Online Continual Learning for Embedded Devices
By Tyler L. Hayes, Christopher Kanan. 2022
73 citations recorded by the source corpus.
Real-time on-device continual learning is needed for new applications such as home robots, user personalization on smartphones, and augmented/virtual reality headsets. However, this setting poses unique challenges: embedded devices have limited memory and compute capacity and conventional machine learning models suffer from catastrophic forgetting when updated on non-stationary data streams. While several online continual learning models have been developed, their effectiveness for embedded applications has not been rigorously studied. In this paper, we first identify criteria that online continual learners must meet to effectively perform real-time, on-device learning. We then study the efficacy of several online continual learning methods when used with mobile neural networks. We measure their performance, memory usage, compute requirements, and ability to generalize to out-of-domain inputs.
Subfields: Artificial Intelligence
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Youngjae Min
- Hyo-Sang Shin
- Namhoon Cho
- A. Rupam Mahmood
- Mohamed Elsayed
- David Picard
- A. Mahmood
- Bertrand Delezoide
Citations in the corpus
- FeTrIL: Feature Translation for Exemplar-Free Class-Incremental Learning (2022, background): “CIL algorithms are needed when data arrive sequentially and/or computational constraints are important [9, 17, 24, 29].”
- Memory-efficient Reinforcement Learning with Knowledge Consolidation (2022, background): “The requirement of a large buffer prevents the application of RL algorithms to the real world since it creates a heavy memory burden, especially for onboard and edge devices (Hayes et al. 2019, Hayes and Kanan 2022).”
- Addressing Loss of Plasticity and Catastrophic Forgetting in Continual Learning (2024, methodology): “Although a few methods address both issues simultaneously, such methods expect known task boundaries, maintain a replay buffer, or require pretraining, which does not fit streaming learning (Hayes et al. 2022).”
- Π-ORFit: One-Pass Learning with Bregman Projection (2024, background): “Experience re-play schemes were investigated in [15], [16] for preventing catastrophic forgetting while mainly concerning memory efficiency.”
- Online Continual Learning For Interactive Instruction Following Agents (2024, background): “…by the agents as indicators of the ‘quality’ of the newly obtained logits ( i.e. , how much they contain accurate knowledge of the corresponding tasks), as empirically observed that high confidence tends to have high accuracy in Figure 3. capacity to store all task data (Hayes & Kanan, 2022).”