Import AI 139: Making better healthcare AI systems via audio de-identification; teaching drones to help humans fight fires; and why language models could be smarter than you think

Stopping poachers with machine learning:
…Finally, an AI-infused surveillance & response system that (most) people approve of…
Researchers and staffers with the University of Southern California, Key Biodiversity Area Secretariat, World Wide Fund for Nature, Wildlife Conservation Society, and Uganda Wildlife Authority, have used machine learning to increase the effectiveness of rangers who try to stop poaching in areas of significant biodiversity.

The project saw USC researchers collaborate with rangers in areas vulnerable to poaching in Uganda and Cambodia, and involved analyzing historical records of incidents of poaching and building predictive models to attempt to predict where poachers might strike next.
  The results are encouraging: the researchers tested their system in Murchison Falls National Park (MFNP) in Uganda and the Srepok Wildlife Sanctuary (SWS) in Cambodia, following an earlier test of the system in the Queen Elizabeth National Park (QENP) in Uganda. In MFNP, using the system, “park rangers detected poaching activity in 38 cells,” they write. Additionally, “the amount of poaching activity observed is highest in the high-risk regions and lowest in the low-risk regions”, suggesting that the algorithm developed by the researchers is learning useful patterns. The researchers also deployed the model in the SWS park in Cambodia and again observed that regions the algorithm classified as high risk had higher incidences of poaching.

Impact: To get a sense of the impact of this technique, we can compare the results of the field tests to typical observations made by the rangers.
  In the SWS park in Cambodia during 2018, the average number of snares confiscated each month was 101. By comparison, during the month the rangers were using the machine learning system they removed 521 snares. “They also confiscated a firearm and they saw, but did not catch, three groups of poachers in the forest”.

Different parks, different data: As most applied scientists know, real world data tends to contain its own surprising intricacies. Here, this manifests as radically different distributions of types of datapoints across the different parks – in the MFNP park in Uganda around 15% of the collected datapoints are for areas where poaching is thought to have occurred, compared to around 4.3% for QENP and ~0.25% for the SWS park in Cambodia. Additionally, the makeup of each dataset changes over the year as the parks change various things that factor into the data collection, like the number of rangers, or the routes they cover, and so on.

Why this matters: Machine learning approaches are giving us the ability to build a sense&respond infrastructure for the world, and as we increase the sample efficiency and accuracy of such techniques we’ll be able to better build systems to help us manage an increasingly unstable planet. It’s deeply gratifying to see ML being used to protect biodiversity.
  Read more: Stay Ahead of Poachers: Illegal Wildlife Poaching Prediction and Patrol Planning Under Uncertainty with Field Test Evaluations (Arxiv).

#####################################################

Want to check-in for your flight in China? Use facial recognition:
I’ve known for some time that China has rolled out facial recognition systems in its airports to speed-up the check-in process. Now, Matthew Brennan has recorded a short video showing what this is like in practice from the perspective of a consumer. I imagine these systems will become standard worldwide within a few years.
  Note: I’ve seen a bunch of people on Twitter saying variations of “this technology makes me deeply uncomfortable and I hate it”. I think I have a slightly different opinion here. If you feel strongly about this would love to better understand your position, so please tweet at me or email me!
  Check out the facial recognition check-in here (Matthew Brennan, Twitter).

#####################################################

Why language models could be smarter than you think:
…LSTMs are smarter than you think, say researchers…
Researchers with the Cognitive Neuroimaging Unit at the ‘NeuroSpin Center’, as well as Facebook AI Research and the University of Amsterdam, have analyzed how LSTMs keep track of certain types of information, when tested for their ability to model language structures at longer timescales. The analysis is meant to explore “whether these generic sequence-processing devices are discovering genuine structural properties of language in their training data, or whether their success can be explained by opportunistic surface-pattern-based heuristics”, the authors write.

The researchers study a pre-trained model “composed of a 650-dimensional embedding layer, two 650-dimensional hidden layers, and an output layer with vocabulary size 50,000”. They evaluate this model on a set of what they call ‘number-agreement tasks’ where they test subject-verb agreement in increasingly challenging setups (eg, a simple case is looking at network activations for ‘the boy greets the guy’, with harder ones taking the form of things like ‘the boy most probably greets the guy’ and ‘the boy near the car kindly greets the guy’, and so on).

