Lifelong Learning with Non-i.i.d. Tasks

Authors: Anastasia Pentina, Christoph H. Lampert

NeurIPS 2015 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental For the second scenario we propose to learn an inductive bias in form of a transfer procedure. We present a generalization bound and show on a toy example how it can be used to identify a beneficial transfer algorithm. Numeric experiments confirm that by optimizing J (θ) with respect to θ one can obtain an advantageous angle: using n = 2, . . . , 11 tasks, each with m = 10 samples, we obtain an average test error of 14.2% for the (n + 1)th task.
Researcher Affiliation Academia Anastasia Pentina IST Austria Klosterneuburg, Austria apentina@ist.ac.at Christoph H. Lampert IST Austria Klosterneuburg, Austria chl@ist.ac.at
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any explicit statements about releasing source code or links to a code repository.
Open Datasets No The paper describes a 'toy example' with generated data ('sampled from a non-stationary environment') but does not specify a public dataset or provide access information (link, DOI, citation) for the data used in the numerical experiments.
Dataset Splits No The paper describes a sequential task-based setup where 'n' observed tasks are used to evaluate performance on a '(n+1)th' future task, but does not provide specific train/validation/test splits (e.g., percentages, counts, or stratified methods) for a single dataset within or across tasks.
Hardware Specification No The paper does not provide any specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with version numbers).
Experiment Setup Yes We use a one-parameter family of transfer algorithms, Aα for α R. Given sample sets Sprev and Scur, any algorithm Aα first rotates Sprev by the angle α, and then trains a linear support vector machine on the union of both sets. [...] For that we set Qi = N(wi, I2), i.e. unit variance Gaussian distributions with means wi. Similarly, we choose all reference prior distributions as unit variance Gaussian with zero mean, Pi = N(0, I2). Analogously, we set the hyper-prior P to be N(0, 10), a zero mean normal distribution with enlarged variance in order to make all reasonable rotations α lie within one standard deviation from the mean. As hyper-posteriors Q we choose N(θ, 1) and the goal of the learning is to identify the best θ.