In this blog post we introduce Population Based Augmentation (PBA), an algorithm that quickly and efficiently learns a state-of-the-art approach to augmenting data for neural network training. PBA matches the previous best result on CIFAR and SVHN but uses one thousand times less compute, enabling researchers and practitioners to effectively learn new augmentation policies using a single workstation GPU. You can use PBA broadly to improve deep learning performance on image recognition tasks.

We discuss the PBA results from our recent paper and then show how to easily run PBA for yourself on a new data set in the Tune framework.

Why should you care about data augmentation?

Recent advances in deep learning models have been largely attributed to the quantity and diversity of data gathered in recent years. Data augmentation is a strategy that enables practitioners to significantly increase the diversity of data available for training models, without actually collecting new data. Data augmentation techniques such as cropping, padding, and horizontal flipping are commonly used to train large neural networks. However, most approaches used in training neural networks only use basic types of augmentation. While neural network architectures have been investigated in depth, less focus has been put into discovering strong types of data augmentation and data augmentation policies that capture data invariances.

An image of the number “3” in original form and with basic augmentations applied.

Recently, Google has been able to push the state-of-the-art accuracy on datasets such as CIFAR-10 with AutoAugment, a new automated data augmentation technique. AutoAugment has shown that prior work using just applying a fixed set of transformations like horizontal flipping or padding and cropping leaves potential performance on the table. AutoAugment introduces 16 geometric and color-based transformations, and formulates an augmentation policy that selects up to two transformations at certain magnitude levels to apply to each batch of data. These higher performing augmentation policies are learned by training models directly on the data using reinforcement learning.

What’s the catch?

AutoAugment is a very expensive algorithm which requires training 15,000 models to convergence to generate enough samples for a reinforcement learning based policy. No computation is shared between samples, and it costs 15,000 NVIDIA Tesla P100 GPU hours to learn an ImageNet augmentation policy and 5,000 GPU hours to learn an CIFAR-10 one. For example, if using Google Cloud on-demand P100 GPUs, it would cost about \$7,500 to discover a CIFAR policy, and \$37,500 to discover an ImageNet one! Therefore, a more common use case when training on a new dataset would be to transfer a pre-existing published policy, which the authors show works relatively well.

Population Based Augmentation

Our formulation of data augmentation policy search, Population Based Augmentation (PBA), reaches similar levels of test accuracy on a variety of neural network models while utilizing three orders of magnitude less compute. We learn an augmentation policy by training several copies of a small model on CIFAR-10 data, which takes five hours using a NVIDIA Titan XP GPU. This policy exhibits strong performance when used for training from scratch on larger model architectures and with CIFAR-100 data.

Relative to the several days it takes to train large CIFAR-10 networks to convergence, the cost of running PBA beforehand is marginal and significantly enhances results. For example, training a PyramidNet model on CIFAR-10 takes over 7 days on a NVIDIA V100 GPU, so learning a PBA policy adds only 2% precompute training time overhead. This overhead would be even lower, under 1%, for SVHN.

CIFAR-10 test set error between PBA, AutoAugment, and the baseline which only uses horizontal flipping, padding, and cropping, on WideResNet, Shake-Shake, and PyramidNet+ShakeDrop models. PBA is significantly better than the baseline and on-par with AutoAugment.

PBA leverages the Population Based Training algorithm to generate an augmentation policy schedule which can adapt based on the current epoch of training. This is in contrast to a fixed augmentation policy that applies the same transformations independent of the current epoch number.

This allows an ordinary workstation user to easily experiment with the search algorithm and augmentation operations. One interesting use case would be to introduce new augmentation operations, perhaps targeted towards a particular dataset or image modality, and be able to quickly produce a tailored, high performing augmentation schedule. Through ablation studies, we have found that the learned hyperparameters and schedule order are important for good results.

How is the augmentation schedule learned?

We use Population Based Training with a population of 16 small WideResNet models. Each worker in the population will learn a different candidate hyperparameter schedule. We transfer the best performing schedule to train larger models from scratch, from which we derive our test error metrics.

