{
  "format_version": 1,
  "repository_type": "fitted_nsd_encoding_and_variance_partitioning_models",
  "artifact_name": "VEDB and Reference SimCLR ResNet-18 — NSD Encoding and Variance-Partitioning Models",
  "framework": {
    "feature_extractor": "pytorch",
    "encoding_model_fitting": "pytorch",
    "artifact_serialization": "numpy"
  },
  "paper": {
    "title": "Eccentricity-Constrained CNN Training Reveals Adaptive Information Coding Around the Visual Field",
    "authors": [
      "Dylan M. Diaz",
      "Margaret M. Henderson"
    ],
    "year": 2026,
    "venue": "Proceedings of the 9th Conference on Cognitive Computational Neuroscience",
    "doi": "10.32470/0416gfsq",
    "arxiv": "2607.19316"
  },
  "upstream_models": {
    "architecture": "resnet18",
    "pretraining_objective": "simclr",
    "vedb_models": [
      {
        "name": "Baseline",
        "pretraining_dataset": "Visual Experience Dataset (VEDB)",
        "repository": "DM-Diaz/VEDB-SimCLR-ResNet18-Baseline",
        "encoding_analysis_identifier": "resnet18-Baseline"
      },
      {
        "name": "Fovea-Gaze",
        "pretraining_dataset": "Visual Experience Dataset (VEDB)",
        "repository": "DM-Diaz/VEDB-SimCLR-ResNet18-Fovea-Gaze",
        "encoding_analysis_identifier": "resnet18-FoveaGaze"
      },
      {
        "name": "Periph",
        "pretraining_dataset": "Visual Experience Dataset (VEDB)",
        "repository": "DM-Diaz/VEDB-SimCLR-ResNet18-Periph",
        "encoding_analysis_identifier": "resnet18-PeriphNonTTM"
      },
      {
        "name": "Periph-NF",
        "pretraining_dataset": "Visual Experience Dataset (VEDB)",
        "repository": "DM-Diaz/VEDB-SimCLR-ResNet18-Periph-NF",
        "encoding_analysis_identifier": "resnet18-PeriphTTM"
      }
    ],
    "reference_models": [
      {
        "name": "STL-10",
        "pretraining_dataset": "STL-10",
        "source": "Spijkervet/SimCLR",
        "source_url": "https://github.com/Spijkervet/SimCLR",
        "externally_provided_checkpoint": true,
        "redistributed_by_project": false,
        "encoding_analysis_identifier": "resnet18-pretrained-simclr"
      },
      {
        "name": "ImageNet-100",
        "pretraining_dataset": "ImageNet-100",
        "dataset_source": "clane9/imagenet-100",
        "repository": "DM-Diaz/SimCLR-ResNet18-ImageNet100",
        "encoding_analysis_identifier": "resnet18-simclr-imgnet100"
      },
      {
        "name": "ImageNet-1K",
        "pretraining_dataset": "ImageNet-1K",
        "dataset_source": "evanarlian/imagenet_1k_resized_256",
        "repository": "DM-Diaz/SimCLR-ResNet18-ImageNet1K",
        "encoding_analysis_identifier": "resnet18-simclr-imgnet1k"
      }
    ],
    "vedb_collection": "DM-Diaz/eccentricity-constrained-simclr-models-vedb"
  },
  "neural_dataset": {
    "name": "Natural Scenes Dataset (NSD)",
    "modality": "7T fMRI",
    "subjects": [
      "S1",
      "S2",
      "S3",
      "S4",
      "S5",
      "S6",
      "S7",
      "S8"
    ],
    "held_out_evaluation_images": 1000,
    "held_out_images_description": "NSD images shared across participants",
    "raw_data_redistributed": false
  },
  "input_preprocessing": {
    "input_resolution": [
      224,
      224
    ],
    "nsd_visual_field_transform_reapplied": false,
    "rescale": "uint8 / 255.0",
    "normalization": {
      "name": "ImageNet",
      "mean": [
        0.485,
        0.456,
        0.406
      ],
      "std": [
        0.229,
        0.224,
        0.225
      ]
    }
  },
  "feature_extraction": {
    "layers": [
      "conv1",
      "layer1.1",
      "layer2.1",
      "layer3.1",
      "layer4.1",
      "avgpool"
    ],
    "convolutional_spatial_reduction": {
      "method": "adaptive_average_pooling",
      "target_pre_pca_features_per_layer": 5000
    },
    "pca": {
      "components_per_layer": 200,
      "fit_separately_by_subject": true,
      "fit_separately_by_model_condition": true,
      "fit_separately_by_layer": true,
      "fit_scope": "full subject-specific feature matrix before encoding-model train/holdout partitioning"
    },
    "layer_features_concatenated": true
  },
  "encoding_model": {
    "type": "voxelwise_ridge_regression",
    "regularization": "L2",
    "candidate_lambda_count": 20,
    "lambda_selection": "nested_holdout",
