Biological learning in key-value memory networks

Authors: Danil Tyulmankov, Ching Fang, Annapurna Vadaparty, Guangyu Robert Yang

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We compute the accuracy as a function of the number of stored stimuli (Figure 2a). Figure 3a shows the performance of the meta-learned algorithms in comparison to the corresponding simplified versions (section 2) for either sequential and random local third factors.
Researcher Affiliation Academia Danil Tyulmankov Columbia University dt2586@columbia.edu Ching Fang Columbia University ching.fang@columbia.edu Annapurna Vadaparty Columbia University Stanford University apvadaparty@gmail.com Guangyu Robert Yang Columbia University Massachusetts Institute of Technology yanggr@mit.edu
Pseudocode No The paper describes the model equations and update rules (e.g., Eq. 1-17) in mathematical notation and prose, but does not include a distinct pseudocode or algorithm block.
Open Source Code No The paper does not provide any statement regarding the release of open-source code for the described methodology, nor does it include links to a code repository.
Open Datasets No The paper states that stimuli are generated, e.g., 'All experiments use random uncorrelated binary random memories xt 2 {+1, 1}d unless otherwise stated.' and 'For the correlated memories (Fig 4c), we generate stimuli by picking a template pattern at random, and for each subsequent pattern we flip each bit with some probability.' It does not refer to a publicly available dataset with specific access information or citations.
Dataset Splits No The paper describes properties of the input stimuli, such as 'The network stores a set of T stimuli {xt} and the query ex is a corrupted version of a stored key (60% of the entries are randomly set to zero)', and mentions 'Training data consists of sequence lengths between T = N/2 and T = 2N'. However, it does not specify explicit training, validation, or test dataset splits in terms of percentages or counts, or refer to standard predefined splits.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory, or cloud resources) used to run the experiments.
Software Dependencies No The paper mentions using 'Adam' for optimization but does not provide specific version numbers for any software dependencies, programming languages, or libraries used for implementation.
Experiment Setup Yes optimizing these parameters using stochastic gradient descent (Adam, [Kingma and Ba, 2014]). Empirically, we find that p 4/N produces desirable performance. a stronger global third factor (qt = 10). we introduce an empirically chosen synaptic decay parameter λ = 0.95 to the Hopfield network. (d = N = 40, from Figure 2 caption).