IC Colloquium: Towards understanding the learning dynamics of neural networks
By: Razvan Pascanu - Google DeepMind
Video of his talk
Abstract
In this topic I will discuss a topic that I've been interested in since my PhD, the learning dynamics of neural networks. Specifically I will try to provide one intuition of why learning tends to be computationally and data inefficient for deep learning, using this as an entry point to introduce the topic of continual learning. The main argument is that interference between gradients coming from different modes of the data can lead to them being learned sequentially even though data needs to be sampled in an IID fashion from the entire distribution. I will summarize how continual learning might relate to this issue and some of the main themes within this growing subfield. One problem that I would spend a bit more time on is that of plasticity loss which has been a focus in continual reinforcement learning in the last couple of years. If time allows, I would touch on the topic of generalization, the main goal of learning, and in particular of compositional generalization as well as the concept of alignment, explored more thoroughly within the space of algorithmic reasoning.
Bio
Razvan Pascanu has been a research scientist at Google DeepMind since 2014. Before this, he did his PhD in Montréal with prof. Yoshua Bengio, working on understanding deep networks, recurrent models and optimization. Since he joined Google DeepMind he has also had significant contributions in deep reinforcement learning, continual learning, meta-learning, graph neural networks as well as continuing his research agenda of understanding deep learning, recurrent models and optimization. Please see his scholar page for specific contributions. He is also actively promoting AI research and education as a main organizer of Conference on Life-long Learning Agents (CoLLAs) lifelong-ml.cc , Eastern European Machine Learning Summer School (EEML) www.eeml.eu and www.workshops.eeml.eu as well as different workshops at NeurIPS, ICML and ICLR.
More information
Video of his talk
Abstract
In this topic I will discuss a topic that I've been interested in since my PhD, the learning dynamics of neural networks. Specifically I will try to provide one intuition of why learning tends to be computationally and data inefficient for deep learning, using this as an entry point to introduce the topic of continual learning. The main argument is that interference between gradients coming from different modes of the data can lead to them being learned sequentially even though data needs to be sampled in an IID fashion from the entire distribution. I will summarize how continual learning might relate to this issue and some of the main themes within this growing subfield. One problem that I would spend a bit more time on is that of plasticity loss which has been a focus in continual reinforcement learning in the last couple of years. If time allows, I would touch on the topic of generalization, the main goal of learning, and in particular of compositional generalization as well as the concept of alignment, explored more thoroughly within the space of algorithmic reasoning.
Bio
Razvan Pascanu has been a research scientist at Google DeepMind since 2014. Before this, he did his PhD in Montréal with prof. Yoshua Bengio, working on understanding deep networks, recurrent models and optimization. Since he joined Google DeepMind he has also had significant contributions in deep reinforcement learning, continual learning, meta-learning, graph neural networks as well as continuing his research agenda of understanding deep learning, recurrent models and optimization. Please see his scholar page for specific contributions. He is also actively promoting AI research and education as a main organizer of Conference on Life-long Learning Agents (CoLLAs) lifelong-ml.cc , Eastern European Machine Learning Summer School (EEML) www.eeml.eu and www.workshops.eeml.eu as well as different workshops at NeurIPS, ICML and ICLR.
More information
Practical information
- General public
- Free
Contact
- Host: Caglar Gulcehre