We propose Wake-Sleep Consolidated Learning (WSCL), a learning strategy
leveraging Complementary Learning System theory and the wake-sleep phases of
the human brain to improve the performance of deep neural networks for visual
classification tasks in continual learning settings. Our method learns
continually via the synchronization between distinct wake and sleep phases.
During the wake phase, the model is exposed to sensory input and adapts its
representations, ensuring stability through a dynamic parameter freezing
mechanism and storing episodic memories in a short-term temporary memory
(similarly to what happens in the hippocampus). During the sleep phase, the
training process is split into NREM and REM stages. In the NREM stage, the
model’s synaptic weights are consolidated using replayed samples from the
short-term and long-term memory and the synaptic plasticity mechanism is
activated, strengthening important connections and weakening unimportant ones.
In the REM stage, the model is exposed to previously-unseen realistic visual
sensory experience, and the dreaming process is activated, which enables the
model to explore the potential feature space, thus preparing synapses to future
knowledge. We evaluate the effectiveness of our approach on three benchmark
datasets: CIFAR-10, Tiny-ImageNet and FG-ImageNet. In all cases, our method
outperforms the baselines and prior work, yielding a significant performance
gain on continual visual classification tasks. Furthermore, we demonstrate the
usefulness of all processing stages and the importance of dreaming to enable
positive forward transfer.
Analysis of Wake-Sleep Consolidated Learning
Wake-Sleep Consolidated Learning (WSCL) is a novel learning strategy that draws inspiration from the Complementary Learning System theory and the wake-sleep phases of the human brain. This approach aims to enhance the performance of deep neural networks in visual classification tasks in continual learning settings. The multi-disciplinary nature of this concept lies in its combination of neuroscience principles and machine learning techniques.
The WSCL method operates through a synchronization between distinct wake and sleep phases, mirroring the behavior of the human brain. During the wake phase, the model is exposed to sensory input and fine-tunes its representations for stability using a dynamic parameter freezing mechanism. This stability ensures that the model retains previously acquired knowledge while adapting to new input. Additionally, episodic memories are stored in a short-term temporary memory, similar to the way the hippocampus functions in humans.
In the sleep phase, the training process is further divided into NREM and REM stages. In the NREM stage, synaptic weights of the model are consolidated by replaying samples from both short-term and long-term memory. This consolidation process activates the synaptic plasticity mechanism, reinforcing important connections and weakening unimportant ones. By actively strengthening relevant connections, the model becomes more effective at classifying visual inputs.
In the REM stage, the model is exposed to previously-unseen realistic visual sensory experiences. This phase activates the dreaming process, which allows the model to explore potential features in the input space. By exploring new patterns and possibilities, the model prepares its synapses to acquire future knowledge. This dreaming stage is crucial for enabling positive forward transfer, where previous learning helps facilitate new learning.
The effectiveness of the WSCL approach has been evaluated on three benchmark datasets: CIFAR-10, Tiny-ImageNet, and FG-ImageNet. In all cases, WSCL outperformed existing baselines and previous approaches, achieving significant performance gains in continual visual classification tasks. This demonstrates the potential of leveraging concepts from neurobiology to drive advancements in machine learning.
Conclusion
Wake-Sleep Consolidated Learning presents a promising strategy for improving the performance of deep neural networks in continual learning settings. By emulating the wake-sleep phases of the human brain and incorporating the Complementary Learning System theory, this approach tackles the challenge of retaining previous knowledge while adapting to new information. The multi-disciplinary nature of this research, combining insights from neuroscience and machine learning, highlights the potential for interdisciplinary collaborations to drive progress in artificial intelligence.