wm-computational-limits

Are there computational limits on human working memory (WM) capacity aside from anatomical limits?

The directory workingmem.task.SIR contains a version of the Store-Ignore-Recall (SIR) task used in human experiments to tax working memory (CITE). The task involves storing and recalling items stored in virtual WM 'slots', here, 'registers'. In humans, the task requires active role-addressable maintenance of information. In computational models, the retrieval may be facilitated by a number of strategies.

To generate a dataset, use: ``` python -m workingmem.task.SIR -h usage: main.py [-h] [--nreg NREG] [--nitems NITEMS] [--seqlen SEQLEN] [--concurrentreg CONCURRENTREG] [--concurrentitems CONCURRENTITEMS] [--heldoutreg HELDOUTREG] [--heldoutitems HELDOUTITEMS] [--locality LOCALITY] [--ignoreprob IGNOREPROB] [--samediffprob SAMEDIFFPROB] [--ntrain NTRAIN] [--nval NVAL] [--ntest NTEST] [--split SPLIT] [--basedir BASEDIR] [--seed SEED] [--generate]

Generate a dataset for the SIR task, or load one if it already exists, and output a few examples

options: -h, --help show this help message and exit --nreg NREG Number of registers (vocab). [100] --nitems NITEMS Number of items (vocab). [100] --seqlen SEQLEN Sequence length (trials in a sequence). [100] --concurrentreg CONCURRENTREG Number of concurrent registers. [2] --concurrentitems CONCURRENTITEMS Number of concurrent items. [4] --heldoutreg HELDOUTREG Held-out registers for testing. --heldoutitems HELDOUTITEMS Held-out items for testing. --locality LOCALITY Locality for the tasks. --ignoreprob IGNOREPROB Probability of ignoring an item. --samediffprob SAMEDIFFPROB Probability of 'same' outcome [.5] --ntrain NTRAIN Number of training samples. [10,000] --nval NVAL Number of validation samples. [2000] --ntest NTEST Number of test samples. [2000] --split SPLIT Dataset split to use. --basedir BASEDIR Base directory to save or load data. --seed SEED random seed for dataset generation --generate generate the dataset splits and store to disk? if not passed, will simply output a single example generated on-the-fly. ```

Examples look of this sort: St reg_54 item_4 diff St reg_54 item_81 diff St reg_60 item_81 diff St reg_60 item_81 diff St reg_54 item_81 same St reg_54 item_81 same St reg_60 item_40 diff Ig reg_54 item_81 same St reg_54 item_81 same Ig reg_54 item_81 diff St reg_60 item_81 diff St reg_54 item_81 diff St reg_60 item_40 diff St reg_60 item_4 diff Ig reg_54 item_81 same

The main module (-m workingmem), implemented with an entrypoint in workingmem/__main__.py does the orchestrating of loading/constructing datasets, training/evaluating models. To see the options, run python -m workingmem -h.

The module allows you to integrate the main script with Weights & Biases (wandb)