Neurons that count, sometimes together: During analysis, they noticed the LSTM had developed “two ‘grandmother’ cells to carry number features from the subject to the verb across the intervening material”. They found that these cells were sometimes used to help the network decide in particularly tricky cases: “The LSTM also possesses a more distributed mechanism to predict number when subject and verb are close, with the grandmother number cells only playing a crucial role in more difficult long-distance cases”.
  They also discovered a cell that “encodes the presence of an embedded phrase separating the main subject-verb dependency, and has strong efferent connections to the long-distance number cells, suggesting that the network relies on genuine syntactic information to regulate agreement-feature percolation”.

Why this matters: “Strikingly, simply training an LSTM on a language-model objective on raw corpus data brought about single units carrying exceptionally specific linguistic information. Three of these units were found to form a highly interactive local network, which makes up the central part of a ‘neural’ circuit performing long-distance number agreement”, they write. “Agreement in an LSTM language-model cannot be entirely explained away by superficial heuristics, and the networks have, to some extent, learned to build and exploit structure-based syntactic representations, akin to those conjectured to support human-sentence processing”
  The most interesting thing about all of this is the apparent sophistication that emerges as we train these networks. It seems to inherently support some of the ideas outlined by Richard Sutton (covered in Import AI #138) about the ability for relatively simple networks to – given sufficient compute – develop very sophisticated capabilities.
  Read more: The emergence of number and syntax units in LSTM language models (Arxiv).

#####################################################

Teaching drones to help humans fight fires:
…Towards a future where human firefighters are guarded by flying machines…
Researchers with the Georgia Institute of Technology have developed an algorithm to let humans and drones work together when fighting fires, with the drones now able to analyze the fire from above and relay that information to firefighters. “The proposed algorithm overcomes the limitations of prior work by explicitly estimating the latent fire propagation dynamics to enable intelligent, time-extended coordination of the UAVs in support of on-the-ground human firefighters,” they write.

The researchers use FARSITE, software for simulating wildfire propagation that is already widely used by the United States’ National Park Service and Forest Service. They use an Adaptive Extended Kalman Filter (AEKF) to make predictions about where the fire is likely to spread to. They eventually develop a basic system that works in simulation which lets them coordinate the actions of drones and humans, so that the drones learn to intelligently inform people about fire propagation. They also implement a “Human Safety Module” which attempts to work out how safe the people are, and how safe they will be as the fire location changes over time.

Three tests: They test the system on three scenarios: stationary fire, moving fire, and one where the fire moves and grows in area over time.The tests mostly tell us that you need to dramatically increase the number of drones to satisfy human safety guarantees (eg, in the case of a moving and growing fire you need over 25 drones to provide safety for 10 humans. Similarly, in this scenario you need at least 7 drones to be able to significantly reduce your level of uncertainty about the map of the fire and how it will change). Their approach outperforms prior state-of-the-art in this domain.

Why this matters: I think in as little as a decade it’s going to become common to see teams of humans and drones working together to deal with wildfires or building fires, and papers like this show how we might use adaptive software systems to increase the safety of human-emergency responders.
  Read more: Safe Coordination of Human-Robot Firefighting Teams (Arxiv).

#####################################################

Google wants it to be easier to erase personal information from audio logs:
…Audio de-ID metric & benchmark should spur research into cleaning up datasets…
Medical data is one of the most difficult types of data for machine learning researchers to work with, due to the fact it is full of personally identifiable health information. This means that people need to invest in tools to remove this personal information from medical data, so that people can build secondary applications on top of it. Coupled with this is the fact that in recent years more and more medical services are accessed digitally, leading to a growth in the amount of digital healthcare-relevant data, all of which needs to have the personal information removed before it can be processed by most machine learning approaches.

Now, researchers with Google are trying to work on this problem in the audio domain by releasing Audio de-ID, a new metric for measuring the success at de-identifying audio logs, and an evaluation benchmark for evaluating systems. The evaluation benchmark tests models against a Google-curated dataset made of the ‘Switchboard’ and ‘Fisher’ datasets with Personal Health Information (PHI) tagged and labelled in the data, and challenges models to automatically slice personally identifiable information out of datasets.

Taking personal information out of audio logs: Removing personally identifiable information from audio logs is a tricky task. Google’s pipeline works as follows: “Our pipeline first produces transcripts from the audio using [Audio Speech Recognition], proceeds by running text-based [Named Entity Recognition] tagging, and then redacts [Personal Health Information] tokens, using the aligned token boundaries determined by ASR. Our tagger relies on the state-of-the-art techniques for solving the audio NER problem of recognizing entities in audio transcripts. We leverage the available ASR technology, and use its component of alignment back to audio.

