Disentangling neural mechanisms for perceptual grouping
Authors: Junkyung Kim*, Drew Linsley*, Kalpit Thakkar, Thomas Serre
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We address this question by systematically evaluating neural network architectures featuring combinations bottom-up, horizontal, and top-down connections on two synthetic visual tasks... We compare model predictions to human psychophysics data on the same visual tasks and show that human judgements are significantly more consistent with image predictions from our TD+H-CNN than Res Net and U-Net models. |
| Researcher Affiliation | Collaboration | Department of Cognitive, Linguistic and Psychological Sciences Brown University Providence, RI 02912, USA junkyung@google.com {drew_linsley,kalpit_thakkar,thomas_serre}@brown.edu ... Deep Mind, London, UK. |
| Pseudocode | No | The paper describes the fGRU module with mathematical formulas and architectural details, but it does not include a clearly labeled 'Pseudocode' or 'Algorithm' block, nor does it present the procedure in a structured, code-like format. |
| Open Source Code | No | The paper does not contain an explicit statement about open-sourcing the code for the described methodology, nor does it provide a link to a code repository. |
| Open Datasets | No | The paper introduces two synthetic visual tasks, the Pathfinder challenge and the cluttered ABC challenge (c ABC), and describes their generation process in detail in Appendix A. However, it does not provide any concrete access information (e.g., URL, DOI, or a specific citation to a public repository) for these datasets. |
| Dataset Splits | Yes | For Pathfinder this was 60K images, and for c ABC this was 45K, and we used these image budgets for all difficulty levels of the respective challenges (with 10% held out for evaluation). |
| Hardware Specification | Yes | All experiments were run with NVIDIA Titan X GPUs. |
| Software Dependencies | No | The paper mentions 'Adam optimizer' and that human experiments were 'implemented using js Psych and custom javascript functions', but it does not provide specific version numbers for these or any other key software components, libraries, or frameworks used for the experiments. |
| Experiment Setup | Yes | Models were trained with batches of 32 images and the Adam optimizer. We trained each model with five random weight initializations and searched over learning rates [1e 3, 1e 4, 1e 5, 1e 6]. ... We also initialized the learnable scale parameter δ of f GRU normalizations to 0.1... Similarly, f GRU parameters for learning additive suppression/facilitation (µ, κ) were initialized to 0, and parameters for learning multiplicative suppression/facilitation (α, ω) were initialized to 0.1. |