Overview of Population Based Training, which discovers hyperparameter schedules by training a population of neural networks. It combines random search (explore) with the copying of model weights from high performing workers (exploit). Source

The population models are trained on the target dataset of interest starting with all augmentation hyperparameters set to 0 (no augmentations applied). At frequent intervals, an “exploit-and-explore” process “exploits” high performing workers by copying their model weights to low performing workers, and then “explores” by perturbing the hyperparameters of the worker. Through this process, we are able to share compute heavily between the workers and target different augmentation hyperparameters at different regions of training. Thus, PBA is able to avoid the cost of training thousands of models to convergence in order to reach high performance.

Example and Code

We leverage Tune’s built-in implementation of PBT to make it straightforward to use PBA.

import ray
def explore(config):
    """Custom PBA function to perturb augmentation hyperparameters."""
    ...

ray.init()
pbt = ray.tune.schedulers.PopulationBasedTraining(
    time_attr="training_iteration",
    reward_attr="val_acc",
    perturbation_interval=3,
    custom_explore_fn=explore)
train_spec = {...}  # Things like file paths, model func, compute.
ray.tune.run_experiments({"PBA": train_spec}, scheduler=pbt)

We call Tune’s implementation of PBT with our custom exploration function. This will create 16 copies of our WideResNet model and train them time-multiplexed. The policy schedule used by each copy is saved to disk and can be retrieved after termination to use for training new models.

You can run PBA by following the README at: https://github.com/arcelien/pba. On a Titan XP, it only requires one hour to learn a high performing augmentation policy schedule on the SVHN dataset. It is also easy to use PBA on a custom dataset as well: simply define a new dataloader and everything else falls into place.

Big thanks to Daniel Rothchild, Ashwinee Panda, Aniruddha Nrusimha, Daniel Seita, Joseph Gonzalez, and Ion Stoica for helpful feedback while writing this post. Feel free to get in touch with us on Github!

This post is based on the following paper to appear in ICML 2019 as an oral presentation:

• Population Based Augmentation: Efficient Learning of Augmentation Policy Schedules
  Daniel Ho, Eric Liang, Ion Stoica, Pieter Abbeel, Xi Chen
  Paper Code

MIT researchers have developed a novel “photonic” chip that uses light instead of electricity — and consumes relatively little power in the process. The chip could be used to process massive neural networks millions of times more efficiently than today’s classical computers do.

Neural networks are machine-learning models that are widely used for such tasks as robotic object identification, natural language processing, drug development, medical imaging, and powering driverless cars. Novel optical neural networks, which use optical phenomena to accelerate computation, can run much faster and more efficiently than their electrical counterparts.  

But as traditional and optical neural networks grow more complex, they eat up tons of power. To tackle that issue, researchers and major tech companies — including Google, IBM, and Tesla — have developed “AI accelerators,” specialized chips that improve the speed and efficiency of training and testing neural networks.

For electrical chips, including most AI accelerators, there is a theoretical minimum limit for energy consumption. Recently, MIT researchers have started developing photonic accelerators for optical neural networks. These chips perform orders of magnitude more efficiently, but they rely on some bulky optical components that limit their use to relatively small neural networks.

In a paper published in Physical Review X, MIT researchers describe a new photonic accelerator that uses more compact optical components and optical signal-processing techniques, to drastically reduce both power consumption and chip area. That allows the chip to scale to neural networks several orders of magnitude larger than its counterparts.

Simulated training of neural networks on the MNIST image-classification dataset suggest the accelerator can theoretically process neural networks more than 10 million times below the energy-consumption limit of traditional electrical-based accelerators and about 1,000 times below the limit of photonic accelerators. The researchers are now working on a prototype chip to experimentally prove the results.

“People are looking for technology that can compute beyond the fundamental limits of energy consumption,” says Ryan Hamerly, a postdoc in the Research Laboratory of Electronics. “Photonic accelerators are promising … but our motivation is to build a [photonic accelerator] that can scale up to large neural networks.”

Practical applications for such technologies include reducing energy consumption in data centers. “There’s a growing demand for data centers for running large neural networks, and it’s becoming increasingly computationally intractable as the demand grows,” says co-author Alexander Sludds, a graduate student in the Research Laboratory of Electronics. The aim is “to meet computational demand with neural network hardware … to address the bottleneck of energy consumption and latency.”

