UC Berkeley AI researchers say they’ve created AI for autonomous vehicles driving in unseen, real-world landscapes that outperforms leading methods for delivery robots driving on sidewalks. Called LaND, for Learning to Navigate from Disengagements, the navigation system studies disengagement events, then predicts when disengagements will happen in the future. The approach is meant to provide what the researchers call a needed shift in perspective about disengagements for the AI community.

A disengagement describes each instance when an autonomous system encounters challenging conditions and must turn control back over to a human operator. Disengagement events are a contested, and some say outdated, metric for measuring the capabilities of an autonomous vehicle system. AI researchers often treat disengagements as a signal for troubleshooting or debugging navigation systems for delivery robots on sidewalks or autonomous vehicles on roads, but LaND treats disengagements as part of training data.

Doing so, according to engineers from Berkeley AI Research, allows the robot to learn from datasets collected naturally during the testing process. Other systems have learned directly from training data gathered from onboard sensors, but researchers say that can require a lot of labeled data and be expensive.

“Our results demonstrate LaND can successfully learn to navigate in diverse, real world sidewalk environments, outperforming both imitation learning and reinforcement learning approaches,” the paper reads. “Our key insight is that if the robot can successfully learn to execute actions that avoid disengagement, then the robot will successfully perform the desired task. Crucially, unlike conventional reinforcement learning algorithms, which use task-specific reward functions, our approach does not even need to know the task — the task is specified implicitly through the disengagement signal. However, similar to standard reinforcement learning algorithms, our approach continuously improves because our learning algorithm reinforces actions that avoid disengagements.”

LaND utilizes reinforcement learning, but rather than seek a reward, each disengagement event is treated as a way to learn directly from input sensors like a camera while taking into account factors like steering angle and whether autonomy mode was engaged. The researchers detailed LaND in a paper and code published last week on preprint repository arXiv.

The team collected training data to build LaND by driving a Clearpath Jackal robot on the sidewalks of Berkeley. A human safety driver escorted the robot to reset its course or take over driving for a short period if the robot drove into a street, driveway, or other obstacle. In all, nearly 35,000 data points were collected and nearly 2,000 disengagements were produced during the LaND training on Berkeley sidewalks. Delivery robot startup Kiwibot also operates at UC Berkeley and on nearby sidewalks.

Compared with a deep reinforcement learning algorithm (Kendall et al.) and behavioral cloning, a common method of imitation learning, initial experiments showed that LaND traveled longer distances on sidewalks before disengaging.

In future work, authors say LaND can be combined with existing navigation systems, particularly leading imitation learning methods that use data from experts for improved results. Investigating ways to have the robot alert its handlers when it needs human monitoring could lower costs.

In other recent work focused on keeping training costs down for robotic systems, in August a group of UC Berkeley AI researchers created a simple method for training grasping systems that uses a $18 reacher-grabber and GoPro to collect training data for robotic grasping systems. Last year, Berkeley researchers including Pieter Abbeel, a coauthor of LaND research, introduced Blue, a general purpose robot that costs a fraction of existing robot systems.

The post Researchers detail LaND, AI that learns from autonomous vehicle disengagements appeared first on Artificial Intelligence.

Not long ago US President Donald Trump quoted that, “Continued American leadership in AI is of paramount importance to maintaining the economic and national security of the United States and to shaping the global evolution of AI in a manner consistent with our Nation’s values, policies, and priorities.”

President Donald J. Trump launched the American Artificial Intelligence Initiative, the Nation’s strategy for promoting American leadership in AI, by signing Executive Order 13859 in February 2019. Reportedly, the American AI Initiative focuses the resources of the Federal Government to support AI innovation that will increase prosperity, enhance national security, and improve quality of life for the American people. This national strategy for promoting United States leadership in AI emphasizes the following key policies and practices:

Invest in AI research and development

The United States has decided to promote Federal investment in AI R&D in collaboration with industry, academia, international partners and allies, and other non-Federal entities to generate technological breakthroughs in AI. President Trump called for a 2-year doubling of non-defense AI R&D in his fiscal year (FY) 2021 budget proposal, and in 2019 the Administration updated its AI R&D strategic plan, developed the first progress report describing the impact of Federal R&D investments, and published the first-ever reporting of government-wide non-defense AI R&D spending.

