Categories
AI clinical

Comment: The danger of AI is weirder than you think.

AI can be really destructive and not know it. So the AIs that recommend new content in Facebook, in YouTube, they’re optimized to increase the number of clicks and views. And unfortunately, one way that they have found of doing this is to recommend the content of conspiracy theories or bigotry. The AIs themselves don’t have any concept of what this content actually is, and they don’t have any concept of what the consequences might be of recommending this content.

Shane, J. (2019). The danger of AI is weirder than you think. TED.

We don’t need to worry about AI that is conscious (yet), only that it is competent and that we’ve given it a poorly considered problem to solve. When we think about the solution space for AI-based systems we need to be aware that the “correct” solution for the algorithm is one that literally solves the problem, regardless of the method.

The danger of AI isn’t that it’s going to rebel against us, but that it’s going to do exactly what we ask it to.

Janelle Shane

This matters in almost every context we care about. Consider the following scenario. ICUs are very expensive for a lot of good reasons; they have a very specialised workforce, a very low staff to patient ratio, the time spent with each patient is very high, and the medication is crazy expensive. We might reasonably ask an AI to reduce the cost of running an ICU, thinking that it could help to develop more efficient workflows, for example. But the algorithm might come to the conclusion that the most cost-effective solution is to kill all the patients. According to the problem we proposed, this isn’t incorrect but it’s clearly not what we were looking for, and any human being on earth, including small children, will understand why.

Before we can ask AI-based systems to help solve problems we care about, we’ll need to first develop a language for communicating with them. A language that includes the common sense parameters that inherently bound all human-human conversation. When I ask a taxi driver to take me to the airport “as quickly as possible”, I don’t also need to specify that we shouldn’t break any rules of driving, and that I’d like to arrive alive. We both understand the boundaries that define the limits of my request. As the video above shows, an AI doesn’t have any “common sense” and this is a major obstacle for progress towards having AI that can address real world problems beyond the narrow contexts where they are currently competent.

Categories
AI

Comment: How Can AI Systems Understand Human Values?

…for ML systems to truly be successful, they need to understand human values. More to the point, they need to be able to weigh our competing desires and demands, understand what outcomes we value most, and act accordingly.

Creighton, J. (2019). How Can AI Systems Understand Human Values? The Future of LIfe Institute blog.

This article identifies three kinds of autonomous agents that help understand why this is an important question.

  • Reflex agents react to predetermined changes in the environment e.g. a thermostat regulates the temperature in a house.
  • Goal-based agents keep working until a predetermined goal has been reached e.g. analyse and identify every image in a set.
  • Utility-based agents make tradeoffs as part of following decision paths that maximise the total rewards e.g. route planning as part of a navigation app on your phone.

AIs will need to have models that allow them to roughly figure out our evaluations in totally novel situations, the kinds of value situations where humans might not have any idea in advance that such situations might show up.

The post makes clear that current ML systems rely on utility-based agents but that these agents must assume a set of priorities that don’t change. To stick with the route planning example, you may want to take the longer route when driving because you prefer to save money even if it costs you in time. However, when you’re late for an important meeting you may value your time more than the money, in which case you’ll want the shorter – more expensive – route. In other words, your values are dynamic and change depending on the context.

We get around this now by being presented with options when we first identify the route we want to take; the phone tells us that the shorter route has a toll road but the longer route will add 15 minutes to our trip. The software is sophisticated enough to know that these are differences that matter to us, but it is impossible for it to know what option is best today, in this moment.

This is just a simple example of finding an optimal route for you while driving, so imagine how complex the decision-making becomes when AI-based systems are implemented in hospitals and schools. While it makes sense to be asked what route you’d prefer when driving to a meeting, we can’t have situations where we’re asked every 5 minutes which of an arbitrary number of choices we’d prefer, given a wide variety of contexts. We’re going to have to give up some of the decision-making authority to machines. Which is why it really matters that we figure out how to get them to include human values in their choices.

Categories
conference PhD research

Research plan for 2011

I’ve just been asked by one of my supervisors for my research goals for 2011. This will include my own work, as well as planning how our undergraduate and Masters students’ work might feed into some of the bigger projects.┬áThe first goal I have for 2011 is to submit 2 articles based on my first PhD objective. I’d planned on sending these off at the end of 2010, but never managed to finish them in time. They’re almost done now, and I’m hoping to submit them before lectures begin in a few weeks.

In order to plan my research activities for 2010, I created a chart to help me visualise the different activities I’d be involved in (see below). I completed most of the tasks, although not necessarily exactly as I’d initially planned them. Looking back at the process I went through in 2010, it’s clear that things don’t always work out the way you planned them and that that isn’t always a bad thing. I ran out of time at the end of the year, and didn’t get to complete everything I’d wanted to. Now I have the dilemma of trying to decide if I still want to do them, and run the risk of biasing the results e.g. realising that students may not have great recall of certain events.

First PhD objective

I’ve created a similar chart for my 2011 progress, and tried to incorporate the results of the 2010 research I conducted, showing how they feed into my┬ásecond PhD objective. This includes a review of the undergraduate curriculum using document analysis, and a Delphi study of physiotherapy educators to plan a blended learning intervention for one (or several) undergraduate modules. The original plan was to identify one module and then develop it using a blended learning approach but now we’re considering the possibility of working with a few, although this isn’t reflected in the diagram below.

I still need to figure out how I’m going to incorporate input from my 4th year research group, as well as the 3 Masters students I’m supervising this year. I can’t really see how they’ll fit in, so I might need to start a few additional projects (Edit 13/01/11: I changed the diagram above to show the 4th year project and one MSc contribution to my study). I find preparing charts like this useful to organise my thoughts around the process. I often struggle to see details but am OK with picturing the overall structure. When I create a flowchart like this, it forces me to think about the specific steps I’m going to need to take to move forward, as well as how all the pieces fit together.

In addition to the research, I also hope to present at 4 international, and 2 local conferences:

  • International Association for Medical Education
  • Education in a Changing Environment
  • Personal Learning Environments
  • World Physical Therapy Congress
  • South African Association for Health Educators
  • Higher Education Learning and Teaching Association of South Africa

I’ve committed to convert each conference presentation into a publication, so hopefully that’ll grease the wheels for funding from the university. It seems like quite a full programme, but since my teaching load has been reduced for this year, I think it’s doable.