Joining Sludds and Hamerly on the paper are: co-author Liane Bernstein, an RLE graduate student; Marin Soljacic, an MIT professor of physics; and Dirk Englund, an MIT associate professor of electrical engineering and computer science, a researcher in RLE, and head of the Quantum Photonics Laboratory.  

Compact design

Neural networks process data through many computational layers containing interconnected nodes, called “neurons,” to find patterns in the data. Neurons receive input from their upstream neighbors and compute an output signal that is sent to neurons further downstream. Each input is also assigned a “weight,” a value based on its relative importance to all other inputs. As the data propagate “deeper” through layers, the network learns progressively more complex information. In the end, an output layer generates a prediction based on the calculations throughout the layers.

All AI accelerators aim to reduce the energy needed to process and move around data during a specific linear algebra step in neural networks, called “matrix multiplication.” There, neurons and weights are encoded into separate tables of rows and columns and then combined to calculate the outputs.

In traditional photonic accelerators, pulsed lasers encoded with information about each neuron in a layer flow into waveguides and through beam splitters. The resulting optical signals are fed into a grid of square optical components, called “Mach-Zehnder interferometers,” which are programmed to perform matrix multiplication. The interferometers, which are encoded with information about each weight, use signal-interference techniques that process the optical signals and weight values to compute an output for each neuron. But there’s a scaling issue: For each neuron there must be one waveguide and, for each weight, there must be one interferometer. Because the number of weights squares with the number of neurons, those interferometers take up a lot of real estate.

“You quickly realize the number of input neurons can never be larger than 100 or so, because you can’t fit that many components on the chip,” Hamerly says. “If your photonic accelerator can’t process more than 100 neurons per layer, then it makes it difficult to implement large neural networks into that architecture.”

The researchers’ chip relies on a more compact, energy efficient “optoelectronic” scheme that encodes data with optical signals, but uses “balanced homodyne detection” for matrix multiplication. That’s a technique that produces a measurable electrical signal after calculating the product of the amplitudes (wave heights) of two optical signals.

Pulses of light encoded with information about the input and output neurons for each neural network layer — which are needed to train the network — flow through a single channel. Separate pulses encoded with information of entire rows of weights in the matrix multiplication table flow through separate channels. Optical signals carrying the neuron and weight data fan out to grid of homodyne photodetectors. The photodetectors use the amplitude of the signals to compute an output value for each neuron. Each detector feeds an electrical output signal for each neuron into a modulator, which converts the signal back into a light pulse. That optical signal becomes the input for the next layer, and so on.

The design requires only one channel per input and output neuron, and only as many homodyne photodetectors as there are neurons, not weights. Because there are always far fewer neurons than weights, this saves significant space, so the chip is able to scale to neural networks with more than a million neurons per layer.

Finding the sweet spot

With photonic accelerators, there’s an unavoidable noise in the signal. The more light that’s fed into the chip, the less noise and greater the accuracy — but that gets to be pretty inefficient. Less input light increases efficiency but negatively impacts the neural network’s performance. But there’s a “sweet spot,” Bernstein says, that uses minimum optical power while maintaining accuracy.

That sweet spot for AI accelerators is measured in how many joules it takes to perform a single operation of multiplying two numbers — such as during matrix multiplication. Right now, traditional accelerators are measured in picojoules, or one-trillionth of a joule. Photonic accelerators measure in attojoules, which is a million times more efficient.

In their simulations, the researchers found their photonic accelerator could operate with sub-attojoule efficiency. “There’s some minimum optical power you can send in, before losing accuracy. The fundamental limit of our chip is a lot lower than traditional accelerators … and lower than other photonic accelerators,” Bernstein says.

In preparation for a recent meeting of the WEF global AI council, we were asked the question:

What do you think are the top three policy and governance issues that face AI/ML currently?

Here are my answers.

1. For me the biggest governance issue facing AI/ML ethics is the gap between principles and practice. The hard problem the industry faces is turning good intentions into demonstrably good behaviour. In the last 2.5 years there has been a gold rush of new ethical principles in AI. Since Jan 2017 at least 22 sets of ethical principles have been published, including principles from Google, IBM, Microsoft and Intel. Yet any evidence that these principles are making a difference within those companies is hard to find – leading to a justifiable accusation of ethics-washing – and if anything the reputations of some leading AI companies are looking increasingly tarnished.

