Code used for paper Cerebro-cerebellar networks facilitate learning through feedback decoupling (see also Cortico-cerebellar networks as decoupling neural interfaces).
Beyond the standard python libraries, you will need Pytorch (we use version 1.6.0, but later should work) to define the neural network models and as well as ignite (we use version 0.2.1) which wraps the training regime.
For the linedraw and seqmnist-based tasks:
- Go into the /scripts folder and choose the corresponding folder according to the experiment you want to run.
Open the train-model.py file. - In the file, manually define the path to the src folder (e.g. Documents/ccDNI/src) and where to save results (e.g. Documents/ccDNI/results) for your system. These are defined in the src_path and root variables respectively.
- To run within IDE:
- Select the hyperparameters of the experiment (see main ones below). E.g. to run the ccRNN model for 5 epochs set args.model = 'DNI_LSTM' and args.epochs = 5
- Run the file (press play button). To run from terminal
- Comment out suggested hyperparameter values (underneath where the args variable is defined)
- Set the train-model.py file location as the current working directory
- Run the python command on the file with desired arguments - e.g. python train-model.py -model=DNI_LSTM -epochs=5
- Check the results are saved in the folder defined in step 2. The resulting numpy file should have shape (a, b, c, d), where a is the number of number of seeds, b is the number of models (usually just one), c is the number of different metrics (e.g. train and validation MSE), and d is the number of epochs.
For the image captioning task:
- Go into the /other/image-captioning folder.
- Run the bash file download.sh to download the dataset with the command ./download.sh.
- Preprocess the data by running the build_vocab.py and resize.py scripts.
- Go into the train.py file. Configure data paths (i.e. vocab_path, image_dir, caption_path) based on destinations of steps 2, 3 above. Configure path to save model (model_path). Replace also the save_path variable defined in the get_fn() method to where you wish the model losses to be saved.
- Run the train.py file with the desired experiment hyperparameters (either directly in IDE or via terminal, see above).
- Check/plot model losses are saved in the save_path defined in 4.
- To sample a caption for an example image post training, run the sample.py file with the decoder_path set as the filepath of the trained model.
Example plotting code can be found in the /plotting directory. To the learning curves and trained model output for the simple line drawing task as shown in Figure 2 in the Neurips paper, run the linedraw_plotting.py file (to be added soon). To plot the learning curves under different levels of cerberal feedback (i.e. backpropagation truncation sizes) for the seqmnist line-drawing and digit-drawing tasks as in Figure S3 in Neurips, run the seqmnist_curves_all.py file.
The primary hyperparameters with which we modify our experimental setup can be listed as
- model - type of model used. Can be standard 'LSTM' (cRNN) or DNI enabled 'DNI_LSTM' (ccRNN)
- bptt - cerebral temporal feedback (when normalised by sequence length of task)/truncation size.
- spars-int - degree of sparseness in teacher feedback for task. E.g. for Fig 2 in Neurips spars-int=2 so that teacher feedback is only available every other timestep.
For the seqmnist-based drawing tasks:
- fixed-mag - Fixed magnitute for each line. Set as true to do line-drawing task.
- digit-draw - set as true to do digit-drawing task
Pytorch experiment structured according to https://github.com/miltonllera/pytorch-project-template
The DNI implementation is based on https://github.com/koz4k/dni-pytorch
References:
Boven, Pemberton et al. bioRxiv, https://www.biorxiv.org/content/10.1101/2022.01.28.477827v1
Pemberton, Boven et al NeurIPS 2021, https://proceedings.neurips.cc/paper/2021/hash/3ffebb08d23c609875d7177ee769a3e9-Abstract.html