Why this matters: The (preliminary) benchmark results in the paper show that Audio Speech Recognition performance is “the main impedance towards achieving results comparable to text de-ID”, suggesting that as we develop more capable ASR models we will see improvements in our ability to automatically clean datasets so we can do more useful things with them.

Want the data? Google says you should be able to get it from this link referenced in the paper, but the link currently 404s (as of Sunday the 24th of March): https://g.co/audio-ner-annotations-data
   Read more: Audio De-identification: A New Entity Recognition Task (Arxiv).

#####################################################

It just got (a little bit) easier to develop & test AI algorithms on robot hardware:
…Gym-gazebo2 lets you simulate robots and provides an OpenAI Gym-like interface…
Researchers with Acutronic Robotics have released gym-gazebo2, software that people can use to develop and compare reinforcement learning algorithms’ performance on robots simulated within the ‘ROS 2’ and ‘Gazebo’ software platforms. Gym-gazebo2 contains simulation tools, middleware software, and a high-fidelity simulation of the ‘MARA’ robot which is a product developed by Acutronic Robotics.

Gym-Gazebo2 ingredients: The software consists of three main chunks of software: ROS 2, Gazebo, and Gym-Gazebo2. Gym-Gazebo2 can be installed via a docker container, which should simplify setup.

MARA, MARA, MARA: Initially, Gym-Gazebo2 ships with four environments based around Acutronic Robotics’s Modular Articulated Robotic Arm (MARA) system. These environments are:

  • MARA: The agent is rewarded for moving its gripper to a target position.
  • MARA Orient: The agent is rewarded for moving its gripper to a target position and specific orientation.
  • MARA Collision: The agent is rewarded in the same way as MARA, but is punished if the robot collides with anything.
  • MARA Collision Orient: The agent is rewared in the same way as MARA Orient, but is punished if it collides with anything.

  (It’s worth noting that these environments are very, very simple: real world robot tasks tend to involve more multi-step scenarios, usually with additional constraints. However, these environments could prove to be useful for validating performance of a given approach early in development.)

Free algorithms: Alongside gym-gazebo2, Acutronic Robotics is also releasing ros2learn, a collection of algorithms (based on OpenAI Baselines), as well as some pre-built experimental scripts for running algorithms like PPO, TRP, and ACKTR on the ‘MARA’ robot platform.

Why this matters: Robotics is about to be revolutionized by AI; following years of development by researchers around the world, deep learning techniques are maturing to the point that they can be applied to robots to solve tasks that were previously impossible or very, very, very expensive and complicated to program. Software like gym-gazebo2 will make it easier for people to validate algorithms in a high-fidelity simulation, and if people happen to like the MARA arm itself can also let them validate stuff in reality (though this tends to be expensive and fraught with various kinds of confounding pain, so it’ll depend on the popularity of the MARA hardware platform).
  Get the ‘ros2learn’ code from GitHub here.
  Get the code for gym-gazebo2 from GitHub here.
  Read more: gym-gazebo2, a toolkit for reinforcement learning using ROS 2 and Gazebo (Arxiv).

#####################################################

AI Policy with Matthew van der Merwe:
…Matthew van der Merwe has kindly offered to write some sections about AI & Policy for Import AI. I’m (lightly) editing them. All credit to Matthew, all blame to me, etc. Feedback: jack@jack-clark.net

Understanding the European Union’s AI ambitions:
In discussions about global leadership in AI, the EU is rarely considered a serious contender. This report by Charlotte Stix at the Leverhulme Center for the Future of Intelligence at the University of Cambridge, explores this view, and provides a detailed picture of the EU’s AI landscape and how this could form the basis for leadership in AI.

What the EU wants: The EU’s strategy is set out in 2 key documents: (1) The European Commission’s Communication on AI, which – among other things – outlines the EU’s ambition to put ‘ethical AI’ at the core of their strategy; (2) The Coordinated Plan on AI, which outlines how member states could coordinate their strategies and resources to improve European competitiveness in AI.

What the EU has: The EU’s central funding commitments for AI R&D remain modest – €1.5bn between 2018-2020. This is fairly trivial when compared with government funding in China (tens of billions of dollars per year), and even private companies (Alphabet’s 2018 R&D spend was $21bn). The Coordinated Plan includes an ambition to reach €20bn total funding per year by 2020, from member states and the private sector, though there are not yet concrete plans for how to achieve this. At present, inbound VC investment in AI is 6x lower than the US. The EU is having problems retaining talent and organizations, with promising companies increasingly acquired by international companies. There are proposals for the EU to establish large-scale projects in AI research, leveraging their impressive track-record in this domain (e.g. CERN, the Human Brain Project), but these remain relatively early-stage.

