Neural Systematic Binder
Authors: Gautam Singh, Yeongbin Kim, Sungjin Ahn
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In experiments, we find that Sys Binder provides significantly better factor disentanglement within the slots than the conventional object-centric methods, including, for the first time, in visually complex scene images such as CLEVR-Tex. |
| Researcher Affiliation | Academia | Gautam Singh1 , Yeongbin Kim2 & Sungjin Ahn2 1Rutgers University 2KAIST |
| Pseudocode | Yes | A PSEUDO-CODE Algorithm 1 Neural Systematic Binder. |
| Open Source Code | Yes | We release all the resources used in this work, including the code and the datasets, at the project link: https://sites.google.com/view/neural-systematic-binder. |
| Open Datasets | Yes | Datasets. We evaluate our model on three datasets: CLEVR-Easy, CLEVR-Hard, and CLEVR-Tex. These are variants of the original CLEVR dataset (Johnson et al., 2017) ... We release our code and the datasets here1. 1https://sites.google.com/view/neural-systematic-binder |
| Dataset Splits | No | The paper mentions training steps and parameters but does not explicitly state the dataset splits (e.g., percentages or counts for training, validation, and test sets). |
| Hardware Specification | No | The paper provides a table of 'Memory Consumption' and 'Time per Training Iteration' but does not specify the hardware (e.g., specific GPU or CPU models) on which these measurements were taken. |
| Software Dependencies | No | The paper mentions software components like GRU, MLP, and implicitly Python/PyTorch through its described architecture and cited works (e.g., SLATE), but it does not specify exact version numbers for any software dependencies. |
| Experiment Setup | Yes | Table 2: Hyperparameters of our model used in our experiments. This table details General, Sys Binder, Transformer Decoder, and d VAE hyperparameters including Batch Size (40), Training Steps (200K, 400K), Block Size (256, 128), # Blocks (8, 16), # Prototypes (64), # Iterations (3), # Slots (4, 6), Learning Rate (0.0001, 0.0003), # Decoder Blocks (8), # Decoder Heads (4, 8), Hidden Size (192), Dropout (0.1), Patch Size (4x4 pixels), Vocabulary Size (4096), Temperature Start (1.0), Temperature End (0.1), and Temperature Decay Steps (30000). |