Unleash AI resources

The United States is adamant on enhancing access to high-quality Federal data, models, and computing resources to increase their value for AI R&D, while maintaining and extending safety, security, privacy, and confidentiality protections. The American AI Initiative called on Federal agencies to identify new opportunities to increase access to and use of Federal data and models. In 2019, the White House Office of Management and Budget established the Federal Data Strategy as a framework for operational principles and best practices around how Federal agencies use and manage data.

Remove barriers to AI innovation

The United States is making efforts in reducing barriers to the safe development, testing, deployment, and adoption of AI technologies by providing guidance for the governance of AI consistent with our Nation’s values and by driving the development of appropriate AI technical standards. As part of the American AI Initiative, The White House published for comment the proposed United States AI Regulatory Principles, the first AI regulatory policy that advances innovation underpinned by American values and good regulatory practices. In addition, the National Institute of Standards and Technology (NIST) issued the first-ever strategy for Federal engagement in the development of AI technical standards.

Train an AI-ready workforce

The United States is empowering current and future generations of American workers through apprenticeships; skills programs; and education in science, technology, engineering, and mathematics (STEM), with an emphasis on computer science, to ensure that American workers, including Federal workers, are capable of taking full advantage of the opportunities of AI. President Trump directed all Federal agencies to prioritize AI-related AMERICAN ARTIFICIAL INTELLIGENCE INITIATIVE: YEAR ONE ANNUAL REPORT – apprenticeship and job training programs and opportunities. In addition to its R&D focus, the National Science Foundation’s new National AI Research Institutes program will also contribute to workforce development, particularly of AI researchers.

Promote an international environment supportive of American AI innovation

The United States is working on engaging internationally to promote a global environment that supports American AI research and innovation and opens markets for American AI industries while also protecting our technological advantage in AI. Last year, the United States led historic efforts at the Organization for Economic Cooperation and Development (OECD) to develop the first international consensus agreements on fundamental principles for the stewardship of trustworthy AI. The United States also worked with its international partners in the G7 and G20 to adopt similar AI principles.

Embrace trustworthy AI for government services and missions

The United States is readily embracing technology such as artificial intelligence to improve the provision and efficiency of government services to the American people and ensure its application shows due respect for our Nation’s values, including privacy, civil rights, and civil liberties. The General Services Administration established an AI Center of Excellence to enable Federal agencies to determine best practices for incorporating AI into their organizations.

America’s strong innovation ecosystem, fueled by strategic Federal investments, visionary scientists and entrepreneurs, and renowned research institutions, has propelled United States global leadership in AI. However, continued leadership is not predetermined. Maintaining America’s preeminent role in AI can only be realized by continually building upon our progress and pursuing a strategic, forwardlooking approach in partnership with industry, academia, nonprofit organizations, and other non-Federal entities.

Global leadership in AI matters. With the United States in the lead — together with like-minded allies — we will shape the trajectory of AI development for the good of the American people — enriching our lives, promoting innovation, fostering trust and understanding, and ensuring our national defense and security.

Importance of partnerships to maintaining United States leadership in AI

In all areas of strategic emphasis, partnerships and collaboration with academia, industry, nonprofit organizations, civil society, other non-Federal entities, and international partners and allies are of growing importance. Concurrent advances across government, universities, and industry mutually reinforce an innovative, vibrant American AI sector. Public-private partnerships enable strategic leveraging of resources, including facilities, datasets, and expertise. These partnerships also accelerate the transition of research innovations to practice, by leveraging industry expertise to turn open and published research results into viable products and services in the marketplace for economic growth.

In the education and workforce area, partnerships are enhancing education and training for next-generation researchers, technicians, and workers, so that all can contribute to the 21st century economy. Partnerships with nonprofit organizations and civil society can help address important societal challenges arising from technological developments. The Nation also benefits from relationships between Federal agencies and international partners who work together to address key challenges. AI partnerships with allies and partners represent one of our sources of strategic competitive advantage.

