experiments repo for adding sample memory to a neural network. The network stores features from previously examined samples up to a threshold amount before it begins to "forget" samples.
Samples are stored and retrieved using the Facebooks Faiss library
The easiest installation is with using anaconda
conda create -yn nnmemories python=3
conda env update -f requirements.yml -n nnmemories
conda activate nnmemories
python main.py
Test set: Average loss: 0.0711, Accuracy: 9771/10000 (98%)
python main.py --epochs 10 --memory_epoch 1
Test set: Average loss: 0.1455, Accuracy: 9576/10000 (96%)
python main.py --epochs 10 --memory_epoch 3
Test set: Average loss: 0.1160, Accuracy: 9658/10000 (97%)
python main.py --epochs 10 --memory_epoch 5
Test set: Average loss: 0.0946, Accuracy: 9716/10000 (97%)