Hey folks,
When, or do, we call an AI actually intelligent? I'm sure many who have been working in the AI field for some time have had the following conversation, or something similar:
Sára the AI dev: "I just built a deep neural network model that can do X, surpassing superhuman level"
Bob the skeptic: "That's wonderful! Though AI is still not better than humans if you alter the task parameters a bit, or if you want it to do any other task"
Indeed, since the canonical examples of chess, Jeopardy, and Go, once a task is solved and the AI magician reveals the secret to their tricks, this once intelligent seeming behavior is no longer deemed intelligent. Perhaps it isn't a bad thing to move the goal back a bit, driving the field forward, and few would argue that any current AI system really is close to human level general intelligence. Would you?
This conversation came up a few times this week. An interesting point was made that if we might consider flipping the script on the problem. Are the mistakes the AI makes similar to those that humans might make? We see it in MNIST letters and other classification images, and further, a group has shown that in specific cases, humans fall victim to adversarial images just as AI would.
So is AI going to take over the world anytime soon? Probably not. Though in some cases, perhaps we might be going a bit hard on our expectations for some systems. That said, I'd much rather my driverless car classify images better than I can, so maybe raising the bar higher isn't a bad thing after all.
|
|
|
|
Presents adversarial images to human participants, who misclassify obvious images. By the Google Brain research team.
|
|
As many in our community feel, Continual learning is also poised to meet some of Bob's concerns, learning multiple tasks and configurations without necessarily needing an expert-like system to do so. Bob may bite his tongue when a system can learn not just chess, Jeopardy, or Go, but chess, Jeopardy, and Go. You also may have also noticed a few image recognition examples above, to pay homage to the upcoming CLVISION conference (more info below).
We're happy to share what we have been working on, towards accelerating research surrounding continual learning AI: a necessary step in the direction of strong AI. Also, keep an eye for a big CLAI announcement coming up in March. Passionate about our mission? Join us on slack if you haven't already and feel free to donate if you are passionate about this goal.
|
|
A Few Recent Announcements
- The Workshop "Continual Learning in Computer Vision" supported by ContinualAI, has open its CALL FOR PAPERS (deadline 20 March), you can already submit your work: short and long articles, archival or not, it does not matter, submit your original content to the workshop! The event is planned to be one of the biggest ever organized in our community, make sure to be there! Also consider submitting to the challenge.
- We had wonderful continual learning meetups the last couple of weeks, Rehearsal-Free Continual Learning (view the recording of the meetup here) and on our collaborative project on Colab 1.0 (view the recording of the meetup here). Stay tuned on slack for the next meetup, or subscribe to the open CL mailing-list Continual Learning for AI, for updates on meetups and more events (including those external to CLAI).
- We've maintained a great reading group line up. Can't make this week's reading group? No worries! See the past papers here, and you can also watch the recordings of all the events that we have had.
- The ContinualAI Lab collaborative team is always looking for contributors to the many open-source projects we have under development. Contact us on Slack if you want to learn more about them and join us! We are always looking for motivated people willing to give back to this awesome community!
- You may also consider submitting to the special issue of J. Imaging "Continual Learning in Computer Vision: Theory and Applications" (more info here)
|
|
Top paper picks:
A paper we think you should read, if you have not yet, as chosen by the community:
Unifying Regularisation Methods for Continual Learning
( Frederik Benzing )
Continual Learning addresses the challenge of learning a number of different tasks sequentially. The goal of maintaining knowledge of earlier tasks without re-accessing them starkly conflicts with standard SGD training for artificial neural networks. An influential method to tackle this problem without storing old data are so-called regularisation approaches. They measure the importance of each parameter for solving a given task and subsequently protect important parameters from large changes. In the literature, three ways to measure parameter importance have been put forward and they have inspired a large body of follow-up work. Here, we present strong theoretical and empirical evidence that these three methods, Elastic Weight Consolidation (EWC), Synaptic Intelligence (SI) and Memory Aware Synapses (MAS), are surprisingly similar and are all linked to the same theoretical quantity. Concretely, we show that, despite stemming from very different motivations, both SI and MAS approximate the square root of the Fisher Information, with the Fisher being the theoretically justified basis of EWC. Moreover, we show that for SI the relation to the Fisher -- and in fact its performance -- is due to a previously unknown bias. On top of uncovering unknown similarities and unifying regularisation approaches, we also demonstrate that our insights enable practical performance improvements for large batch training.
|
|
|
|
