Continual Learning Through Synaptic Intelligence

Authors: Friedemann Zenke, Ben Poole, Surya Ganguli

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experiments We evaluated our approach for continual learning on the split and permuted MNIST (Le Cun et al., 1998; Goodfellow et al., 2013), and split versions of CIFAR-10 and CIFAR-100 (Krizhevsky & Hinton, 2009).
Researcher Affiliation Academia 1Stanford University.
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No No explicit statement or link providing access to source code for the methodology was found.
Open Datasets Yes We evaluated our approach for continual learning on the split and permuted MNIST (Le Cun et al., 1998; Goodfellow et al., 2013), and split versions of CIFAR-10 and CIFAR-100 (Krizhevsky & Hinton, 2009).
Dataset Splits Yes However, here we used ξ = 0.1 and the value for c = 0.1 was determined via a coarse grid search on a heldout validation set.
Hardware Specification No No specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running experiments were mentioned.
Software Dependencies No The paper mentions using 'Adam (Kingma & Ba, 2014)' but does not specify software versions for programming languages, libraries, or other dependencies.
Experiment Setup Yes We used a small multi-layer perceptron (MLP) with only two hidden layers consisting of 256 units each with Re LU nonlinearities, and a standard categorical cross-entropy loss function plus our consolidation cost term (with damping parameter ξ = 1 10 3). To avoid the complication of crosstalk between digits at the readout layer due to changes in the label distribution during training, we used a multi-head approach in which the categorical cross entropy loss at the readout layer was computed only for the digits present in the current task. Finally, we optimized our network using a minibatch size of 64 and trained for 10 epochs. To achieve good absolute performance with a smaller number of epochs we used the adaptive optimizer Adam (η = 1 10 3, β1 = 0.9, β2 = 0.999).