Combining Observational Data and Language for Species Range Estimation
Authors: Max Hamilton, Christian Lange, Elijah Cole, Alexander Shepard, Samuel Heinrich, Oisin Mac Aodha, Grant Van Horn, Subhransu Maji
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluated on held-out species, our zero-shot SRMs significantly outperform baselines and match the performance of SRMs obtained using tens of observations. Our approach also acts as a strong prior when combined with observational data, resulting in more accurate range estimation with less data. We present extensive quantitative and qualitative analyses of the learned representations in the context of range estimation and other spatial tasks, demonstrating the effectiveness of our approach. |
| Researcher Affiliation | Collaboration | Max Hamilton1 Christian Lange2 Elijah Cole3 Alexander Shepard4 Samuel Heinrich5 Oisin Mac Aodha2 Grant Van Horn1 Subhransu Maji1 1UMass Amherst 2University of Edinburgh 3Altos Labs 4i Naturalist 5Cornell University |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code and data is publicly available at: https://github.com/cvl-umass/le-sinr |
| Open Datasets | Yes | We use the dataset proposed in SINR [10], consisting of 35.5 million observations covering 47,375 species observed prior to 2022 on the i Naturalist platform. ... We source our text descriptions from Wikipedia [4]. |
| Dataset Splits | No | The paper clearly defines the training and test (evaluation) sets by species count, but does not explicitly define a separate validation dataset split with specific percentages or counts. It mentions training with presence observations and sampling negatives during training, and evaluation on held-out species. |
| Hardware Specification | Yes | Training a single LE-SINR model from scratch using all the text and observational data takes about 10 hours on a single NVIDIA RTX 2080ti GPU occupying about 10GB of VRAM. |
| Software Dependencies | No | The paper mentions using specific models like Grit LM and Llama-3, and refers to the Adam optimizer and ReLU activation, but does not specify version numbers for any software dependencies like PyTorch, Python, or CUDA. |
| Experiment Setup | Yes | We train with the Adam optimizer for 10 epochs with a learning rate of 0.0005. The species network has three linear layers with Re LU activation. The input text embedding dimension is 4,096, the hidden dimension is 512, and the output species embedding dimension is 256. ... For logistic regression in the later few-shot evaluation experiments, we use a regularization strength λ = 20. |