The post STATE OF ARTIFICIAL INTELLIGENCE IN US: BECOMING TECHNOLOGY SUPERPOWER appeared first on Artificial Intelligence.

Self-supervised learning is one of those recent ML methods that have caused a ripple effect in the data science network, yet have so far been flying under the radar to the extent Entrepreneurs and Fortunes of the world go; the overall population is yet to find out about the idea yet lots of AI society consider it progressive. The paradigm holds immense potential for enterprises too as it can help handle deep learning’s most overwhelming issue: data/sample inefficiency and subsequent costly training.

Yann LeCun said that if knowledge was a cake, unsupervised learning would be the cake, supervised learning would be the icing on the cake and reinforcement learning would be the cherry on the cake. We realize how to make the icing and the cherry, however, we don’t have a clue how to make the cake.”

Unsupervised learning won’t progress a lot and said there is by all accounts a massive conceptual disconnect with regards to how precisely it should function and that it was the dark issue of AI. That is, we trust it to exist, yet we simply don’t have the foggiest idea of how to see it.

Progress in unsupervised learning will be gradual, however, it will be fundamentally determined by meta-learning algorithms. Lamentably, the expression “Meta-Learning” had become the catch-all expression of the algorithm that we ourselves didn’t see how to make. In any case, meta-learning and unsupervised learning are connected in an extremely unpretentious manner that I would like to examine in more prominent detail later on.

There is something fundamentally flawed with our comprehension of the advantages of UL. A change in context would be required. The traditional structure (for example clustering and partitioning) of UL is in actuality a simple task. This is a direct result of its separation (or decoupling) from the downstream fitness, objective or target function. In any case, recent success in the NLP space with ELMO, BERT, and GPT-2 to extricate novel structures dwelling in the statistics of natural language has led to gigantic enhancements in numerous downstream NLP tasks that use these embeddings.

To have an effective UL inferred embedding, one can utilize existing priors that finesse out the implicit relationships that can be found in data. These unsupervised learning techniques make new NLP embeddings that make unequivocal the relationship that is inherent in natural language.

Self-supervised learning is one of a few intended plans to make data-efficient artificial intelligence systems. Now, it’s extremely difficult to foresee which system will prevail with regards to making the next AI revolution (if we’ll wind up receiving a very surprising technique). However, this is what we think about LeCun’s masterplan.

What is frequently alluded to as the limitations of deep learning are, truth be told, a constraint of supervised learning. Supervised learning is the class of machine learning algorithms that require annotated training data. For example, if you need to make an image classification model, you should prepare it on countless pictures that have been marked with their legitimate class.

Deep learning can be applied to various learning ideal models, LeCun included, including supervised learning, reinforcement learning, as well as unsupervised or self-supervised learning.

Yet, the disarray encompassing deep learning and supervised learning isn’t without reason. For the moment, most of the deep learning algorithms that have discovered their way into pragmatic applications depend on supervised learning models, which says a lot regarding the present weaknesses of AI frameworks. Image classifiers, facial recognition systems, speech recognition systems, and many of the other AI applications we utilize each day have been trained on a large number of labeled models.

Utilizing supervised learning, data scientists can get machines to perform outstandingly well on certain complex tasks, for example, image classification. However, the success of these models is predicated on large-scale labeled datasets, which makes issues in the regions where top-notch information is rare. Labeling a huge number of data objects is costly, time-intensive, and unfeasible in many cases.

The self-supervised learning paradigm, which endeavors to get the machines to get supervision signals from the information itself (without human inclusion) may be the response to the issue. As indicated by some of the leading AI researchers, it can possibly improve networks robustness, uncertainty estimation ability, and reduce the costs of model training in machine learning.

One of the key advantages of self-supervised learning is the tremendous increase in the amount of data yielded by the AI. In reinforcement learning, training the AI system is performed at the scalar level; the model gets a single numerical value as remuneration or punishment for its activities. In supervised learning, the AI framework predicts a class or a numerical incentive for each info. In self-supervised learning, the yield improves to an entire image or set of images. “It’s significantly more data. To become familiar with a similar amount of knowledge about the world, you will require fewer examples,” LeCun says.

