The train, validation, and test sets consist of 9300, 900, and 1800 problems, respectively. The test set is further subdivided into four splits... For Bongard-HOI... The dataset authors defined train, validation, and test sets consisting of 23041, 17187, and 13941 problems, respectively. We arbitrarily select the final positive and negative images as queries... For Bongard-Classic... we select one positive and one negative example as queries. |
| Hardware Specification | Yes | We ran all experiments on NVIDIA Titan RTX, Titan XP, Ge Force, or A5000 GPUs. Each experiment used at most 1 GPU and 30GB of GPU memory. |
| Software Dependencies | No | The paper mentions various models and tools used (e.g., CLIP, ResNet, Transformer, GPT-4o, DINO) and references the Open AI API, but does not provide specific version numbers for any key software libraries or dependencies like PyTorch, TensorFlow, or CUDA. |
| Experiment Setup | Yes | To train our models, we used the Adam W optimizer (Loshchilov & Hutter, 2017), a maximum learning rate of 5e 5, and a 1-cycle learning rate policy (Smith & Topin, 2019)... We train models for 500,000 iterations with a batch size of 2. We resize images to 512 512 pixels and apply random cropping and horizontal flipping to augment... We train for 10,000 iterations with a batch size of 16 and weight decay of 0.01. We use augmentations at training time, including horizontal flips, random grayscale, color jitter, and random rescaling and cropping (to 224 224), and apply support dropout and label noise. |