2. Like others I am deeply concerned by the acute gender imbalance in AI (estimates of the proportion of women in AI vary between ~12% and ~22%). This is not just unfair, I believe it too be positively dangerous, since it is resulting in AI products and services that reflect the values and ambitions of (young, predominantly white) men. This makes it a governance issue. I cannot help wondering if the deeply troubling rise of surveillance capitalism is not, at least in part, a consequence of male values.

3. A major policy concern is the apparently very poor quality of many of the jobs created by the large AI/ML companies. Of course the AI/ML engineers are paid exceptionally well, but it seems that there is a very large number of very poorly paid workers who, in effect, compensate for the fact that AI is not (yet) capable of identifying offensive content, nor is it able to learn without training data generated from large quantities of manually tagged objects in images, nor can conversational AI manage all queries that might be presented to it. This hidden army of piece workers, employed in developing countries by third party sub contractors and paid very poorly, are undertaking work that is at best extremely tedious (you might say robotic) and at worst psychologically very harmful; this has been called AI’s dirty little secret and should not – in my view – go unaddressed.

The city of Amsterdam envisions a future where fleets of autonomous boats cruise its many canals to transport goods and people, collect trash, or self-assemble into floating stages and bridges. To further that vision, MIT researchers have given new capabilities to their fleet of robotic boats — which are being developed as part of an ongoing project — that lets them target and clasp onto each other, and keep trying if they fail.

About a quarter of Amsterdam’s surface area is water, with 165 canals winding alongside busy city streets. Several years ago, MIT and the Amsterdam Institute for Advanced Metropolitan Solutions (AMS Institute) teamed up on the “Roboat” project. The idea is to build a fleet of autonomous robotic boats — rectangular hulls equipped with sensors, thrusters, microcontrollers, GPS modules, cameras, and other hardware — that provides intelligent mobility on water to relieve congestion in the city’s busy streets.

One of project’s objectives is to create roboat units that provide on-demand transporation on waterways. Another objective is using the roboat units to automatically form “pop-up” structures, such as foot bridges, performance stages, or even food markets. The structures could then automatically disassemble at set times and reform into target structures for different activities. Additionally, the roboat units could be used as agile sensors to gather data on the city’s infrastructure, and air and water quality, among other things.

In 2016, MIT researchers tested a roboat prototype that cruised around Amsterdam’s canals, moving forward, backward, and laterally along a preprogrammed path. Last year, researchers designed low-cost, 3-D-printed, one-quarter scale versions of the boats, which were more efficient and agile, and came equipped with advanced trajectory-tracking algorithms. 

In a paper presented at the International Conference on Robotics and Automation, the researchers describe roboat units that can now identify and connect to docking stations. Control algorithms guide the roboats to the target, where they automatically connect to a customized latching mechanism with millimeter precision. Moreover, the roboat notices if it has missed the connection, backs up, and tries again.

The researchers tested the latching technique in a swimming pool at MIT and in the Charles River, where waters are rougher. In both instances, the roboat units were usually able to successfully connect in about 10 seconds, starting from around 1 meter away, or they succeeded after a few failed attempts. In Amsterdam, the system could be especially useful for overnight garbage collection. Roboat units could sail around a canal, locate and latch onto platforms holding trash containers, and haul them back to collection facilities.

“In Amsterdam, canals were once used for transportation and other things the roads are now used for. Roads near canals are now very congested — and have noise and pollution — so the city wants to add more functionality back to the canals,” says first author Luis Mateos, a graduate student in the Department of Urban Studies and Planning (DUSP) and a researcher in the MIT Senseable City Lab. “Self-driving technologies can save time, costs and energy, and improve the city moving forward.”

“The aim is to use roboat units to bring new capabilities to life on the water,” adds co-author Daniela Rus, director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science. “The new latching mechanism is very important for creating pop-up structures. Roboat does not need latching for autonomous transporation on water, but you need the latching to create any structure, whether it’s mobile or fixed.”

Joining Mateos on the paper are: Wei Wang, a joint postdoc in CSAIL and the Senseable City Lab; Banti Gheneti, a graduate student in the Department of Electrical Engineering and Computer Science; Fabio Duarte, a DUSP and Senseable City Lab research scientist; and Carlo Ratti, director of the Senseable City Lab and a principal investigator and professor of the practice in DUSP.

