Engram Memory Network: Brain-Inspired Prototype Explanations

Official implementation of "Engram Memory Network: Brain-Inspired Prototype Explanations" (IJCNN 2026, WCCI 2026).

Requirements

pip install -r requirements.txt

Training

First train the backbone (encoder + decoder, without the engram memory module), then train EMN.

Supported datasets: CIFAR10, CIFAR100, CIFAR100SUPER, MNIST, HAM, CUB. Other datasets can be added by extending dataloader.py.

# Train backbone
python run_backbone_train.py --opt deit_small_bb_fr --dataset_name CIFAR10 --number 1

# Train EMN
python run_train.py \
  --dataset_name CIFAR10 \
  --backbone_name deit_small_bb_fr \
  --backbone_path backbone_weights/deit_small_bb_fr/CIFAR10_1 \
  --marker CIFAR10_1

Evaluation

python eval.py --load_model_path logs/CIFAR10_CIFAR10_1 --device_id 0

Computes accuracy, LPIPS (AlexNet/VGG), and DISTS. Results are saved to eval.csv.

Results

Visual Similarity on CIFAR-10 (x10⁻¹, lower is better)

Model	LPIPS Top-1	LPIPS Top-3	DISTS Top-1	DISTS Top-3
ProtoVAE	7.50 ± 0.06	7.50 ± 0.07	6.75 ± 0.11	6.76 ± 0.11
ProtoPNet	6.95 ± 0.19	7.00 ± 0.32	4.07 ± 0.13	4.02 ± 0.12
ProtoViT	6.26 ± 0.10	6.31 ± 0.07	3.67 ± 0.03	3.69 ± 0.03
ProtoPFormer	6.23 ± 0.02	6.24 ± 0.02	3.68 ± 0.01	3.69 ± 0.02
EMN (ours)	5.38 ± 0.02	5.50 ± 0.02	3.22 ± 0.01	3.28 ± 0.01

Classification Accuracy (%)

Model	CIFAR-10	CIFAR-100	MNIST	HAM10000	CUB-200
ProtoVAE	81.59	50.37	99.05	—	—
ProtoPNet	94.30	68.63	99.06	62.92	22.43
ProtoViT	97.63	86.41	99.09	80.49	74.09
ProtoPFormer	97.85	87.83	99.41	66.95	50.51
EMN (ours)	96.39	85.28	99.19	88.51	68.74

Subclass Alignment on CIFAR-100S (%, higher is better)

Model	Top-1	Top-3
ProtoPNet	2.28 ± 0.91	1.95 ± 0.47
ProtoViT	15.93 ± 6.55	14.03 ± 7.04
ProtoPFormer	19.49 ± 1.60	19.42 ± 1.38
EMN (ours)	39.87 ± 0.72	36.07 ± 0.36

Project Structure

├── run_train.py              # EMN training entry point
├── run_backbone_train.py     # Backbone training entry point
├── train.py                  # EMN training loop
├── eval.py                   # Evaluation (accuracy, LPIPS, DISTS)
├── dataloader.py             # Dataset loading
├── requirements.txt
├── args/
│   └── get_args.py           # Argument parser
├── emn/
│   └── emn.py                # EMN model (Eq. 1-4)
├── backbone/
│   ├── utils.py              # Encoder/decoder factory
│   ├── backbone_train.py     # Backbone training loop
│   ├── backbone_model.py     # Backbone model (encoder+decoder)
│   ├── encoder.py            # DeiT-Small encoder
│   ├── decoder.py            # ViT decoder
│   └── module.py             # Transformer blocks
├── visualize/                # Explanation visualization
└── assets/                   # Figures for README

Acknowledgements

• LPIPS — Learned Perceptual Image Patch Similarity
• DISTS — Deep Image Structure and Texture Similarity
• ProtoPNet, ProtoVAE, ProtoPFormer, ProtoViT — Baseline implementations

Citation

@inproceedings{emn2026,
  title={Engram Memory Network: Brain-Inspired Prototype Explanations},
  author={Kim, Hyunjun and Ha, Myoung Hoon},
  booktitle={International Joint Conference on Neural Networks (IJCNN)},
  year={2026}
}