RanDumb: Random Representations Outperform Online Continually Learned Representations

Authors: Ameya Prabhu, Shiven Sinha, Ponnurangam Kumaraguru, Philip Torr, Ozan Sener, Puneet Dokania

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical evaluations, summarized in Table 1 (left, top), reveal that despite replacing the representation learning with a pre-defined random representation, Ran Dumb surpasses current stateof-the-art methods in latest online continual learning benchmarks [76].
Researcher Affiliation Collaboration 1University of Oxford 2IIIT Hyderabad 3Apple
Pseudocode No The paper includes equations and figures (e.g., Figure 1, Figure 2) that illustrate the method, but it does not contain a formal pseudocode block or algorithm description.
Open Source Code Yes https://github.com/drimpossible/RanDumb
Open Datasets Yes We evaluate Ran Dumb using five datasets: MNIST, CIFAR10, CIFAR100, Tiny Image Net200, and mini Image Net100.
Dataset Splits Yes We use standard datasets and splits, we provide hyperparameters in experimental details along with ablations in experiment sections to understand the contribution of each component in our algorithm.
Hardware Specification Yes All experiments were conducted on a CPU server with a153 48-core Intel Xeon Platinum 8268 CPU and 392GB of RAM, requiring less than 30 minutes per experiment.
Software Dependencies No The paper mentions using 'Scikit-Learn implementation of Random Fourier Features' and 'LAMDA-PILOT codebase', but it does not specify the version numbers for these software components.
Experiment Setup Yes We used the Scikit-Learn implementation of Random Fourier Features [52] with 25K embedding size, γ = 1.0. We use progressively increasing ridge regression parameter (λ) with dataset complexity, λ = 10 6 for MNIST, λ = 10 5 for CIFAR10/100 and λ = 10 4 for Tiny Image Net200/mini Image Net100.