Making the connection

Each roboat is equipped with latching mechanisms, including ball and socket components, on its front, back, and sides. The ball component resembles a badminton shuttlecock — a cone-shaped, rubber body with a metal ball at the end. The socket component is a wide funnel that guides the ball component into a receptor. Inside the funnel, a laser beam acts like a security system that detects when the ball crosses into the receptor. That activates a mechanism with three arms that closes around and captures the ball, while also sending a feedback signal to both roboats that the connection is complete.

On the software side, the roboats run on custom computer vision and control techniques. Each roboat has a LIDAR system and camera, so they can autonomously move from point to point around the canals. Each docking station — typically an unmoving roboat — has a sheet of paper imprinted with an augmented reality tag, called an AprilTag, which resembles a simplified QR code. Commonly used for robotic applications, AprilTags enable robots to detect and compute their precise 3-D position and orientation relative to the tag.

Both the AprilTags and cameras are located in the same locations in center of the roboats. When a traveling roboat is roughly one or two meters away from the stationary AprilTag, the roboat calculates its position and orientation to the tag. Typically, this would generate a 3-D map for boat motion, including roll, pitch, and yaw (left and right). But an algorithm strips away everything except yaw. This produces an easy-to-compute 2-D plane that measures the roboat camera’s distance away and distance left and right of the tag. Using that information, the roboat steers itself toward the tag. By keeping the camera and tag perfectly aligned, the roboat is able to precisely connect.

The funnel compensates for any misalignment in the roboat’s pitch (rocking up and down) and heave (vertical up and down), as canal waves are relatively small. If, however, the roboat goes beyond its calculated distance, and doesn’t receive a feedback signal from the laser beam, it knows it has missed. “In challenging waters, sometimes roboat units at the current one-quarter scale, are not strong enough to overcome wind gusts or heavy water currents,” Mateos says. “A logic component on the roboat says, ‘You missed, so back up, recalculate your position, and try again.’”

Future iterations

The researchers are now designing roboat units roughly four times the size of the current iterations, so they’ll be more stable on water. Mateos is also working on an update to the funnel that includes tentacle-like rubber grippers that tighten around the pin — like a squid grasping its prey. That could help give the roboat units more control when, say, they’re towing platforms or other roboats through narrow canals.

In the works is also a system that displays the AprilTags on an LCD monitor that changes codes to signal multiple roboat units to assemble in a given order. At first, all roboat units will be given a code to stay exactly a meter apart. Then, the code changes to direct the first roboat to latch. After, the screen switches codes to order the next roboat to latch, and so on. “It’s like the telephone game. The changing code passes a message to one roboat at a time, and that message tells them what to do,” Mateos says.

Darwin Caldwell, the research director of Advanced Robotics at the Italian Institute of Technology, envisions even more possible applications for the autonomous latching capability. “I can certainly see this type of autonomous docking being of use in many areas of robotic ‘refuelling’ and docking … beyond aquatic/naval systems,” he says, “including inflight refuelling, space docking, cargo container handling, [and] robot in-house recharging.”

The research was funded by the AMS Institute and the City of Amsterdam.

By Eugene Vinitsky

We are in the midst of an unprecedented convergence of two rapidly growing trends on our roadways: sharply increasing congestion and the deployment of autonomous vehicles. Year after year, highways get slower and slower: famously, China’s roadways were paralyzed by a two-week long traffic jam in 2010. At the same time as congestion worsens, hundreds of thousands of semi-autonomous vehicles (AVs), which are vehicles with automated distance and lane-keeping capabilities, are being deployed on highways worldwide. The second trend offers a perfect opportunity to alleviate the first. The current generation of AVs, while very far from full autonomy, already hold a multitude of advantages over human drivers that make them perfectly poised to tackle this congestion. Humans are imperfect drivers: accelerating when we shouldn’t, braking aggressively, and make short-sighted decisions, all of which creates and amplifies patterns of congestion.

