The concept of "AGI" (a system that can match or exceed human performance across all tasks) shares all of the defects of the Turing Test. It defines "intelligence" entirely in terms of human performance. 1/ It says that the most important AI system capabilities to create are exactly those things that people can do well. But is this what we want? Is this what we need? 2/

Yoshua Bengio makes many good points in his latest blog post. I would like to share a few thoughts yoshuabengio.org/2023/05/22/how… Yoshua devotes a fair amount of space to discussing alignment and work on AI safety. I strongly agree that building good tools, techniques, and "safety culture" will reduce the risk that we create "rogue" systems accidentally. 2/

@EzraJNewman 1. A model of the world, both the current state (the things it believes are true) and the dynamics (the things it believes will be true in the future, conditioned on its actions, the actions of other agents, and existing active processes) 1/ @EzraJNewman 2. A model of itself and of other individual people (theory of mind). What they believe right now; theirs preferences and desired. Dynamics of social interactions. 2/

In a new paper, Jesse Hostetler (SRI) and I make a contribution to the growing literature on conformal prediction. We provide *prospective guarantees* on the behavior of an agent executing a fixed policy. 1/ Suppose you wish to dispatch your personal robot to walk across campus to the cafeteria to pick up your breakfast and bring it back to your room. The robot has learned a policy pi via reinforcement learning that it will execute to perform this task 2/

"Sentient" is being misapplied by many ML folks. It means "the ability to perceive or feel things" or "capable of experiencing things through its senses". Like many other categories, it is a matter of degree. 1/ The simplest form is a reflex. When a sensor measures a value above some threshold, the system (e.g., a thermostat) responds. The system lacks any awareness, memory, or ability to reflect on this "experience". 2/

Thoughts upon reading arxiv.org/abs/2106.15590:
In this paper, the authors compare highly-cited papers from 2008-2009 with papers from 2018-2019 published in NeurIPS and ICML and summarize the values (i.e., desirable aspects) highlighted in those papers. 1/ The values are exactly what you would expect of an engineering field: ML researchers want algorithms that work well, that are computationally efficient, that generalize beyond the training data, that are novel, and that apply to the real world. 2/

I like this Declaration of Interdependence. I strongly endorse the design principle that systems be "designed according to a notion of ways in which the people depend on the computational systems and ways in which computational systems depend on the people" 1/

https://twitter.com/ddwoods2/status/1430213797881557001

However, I think the declaration (and the table at the end) over-estimates human capabilities. We create human organizations precisely because individual humans cannot scale their sensitivity to context, to anomalies, and ability to adapt. 2/

Question that arose while discussing the @emilymbender et al. digital parrots paper: Is anyone studying how to condition language models on personas? If I ask it to compose a document on my behalf, I want it to write in an AI professor voice, for example. 1/ When translating my speech during a visit to another country, I want it to be polite and not profane or insulting. I might also want to use such a model to check my writing--to find aspects that might offend others. 2/

A review of “Underspecification Presents Challenges for Credibility in Modern Machine Learning” by D’Amour et al.

arxiv.org/abs/2011.03395 0/ Main claim: ML pipelines exhibit “underspecification”. i.e., they can produce a range of fitted models that all give similar iid test set performance but that give very different performance on “stress tests”. 1/

Excellent story by @Oregonian on efforts to reduce gun suicide. oregonlive.com/pacific-northw… Addressing suicide by improving safety culture is probably the most effective way of reducing gun deaths. 3 out of 5 gun deaths are suicides (more in Oregon). Estimates are that suicides can be reduced 30-50% by making guns harder to access during suicidal crises. 2/

I've been trying to imagine how the ML research and publication enterprise could be re-organized. Here are some initial thoughts. Feedback welcome! 1/ My proposal has four components. At the core would be a wiki containing three major components: (a) an annotated and organized bibliography of published papers, (b) an organized set of ML experiment designs and analysis pipelines (with code), 2/

A few reflections on Kalev Leetaru's opinion piece in Forbes 1/
forbes.com/sites/kalevlee… I strongly agree that we need to fight the hype about Deep Learning, and Leetaru correctly criticizes the rampant anthropomorphism underlying much of this hype. I also share deeply Leetaru's concern about deploying DL in high risk applications. 2/

I love Cynthia Rudin's work, but I don't think she is entirely fair to machine learning in this presentation 1/

https://twitter.com/berkustun/status/1052643311612235776

The goal of interpretable ML is important, and I am a huge fan of decision tree (CART) and rule-learning methods. However, interpretability can be very hard to achieve, whereas things like random forests and deep nets can be easily trained and give high performance 2/

I think people (including the authors) are over-interpreting arxiv.org/abs/1711.11561 "Measuring the tendency of CNNs to learn surface statistical regularities" 1/ The authors claim "deep CNNs tend to learn surface statistical regularities in the dataset RATHER THAN higher-level abstract concepts". [emphasis mine] 2/

Disappointing article by @GaryMarcus. He barely addresses the accomplishments of deep learning (eg NL translation) and minimizes others (eg ImageNet with 1000 categories is small ("very finite") ?). 1/

External Tweet loading...
If nothing shows, it may have been deleted
by @GaryMarcus view original on Twitter

DL learns representations as well as mappings. Deep machine translation reads the source sentence, represents it in memory, then generates the output sentence. It works better than anything GOFAI ever produced. 2/

.@katecrawford makes many very important points. Here is my attempt to translate some of them into engineering terminology. 1/

External Tweet loading...
If nothing shows, it may have been deleted
by @cjmay4754 view original on Twitter

As ML researchers and practitioners, we tend to formulate the problem as learning a mapping from inputs X to outputs Y. Crawford's point about bias in the data is that we aren't observing the real inputs or the real outputs 2/