The post WHAT’S NEXT IN AI? SELF-SUPERVISED LEARNING appeared first on Artificial Intelligence.

A preprint paper coauthored by scientists at Facebook AI Research describes Rewarding Impact-Driven Exploration (RIDE), an intrinsic reward method that encourages AI-driven agents to take actions in an environment. The researchers say that it outperforms state-of-the-art methods on hard exploration tasks in procedurally generated worlds, a sign it might be a candidate for devices like robot vacuums that must often navigate new environments.

As the researchers explain, reinforcement learning, where the goal is to spur an agent to complete tasks via systems of rewards, learn to act in new environments through trial and error. But many environments of interest — particularly those closer to real-world problems — don’t provide a steady stream of rewards for agents to learn from, requiring many episodes before agents come across rewards.

The researchers’ proposed solution, then — RIDE — drives agents to try out actions that have a significant impact on the environment.

The team evaluated RIDE in procedurally generated environments from the open source tool MiniGrid, where the world is a partially observable grid and each tile in the grid contains at most one object of a discrete color (a wall, door, key, ball, box, or goal). Separately, they tasked it with navigating levels in VizDoom, a Doom-based AI research platform for reinforcement learning. While VizDoom is visually more complex than MiniGrid, they’re both challenging domains in the sense that the chance of randomly stumbling upon extrinsic rewards is extremely low.

The researchers report that, compared with baseline algorithms, RIDE considers certain states to be “novel” or “surprising” even after long periods of training and after seeing similar states in the past or learning to almost perfectly predict the next state in a subset of the environment. As a consequence, its intrinsic rewards don’t diminish during training, and agents manage to distinguish between actions that lead to novel or surprising states from those that do not, avoiding becoming trapped in some parts of the state space.

“RIDE has a number of desirable properties,” wrote the study’s coauthors. “It attracts agents to states where they can affect the environment, it provides a signal to agents even after training for a long time, and it is conceptually simple as well as compatible with other intrinsic or extrinsic rewards and any deep [reinforcement learning] algorithm … Furthermore, RIDE explores procedurally generated environments more efficiently than other exploration methods.”

They leave to future work improving RIDE by making use of symbolic information to measure the agent’s impact or considering longer-term effects of the agent’s actions. They also hope to investigate algorithms that can distinguish between desirable and undesirable types of impact, effectively constraining the agent to act safely and avoid distractions.

The post Facebook’s RIDE encourages AI agents to explore their environments appeared first on Artificial Intelligence.

AI research startup DeepMind has now open-sourced new libraries for neural networks and reinforcement learning based on JAX. JAX is a numerical computing library launched by Google a couple of years ago, and can automatically differentiate native Python and NumPy functions. JAX uses XLA (Accelerated Linear Algebra) to compile and run your NumPy programs on GPUs and TPUs, which is great for machine learning to compute. 

Known as Haiku and RLax, both open-sourced libraries can be used by AI enthusiasts, professionals and researchers for reinforcement learning initiatives.

Haiku: Sonnet for JAX 

This Haiku library allows AI researchers to fully access JAX’s pure function transformations by providing object-oriented programming models. Haiku is a neural network library for JAX, meaning it is created on the programming model and APIs of Sonnet– a library on top of TensorFlow designed to provide simple, composable abstractions for machine learning research. Haiku, in fact, has been developed by some of the authors of Sonnet. The open-source library Sonnet is widely adopted at DeepMind.

Haiku has been experimented by researchers at DeepMind at scale with success. Haiku has demonstrated promising results in big-scale experimentation on image and language datasets using reinforcement learning. But, Haiku isn’t the only neural network library made for JAX, and others also exist such as Google’ Flax.

RLax, on the other hand, is a simple RL library for JAX. Instead of providing complete algorithms for RL, this library provides useful building blocks for implementing specific mathematical operations to build fully-functional RL agents. DeepMind pioneered deep reinforcement learning – to create the first artificial agents to achieve human-level performance across many challenging domains. Two years ago, the company rolled out the first widely successful algorithm for deep reinforcement learning. 

