flooo

Now consider that the genius cannot physically interact with the world or the people therein, and uses her eyes only for reading text.

I have several chat groups with friends and even with my wife to avoid this topic. With my wife, I have several now: one to share recipes, one to share the weeks meal plan, one to talk about the activities of our two year old and one to share houses we might be interested in buying.

It’s great to separate for record keeping, but mostly to avoid forcing the conversation on some organisational thing when the other just needs eg to vent about something at work

To me, the vast majority of people in the field seem to hold on to the idea that different technology suits different problems.

Also, are you aware that one of the most prominent AI tools of this month (ChatGPT) was obtained with RL!

How about the women whose profile pics he “harvested” (stole)?

Isn’t there a huge ethical issue there?

Most novel AI serves the user perfectly well. And since this type of AI requires lots of labelled data, these users are typically large data harvesting organisations.

Some other applications that you may use daily are translate and face unlock/recognition.

It’s interesting that you mention search and spam filtering which both include an adversarial component. It seems to me that the adversarial AI has become better, in line of expectation from the democratisation of AI tools and knowledge.

Sounds like excel flash fill might be similar to what you’re interested in. It’s a well documented feature described in papers and talks, it comes from a pretty cool group in MS research.

Stochastic methods may not be able to do so in isolation, but they can be used in tandem with other approaches such as in “A Deep Reinforcement Learning Approach to First-Order Logic Theorem Proving” https://ojs.aaai.org/index.php/AAAI/article/view/16780/16587

I’m a researcher working in the fraud detection domain.

Do you have some pointers on scaling GNNs to such large problems?

A colleague of mine recently published a paper on ML-based cell counting for human cancer cells. It's open access and available here https://www.mdpi.com/2076-3417/11/11/4912

I am currently interested in infusing reinforcement learners with symbolic knowledge, with safety constraints as a special case.

I hope this helps cases where learners could come up with better solutions if it were not for pathological failures that we know to avoid.

Also, I try to keep expectations around AI reasonable.

As a researcher in AI, I accept that a lot of currently unsolved challenged are thought of as AI. But lately, I feel that AI is the problem description for all currently unsolved problems. And then some...

This surprises me, because most AI technologies have been around for a long time. Now with blockchain a couple of years ago, I could at least rationalize all excitement as people throwing new technology at an old problem. But with AI I am continually surprised by the reasons why 'an AI' would be able to solve it.

Vowpal Wabbit is used but many build something in-house into existing rule-based or supervised ml systems

In green fields deployments, Azure personalizer may be a nice place to start looking.

On the academic side, this paper provides an overview https://content.iospress.com/articles/data-science/ds200028

Not entirely the same, but this may be helpful https://spinningup.openai.com/en/latest/user/introduction.ht...

Previously discussed on HN: 1. https://news.ycombinator.com/item?id=18408360 2. https://news.ycombinator.com/item?id=24184270

RL, including contextual bandits, is becoming more popular for personalization, i.e. adapting some system to the preferences of (groups of) individuals.

Plug/Source: I did a lit. review on this topic https://doi.org/10.3233/DS-200028

There is an interesting relation to this with regard to reinforcement learning (RL) where the trade-off between exploration and exploitation is one of the fundamental issues. Systematic or structured exploration has shown some success here as well.

What 'structured' means here deeply depends on the representation the agent has of the world, i.e. which states and actions are similar to each other. Current RL setups typically learn such representations on the go. Abstractions can be very powerful here.

Arguably, people also perform structured exploration on learned representations. It may be the case that the adults have learned a representation of the 'game' and their lies little exploratory value in suboptimal actions.

Critiques that violent games and movies numb youths simply by exposure and lead them to become violent in real life eventually follow this exact same patterns.

I'm surprised there is no mention of Postgres' Genetic Query Optimizer (GEQO) here. It kicks in when there is a large number of joins and reduces query planning time at the cost of query execution time.

Another post in this thread mentions adaptive query planning and mistakenly imply that the GEQO is a module for this. My hands have been itching to look into experimenting on some improvements on the GEQO, specifically by improving the genetic algorithms used. When there are many similar queries, so in the adaptive query planning setting, one could also use reinforcement learning to improve query planning over time.

Colleagues in my research group are working on evolving 'creatures' in physical space, e.g. evolvable robots [0]. They address deep questions regarding the interplay between evolution/the body, learning/the brain and the environment using evolutionary algorithms and machine learning.

[0] https://www.york.ac.uk/robot-lab/are/

A friend of mine has been doing this for some time. It works well but you're gonna need some iterations with stakeholders to make it work. This includes medical staff, transport experts etc. What's quite cool about my friends engineering is that they manage to keep the containers sterile by pressure https://hospitainer.com/

I was instantaneously thinking about machine learning from UCI data set repository.

Talking about dangers of AI!

Integrating control in adaptive systems. We see a big trend where (parts of) systems can adapt their behaviour to particularities of the environment they are deployed in with the biggest technical driver being machine learning.

Controlling this adaptivity is usually done in an ad hoc fashion by writing code surrounding the adaptive component, usually a ml model

This yields many problems ranging from "artificial stupidity" to problems with safety and degraded power of users.

By integrating control into adaptive systems, we might get the best of both worlds: tell the program what (not) to do where you know this upfront and let it figure out how to act by itself elsewhere

There is a link to the original work by OpenAI: https://openai.com/blog/reinforcement-learning-with-predicti...

This could be achieved by following the edges of a decision tree from the root to all leaf nodes [0]. Results are not always easily interpreted by humans.

[0] http://www.saedsayad.com/decision_tree.htm

This style using try/except instead of if/else is known as EAFP (Easier to Ask Forgiveness than Permission). It's actually recommended in the Python glossary [0].

[0] https://docs.python.org/3.5/glossary.html#term-eafp