Loss of plasticity in deep continual learning
A study in Nature reveals that standard deep learning methods lose plasticity over time, leading to performance declines. It proposes continual backpropagation to maintain adaptability without retraining from scratch.
Read original article
The study published in Nature investigates the loss of plasticity in deep continual learning, revealing that standard deep-learning methods gradually lose their ability to learn effectively when exposed to new data over time. This phenomenon, termed "loss of plasticity," was demonstrated using classic datasets like ImageNet and CIFAR-100, where traditional backpropagation methods resulted in performance declines, ultimately performing no better than shallow networks. The research highlights that while deep learning has achieved significant advancements, it struggles with continual learning, often requiring networks to be retrained from scratch when new data becomes available. The authors propose a solution through a new algorithm called continual backpropagation, which maintains plasticity by randomly reinitializing a small fraction of less-used units during training. This approach allows networks to adapt continuously without the performance degradation seen in conventional methods. The findings suggest that sustaining deep learning requires integrating a random, non-gradient component alongside traditional gradient descent techniques to preserve variability and adaptability in learning.
- Standard deep-learning methods lose plasticity over time in continual learning settings.
- Loss of plasticity leads to performance levels below that of shallow networks.
- Continual backpropagation can maintain plasticity by reinitializing less-used units.
- The study emphasizes the need for new strategies to enable effective continual learning in deep networks.
- Traditional methods often require retraining from scratch when new data is introduced.
Related
Analysing 16,625 papers to figure out where AI is headed next (2019)
MIT Technology Review analyzed 16,625 AI papers, noting deep learning's potential decline. Trends include shifts to machine learning, neural networks' rise, and reinforcement learning growth. AI techniques cycle, with future dominance uncertain.
AI models collapse when trained on recursively generated data
Recent research in Nature reveals "model collapse" in AI, where training on data from previous models leads to irreversible defects and misrepresentation of original data, emphasizing the need for genuine human-generated data.
The problem of 'model collapse': how a lack of human data limits AI progress
Research shows that using synthetic data for AI training can lead to significant risks, including model collapse and nonsensical outputs, highlighting the importance of diverse training data for accuracy.
Brain found to store three copies of every memory
A University of Basel study found the brain uses three neuron types for memory storage, revealing potential therapeutic applications for traumatic memories and highlighting the complexity of memory formation and retrieval.
When A.I.'s Output Is a Threat to A.I. Itself
A.I. systems face quality degradation from training on their own outputs, risking "model collapse." Ensuring diverse, high-quality real-world data is essential to maintain effectiveness and reliability in A.I. applications.
0 commentsRelated
Analysing 16,625 papers to figure out where AI is headed next (2019)
MIT Technology Review analyzed 16,625 AI papers, noting deep learning's potential decline. Trends include shifts to machine learning, neural networks' rise, and reinforcement learning growth. AI techniques cycle, with future dominance uncertain.
AI models collapse when trained on recursively generated data
Recent research in Nature reveals "model collapse" in AI, where training on data from previous models leads to irreversible defects and misrepresentation of original data, emphasizing the need for genuine human-generated data.
The problem of 'model collapse': how a lack of human data limits AI progress
Research shows that using synthetic data for AI training can lead to significant risks, including model collapse and nonsensical outputs, highlighting the importance of diverse training data for accuracy.
Brain found to store three copies of every memory
A University of Basel study found the brain uses three neuron types for memory storage, revealing potential therapeutic applications for traumatic memories and highlighting the complexity of memory formation and retrieval.
When A.I.'s Output Is a Threat to A.I. Itself
A.I. systems face quality degradation from training on their own outputs, risking "model collapse." Ensuring diverse, high-quality real-world data is essential to maintain effectiveness and reliability in A.I. applications.