    "selection_scope": "independently_per_voxel",
    "feature_normalization": {
      "method": "z_score",
      "statistics_fit_on": "training_plus_nested_holdout",
      "final_held_out_evaluation_excluded": true
    },
    "intercept": {
      "included": true,
      "implementation": "column_of_ones_appended_to_feature_matrix",
      "saved_weight_location": "final_row_of_weights"
    },
    "evaluation_metrics": [
      "r2",
      "corr"
    ]
  },
  "encoding_model_release": {
    "model_count": 7,
    "subjects_per_model": 8,
    "vedb_model_count": 4,
    "reference_model_count": 3,
    "vedb_encoding_fit_count": 32,
    "reference_encoding_fit_count": 24,
    "total_encoding_fit_count": 56,
    "variance_partitioning_fit_count": 16,
    "total_npy_artifact_count": 72
  },
  "variance_partitioning": {
    "method": "voxelwise_encoding_model_variance_partitioning",
    "reported_comparisons": [
      {
        "model1": "Fovea-Gaze",
        "model2": "Periph",
        "paper_figure": "Figure 4C",
        "repository_path": "variance-partitioning/fovea-gaze-vs-periph"
      },
      {
        "model1": "Periph",
        "model2": "Periph-NF",
        "paper_figure": "Figure 4D",
        "repository_path": "variance-partitioning/periph-vs-periph-nf"
      }
    ],
    "subjects_per_comparison": 8,
    "released_artifact_count": 16,
    "fits_per_comparison": [
      "model1_only",
      "model2_only",
      "combined"
    ],
    "combined_feature_space": "concatenation_of_model1_and_model2_feature_spaces",
    "feature_dimensions": {
      "single_model_without_intercept": 1200,
      "single_model_with_intercept": 1201,
      "combined_without_intercept": 2400,
      "combined_with_intercept": 2401
    },
    "regularization": {
      "type": "L2_ridge_regression",
      "candidate_lambda_count": 20,
      "lambda_selection": "nested_holdout",
      "selection_scope": "independently_per_voxel"
    },
    "evaluation_metrics": [
      "r2",
      "corr"
    ],
    "purpose": "estimate variance uniquely and jointly explained by paired representation spaces"
  },
  "encoding_model_artifact": {
    "file_format": ".npy",
    "serialization": "numpy_saved_python_dictionary",
    "load_with_allow_pickle": true,
    "fit_fields": [
      "subject",
      "model",
      "features_file_list",
      "lambdas",
      "voxel_mask",
      "voxel_index",
      "voxel_nc",
      "brain_nii_shape",
      "weights",
      "r2",
      "corr",
      "best_lambda_inds"
    ],
    "contains_fitted_voxelwise_weights": true,
    "contains_held_out_metrics": true,
    "contains_raw_nsd_stimuli": false,
    "contains_raw_fmri_data": false,
    "contains_fitted_pca_transforms": false,
    "contains_feature_normalization_statistics": false,
    "turnkey_new_image_to_voxel_prediction": false
  },
  "variance_partitioning_artifact": {
    "file_format": ".npy",
    "serialization": "numpy_saved_python_dictionary",
    "load_with_allow_pickle": true,
    "fit_fields": [
      "subject",
      "model1",
      "model2",
      "features_file_list1",
      "features_file_list2",
      "lambdas",
      "voxel_mask",
      "voxel_index",
      "voxel_nc",
      "brain_nii_shape",
      "weights_varpart",
      "r2_varpart",
      "corr_varpart",
      "best_lambda_inds_varpart"
    ],
    "weights_varpart_entries": [
      "model1-only",
      "model2-only",
      "combined"
    ],
    "contains_fitted_voxelwise_weights": true,
    "contains_held_out_metrics": true,
    "contains_raw_nsd_stimuli": false,
    "contains_raw_fmri_data": false,
    "contains_fitted_pca_transforms": false,
    "contains_feature_normalization_statistics": false,
    "turnkey_new_image_to_voxel_prediction": false
  },
  "artifact_scope_note": "The repository contains 56 subject-specific fitted voxelwise encoding models derived from seven pretrained visual models (four VEDB-pretrained models and three non-egocentric reference models), plus 16 variance-partitioning fits reported in the associated study, for a total of 72 fitted .npy artifacts. Reproducing predictions for new images or refitting the analyses additionally requires the corresponding pretrained ResNet-18 checkpoints and the original feature-extraction, spatial-pooling, PCA, concatenation, normalization, and model-fitting procedures."
}