Too early for continual learning: harder tasks and agent systems

I was recently fortunate to have interned with Jason Weston and Stephen Roller at FAIR. Jason is well known for Memory Networks (though he should be even more well known for NLP almost from scratch). Memory Networks, Neural Turing Machines, and Stack NNs came out several years ago. Despite their popularity, they haven't quite had staying power.

One reason why memory networks may have fallen out of favor is that benchmark tasks haven't actually required memory.

I believe continual learning is in a similar place.

There has been some interesting work in past years on continual learning, including elastic weight consolidation, progressive neural networks, PathNet, and efficient lifelong learning with A-GEM, with more recent papers especially focusing on evaluation tasks and protocols.

However, the evaluation setups at present are a bit contrived, and I think may not be motivating enough for the field.

One of the first projects in NLP I was aware of was NELL, the Never-ending Language Learner. It's a continuously running system to extract structured, relational knowledge from the ever-growing web. I think this work hints at the future - continual learning will benefit significantly and re-emerge once we have capable, deep-learning based agents.