Why it matters: The EU is making a bet on leadership in ethical and human-centric AI, as a way to play an important role in the global AI landscape. This is an interesting strategy which sets it apart from other actors, e.g. US and China, who are focussed more on leadership in AI capabilities. The EU has expressed ambitions to cooperate closely with other countries and groups that are aligned with their vision (e.g. via the new International Panel on AI), which is encouraging for global coordination on these issues. Being made up of 27/28 member states, the EU’s experience in coordinating between actors could prove an advantage in this regard.
  Read more: Survey of the EU’s AI ecosystem (Charlotte Stix)  

#####################################################

New AI institute at Stanford launches + Bill Gates on AI risk
The Stanford Institute for Human-Centered AI (HAI) has formally launched. HAI is led by Fei-Fei Li, former director of Stanford AI lab and co-founder of AI4ALL, and John Etchemendy, philosopher and former Stanford provost.

The mission: HAI’s mission is informed by three overarching principles – that we must understand and forecast AI’s human impact, and guide its development on this basis; that AI should augment, rather than replace, humans; and that we must develop (artificial) intelligence as “subtle and nuanced as human intelligence.” The institute aims to “bring humanities and social thinking into tech,” to address emerging issues in AI.

Bill Gates on AI risk: At HAI’s launch event, Gates used his keynote speech to describe his hopes and fears for advanced AI. He said “the world hasn’t had that many technologies that are both promising and dangerous”, comparing AI to nuclear fission, which yielded both benefits in terms of energy, and serious risks from nuclear weapons. Gates had previously expressed some scepticism about AI risk, and this speech suggests he has moved towards a more cautious outlook.
  Read more: Launch announcement (Stanford).
  Read more: Bill Gates – AI is like nuclear weapons and nuclear energy in danger and promise (Vox).

#####################################################

Tech Tales:

Beaming In

We had been climbing the mountain for half an hour, and already a third of the people had cut their feeds and faded out of the country – they had discussed enough to come to a decision, and had disappeared to do the work. I stayed, looking at butterflies playing near cliff edges, and at birds diving between different trees, while talking about moving vast amounts of money for the purposes of buttressing a particular ideology and crushing another. The view was clear today, and I could see in my overlay the names of each of the towns and mountains and train stations on the horizon.

As I walked, I talked with the others. I suppose similar hikes have happened throughout history – more than a thousand years ago, a group of Icelandic elders hiked to a desolate riverbank in a rift valley where they formed the world’s oldest democracy, the ‘Althing’. During various world wars many leaders have gathered at various country parks and exchanged ideas with one another, making increasingly strange and influential plans. And of course there are the modern variants: Bilderberg hikes in various thinly-disclosed locations around Europe. Perhaps people use nature and walking within it as a kind of stimulant, to help them make decisions of consequence? Perhaps, they use nature because they have always used nature?

Now, on this walk, after another hour, most of the people have faded out, till it’s just me, and two others. One of them runs a multi-national with interests in intelligence, and the other works on behalf of both the US government and Chinese government on ‘track two’ diplomacy initiatives. We discuss things I cannot talk about, and make plans that will influence many people.

—–

But if you were to look at us from a distance, you wouldn’t see anybody at all. You’d see a mass of drones, hauling around the smell, texture and sense equipment which is needed to render high-fidelity simulations and transmit them to each of us, dialing in from our home offices and yachts and (in a few, rare cases) bunkers and jets. We use our machines to synthesize a beautiful reality for us, and we ‘walk’ within it, together making plans, appearing real to eachother, but appearing to people in the places that we walk as a collection of whirring metal. I am not sure when the last walk occurred with this group where a human decision-maker attended.

We do this because we believe in nature, and we believe that because we are in the presence of nature, our decisions are continuous with those of the past. I think habit plays a part, but so does a desire for absolution, and the knowledge that though trees and butterflies and crows may judge us for what we do, they cannot talk back.

Things that inspired this story: The general tradition of walking through history with particular emphasis on modern walks such as those taken by the Bilderberg group; virtual reality and augmented reality; increasingly capable drones; 5G; force feedback systems; the tendency for those in power to form increasingly closed-off nested communities.