On the other hand, AVs are free of these constraints: they have low reaction times, can potentially coordinate over long distances, and most importantly, companies can simply modify their braking and acceleration patterns in ways that are congestion reducing. Even though only a small percentage of vehicles are currently semi-autonomous, existing research indicates that even a small penetration rate, 3-4%, is sufficient to begin easing congestion. The essential question is: will we capture the potential gains, or will AVs simply reproduce and further the growing gridlock?

Given the unique capabilities of AVs, we want to ensure that their driving patterns are designed for maximum impact on roadways. The proper deployment of AVs should minimize gridlock, decrease total energy consumption, and maximize the capacity of our roadways. While there have been decades of research on these questions, there isn’t an existing consensus on the optimal driving strategies to employ, nor easy metrics by which a self-driving car company could assess a driving strategy and then choose to implement it in their own vehicles. We postulate that a partial reason for this gap is the absence of benchmarks: standardized problems which we can use to compare progress across research groups and methods. With properly designed benchmarks we can examine an AV’s driving behavior and quickly assign it a score, ensuring that the best AV designs are the ones to make it out onto the roadways. Furthermore, benchmarks should facilitate research, by making it easy for researchers to rapidly try out new techniques and algorithms and see how they do at resolving congestion.

In an attempt to fill this gap, our CORL paper proposes 11 new benchmarks in centralized mixed-autonomy traffic control: traffic control where a small fraction of the vehicles and traffic lights are controlled by a single computer. We’ve released these benchmarks as a part of Flow, a tool we’ve developed for applying control and reinforcement learning (via using RLlib and rllab as the reinforcement learning libraries) to autonomous vehicles and traffic lights in the traffic simulators SUMO and AIMSUN. A high score in these benchmarks means an improvement in real-world congestion metrics such as average speed, total system delay, and roadway throughput. By making progress on these benchmarks, we hope to answer fundamental questions about AV usage and provide a roadmap for deploying congestion improving AVs in the real world.

The benchmark scenarios, depicted at the top of this post, cover the following settings:

• A simple figure eight, representing a toy intersection, in which the optimal solution is either a snaking behavior or learning to alternate which direction is moving without conflict.

• A resizable grid of traffic lights where the goal is to optimize the light patterns to minimize the average travel time.

• An on-ramp merge in which a vehicle aggressive merging onto the main highway causes a shockwave that lowers the average speed of the system.

• A toy model of the San-Francisco to Oakland Bay Bridge where four lanes merge to two and then to one. The goal is to prevent congestion from forming so to maximize the number of exiting vehicles.

As an example of an exciting and helpful emergent behavior that was discovered in these benchmarks, the following GIF shows a segment of the bottleneck scenario in which the four lanes merge down to two, with a two-to-one bottleneck further downstream that is not shown. In the top, we have the fully human case in orange. The human drivers enter the four-to-two bottleneck at an unrestricted rate, which leads to congestion at the two-to-one bottleneck and subsequent congestion that slows down the whole system. In the bottom video, there is a mix of human drivers (orange) and autonomous vehicles (red). We find that the autonomous vehicles learn to control the rate at which vehicles are entering the two-to-one bottleneck and they accelerate to help the vehicles behind them merge smoothly. Despite only one in ten vehicles being autonomous, the system is able to remain uncongested and there is a 35% improvement in the throughput of the system.

Once we formulated and coded up the benchmarks, we wanted to make sure that researchers had a baseline set of values to check their algorithms against. We performed a small hyperparameter sweep and then ran the best hyperparameters for the following RL algorithms: Augmented Random Search, Proximal Policy Optimization, Evolution Strategies, and Trust Region Policy Optimization. The top graphs indicate baseline scores against a set of proxy rewards that are used during training time. Each graph corresponds to a scenario and the scores the algorithms achieved as a function of training time. These should make working with the benchmarks easier as you’ll know immediately if you’re on the right track based on whether your score is above or below these values.

From an impact on congestion perspective however, the graph that really matters is the one at the bottom, where we score the algorithms according to the metrics that genuinely affect congestion. These metrics are: average speed for the Figure Eight and Merge, average delay per vehicle for the Grid, and total outflow in vehicles per hour for the bottleneck. The first four columns are the algorithms graded according to these metrics and in the last column we list the results of a fully human baseline. Note that all of these benchmarks are at relatively low AV penetration rates, ranging from 7% at the lowest to 25% at the highest (i.e. ranging from 1 AV in every 14 vehicles to 1 AV in every 4). The congestion metrics in the fully human column are all sharply worse, suggesting that even at very low penetration rates, AVs can have an incredible impact on congestion.

So how do the AVs actually work to ease congestion? As an example of one possible mechanism, the video below compares an on-ramp merge for a fully human case (top) and the case where one in every ten drivers is autonomous (red) and nine in ten are human (white). In both cases, a human driver is attempting to aggressively merge onto the ramp with little concern for the vehicles on the main road. In the fully human case, the vehicles are packed closely together, and when a human driver sharply merges on, the cars behind need to brake quickly, leading to “bunching”. However, in the case with AVs, the autonomous vehicle accelerates with the intent of opening up larger gaps between the vehicles as they approach the on-ramp. The larger spaces create a buffer zone, so that when the on-ramp vehicle merges, the vehicles on the main portion of the highway can brake more gently.

There is still a lot of work to be done; while we’re unable to prove it mathematically, we’re fairly certain that none of our results achieve the optimal top scores and the full paper provides some arguments suggesting that we’ve just found local minima.

There’s a large set of totally untackled questions as well. For one, these benchmarks are for the fully centralized case, when all the cars are controlled by one central computer. Any real road driving policy would likely have to be decentralized: can we decentralize the system without decreasing performance? There are also notions of fairness that aren’t discussed. As the video below shows, bottleneck outflow can be significantly improved by fully blocking a lane; while this driving pattern is efficient, it severely penalizes some drivers while rewarding others, invariably leading to road rage. Finally, there is the fascinating question of generalization. It seems difficult to deploy a separate driving behavior for every unique driving scenario; is it possible to find one single controller that works across different types of transportation networks? We aim to address all of these questions in a future set of benchmarks.

If you’re interested in contributing to these new benchmarks, trying to beat our old benchmarks, or working towards improving the mixed-autonomy future, get in touch via our GitHub page or our website!

Thanks to Jonathan Liu, Prastuti Singh, Yashar Farid, and Richard Liaw for edits and discussions. Thanks to Aboudy Kriedieh for helping prepare some of the videos. This article was initially published on the BAIR blog, and appears here with the authors’ permission.

Europe is gearing up to launch an Artificial Intelligence Public Private Partnership (AI PPP) that brings together AI, data, and robotics. At its core is a drive to lead the world in the development and deployment of trustworthy AI based on EU fundamental rights, principles and values.

The effort is being led by two well-established associations representing over 400 European organisations from Industry and Research: the Big Data Value Association and euRobotics. A first step in this process saw the launch of a consultation document in Brussels last week entitled “Strategic Research, Innovation and Deployment Agenda for an AI PPP”.

The opportunity for Europe

The strategy document comes on the backdrop of international competition, with every country vying to take the lead in AI.

Roberto Viola, Director-General of the European Commission, DG CONNECT, linked it to the recently disputed European Champions League match between Tottenham and Liverpool, which saw an unexpected roster of teams make the final. “Are you sure that China and the US will lead AI? Watch our planes, watch our industrial robots, watch our cars – why are we so shy about our successes? I’ve been around the world, and I’m always asked how Europe can work with other countries in AI. I don’t think we’ve lost the race, in fact I don’t think it’s a race. If it’s a race, it’s about delivering good services in AI to Europeans. We can be in the final of the Champions league.”

Viola says harmonious cross-fertilisation across three dimensions is needed to make AI a success in Europe. First, the EU needs to mainstream AI in society and the economy. Companies have to find their place, and there is a big role for the public to support AI and its introduction in real scenarios. This is the part missing in the EU, compared to US and China. There is a need for more big public procurement. Second, there is a need to push for research and development in AI through European funding programmes, and third, policy is needed to accompany the development of AI in society.

Photo: Roberto Viola, Director-General of the European Commission, DG CONNECT; Credits: Octavian Carare

Trustworthy AI

Viola says “Europe was one of the first to develop ethical codes in AI, and now everyone is doing it. It shows that we know what is important and the impact and relevance of AI for society.”  Responsible AI was the leitmotif of the day, with everyone highlighting it as Europe’s unique selling point. The message was clear, Europe can do excellent AI, but is leading in terms of deploying it with society in mind. Juha Heikkilä, Head of the Robotics and Artificial Intelligence Unit at the European Commission says “Europe is doing things in a different way – that takes the citizen into account.”

AI Market Opportunity

The combination of big data, advances in algorithms, computing power, and advanced robotics is opening up new market opportunities for AI enabled systems.

Photo: David Bisset, Director – euRobotics, Sonja Zillner, Siemens AG; Credits: Octavian Carare

Sonja Zillner from Siemens, Co-Chief Editor for the strategy, says “the vision is to boost EU industrial competitiveness and lead the world in development and deployment value-driven trustworthy AI”. When asked for their preferred applications for the development of AI, the public requested improved healthcare services, energy efficiency, and availability of trains, as well as increased productivity in digital factories. This led to the realisation, Zillner adds, that “AI is across all sectors. All the sectors are investing. This is an important take-home – working across sectors is really central. We want to leverage AI driven market opportunity, across all sectors.”

A European AI Framework

David Bisset, Executive Director of euRobotics and Co-Chief Editor of the strategy presented an AI Framework that builds on the legal and societal fabric that underpins the impact of AI in Europe. Central to the success of this framework will be an ecosystem that brings together skills, data, and environments to experiment and deploy the technology.  Bisset says “we need data stores, regulatory sandboxes that allow us to test robots in cities without breaking the rules, we need EU regulation that creates a level playing field”. New technologies that work across sectors are needed for sensing, measurement and perception, continuous and integrated knowledge, trustworthy, hybrid decision making, physical and human action and interaction, systems, methodologies, and hardware.

Boosting the adoption of AI in Europe faces several challenges however, including the lack of skills, technology problems, lack of private investment, complexity of deploying products, and the policy and regulation landscape. Bisset says “We need a collective action from all stakeholders to address these challenges – we need an AI Innovation ecosystem”. Stakeholders include researchers, innovators, technology creators, regulators, users, citizens, investors, data suppliers, and application providers.

The implementation of the AI PPP will address the following Working Areas.

WA1: Mobilising the EU AI ecosystem

WA2: Skills and acceptance

WA3: Innovation and market enablers

WA4: Guiding standards and regulation

WA5: Promoting Research Excellence

Words of Wisdom

The event closed with a panel discussion, here are some nuggets of wisdom.

Photo: Panel; Credits: Octavian Carare

“We will not have enough people to treat patients. Without AI, without robotics, it will be a disaster for patients. Whatever the cost, patients will demand their treatments. Treatments will be better, more precise with AI.”  Rolf Roussaint, Director of Anaesthesiology at University Hospital Aachen.

“We need to measure that AI is bringing benefits – measurable AI is important, maybe more important than explainable AI” Henk-Jan Vink, TNO Managing Director Unit ICT.

“We should not be too shy about what the EU is doing in AI – we should be proud.”

“We need to be inclusive, it’s not robotics vs AI. We see robotics and AI as two sides of the same coin. One is a physical instantiation of AI. We need to join forces.” Juha Heikkilä, Head of Unit, Robotics and Artificial Intelligence, DG CONNECT, European Commission.

“Don’t fall in love with the technology – researchers are in love with the technology – and industry with profit. Instead we need use cases proving the new benefits for services and impacts on quality of life that AI can bring.” Gabriella Cattaneo, Associate Vice President of IDC4EU European Government Consulting Unit.

“There will be no progress for AI if we can’t find a way for researchers and startups to have access to data.” Hubert Tardieu, ATOS, Advisor to the CEO.

“In China and the US data is not an issue. The EU doesn’t have that data however – or it’s not shared due to concern. Focussing on the citizen is really important, but we also need to push for access to data.” Federico Milani, Deputy Head of Unit, Data Policies and Innovation, DG CONNECT, European Commission.

Call for collaboration

To conclude the event, Thomas Hahn, President of BDVA and Bernd Liepert, President of euRobotics, moderators of this event and the panel, launched a call for participation and collaboration to all European players active in this domain and committed to boost AI adoption in Europe!

Photo: Bernd Liepert, President of euRobotics, Thomas Hahn, President of BDVA, Credits: Octavian Carare