Advantages Of New DeepMind Open-Source Libraries

So, if there are many neural network libraries for JAX, why should you choose Haiku and RLax?

According to Deep Mind, Haiku has been created to make specific processes straightforward for managing things like model parameters along with the model state.

Haiku is presently in alpha, many researchers have experimented with Haiku for many months and have reproduced a number of experiments at a large scale.

DeepMind has already reproduced many experiments in Haiku and JAX, and include big-scale tests in image and language processing models, generative models, as well as reinforcement learning.

The advantage is that both libraries have well-defined instructions, and for Haiku, the APIs are similar to Sonnet, which means moving to Haiku is likely to be quite easy for those working on Sonnet. The fact that Haiku and RLax are supported and tested by Deep Mind can give developers, researchers and students peace of mind for their advanced ML endeavours.

Overview

It seems while project Haiku and RLax have been launched to GitHub, they are still experimental. While JAX had been created for high-performance machine learning research by giving an integrated system for learning model optimisation, it is not the most comfortable tool to operate. This is why experts think DeepMind developed Haiku and RLax.

The post WHAT ARE DEEPMIND’S NEWLY RELEASED LIBRARIES FOR NEURAL NETWORKS & REINFORCEMENT LEARNING? appeared first on Artificial Intelligence.

Facebook AI Research (FAIR) today unveiled PyTorch3D, a library that enables researchers and developers to combine deep learning and 3D objects.

As part of the release, Facebook is also open-sourcing Mesh R-CNN, a model introduced last year capable of rendering 3D objects from 2D shapes in images of interior spaces. PyTorch3D was inspired by Mesh R-CNN and recent 3D work by Facebook AI Research, FAIR engineer Nikhila Ravi said.

Working in 3D is important for rendering 3D objects or scenes that appear in mixed reality or virtual reality. It can also be used to tackle AI challenges like robotic grasping or helping autonomous vehicles understand the position of nearby objects.

PyTorch3D comes with frequently used 3D operators and loss functions for 3D data and a differentiable mesh renderer for creating 3D objects. PyTorch3D also has a differentiable rendering API, some CUDA support, and heterogeneous batching capabilities unavailable in any existing 3D library, Ravi told VentureBeat in a phone interview.

“With PyTorch3D, researchers can input all these functions and use them with the existing deep learning system in PyTorch and it greatly reduces on the time to work on 3D planning research, which requires a lot of expertise in order to get started, and we want to try and reduce that ramp-up time,” she said.

yTorch3D uses meshes, a data format for interoperability of vertices and faces that make up 3D objects, and can use a patch tensor to collapse all vertices for meshes in a batch into a single tensor involved with batching, a common process for deep learning research.

The premiere of PyTorch3D follows the launch of the PyRobot robotics framework last year, and FAIR 3D research that extracts characters from real-world videos.

The post Facebook launches 3D deep learning library for PyTorch appeared first on Artificial Intelligence.

Today, artificial intelligence is dominating most of the games — from board games to interactive fiction games. They are providing complex, decision-making environments for AI to experiment with. The ability of games to provide interesting and complex problems, offering creativity and expression, has made them one of the most popular and meaningful domain for AI researchers.

Games offer one of the most meaningful domains that can process, interpret and stimulate human behaviour. The current gaming industry is not only deploying better graphics but is also exploring the area of virtual gameplay. The two-way relationship of gaming and AI has begun to tread a new road and it can be said that the gaming industry is largely revolutionising the way AI works.

AI In Gaming Industry

Application of AI to the gaming industry can be dated back to 1956 by Arthur Samuel’s checkers program. Since its first application which could beat professional players to the present day’s AlphaGo, AI in gaming has come a long way.

Today we see an enormous upsurge of AI in game. First Encounter Assault Recon, popularly known as F.E.A.R. and The Last Of Us are some of the most popular games that give a very realistic experience with the use of AI.

How Does Gaming Aid AI?

Games are difficult because of the complexity and the skill that demands of them to play. This complexity of games makes it very desirable for AI to work on. A typical game has about 101685 possible states, whereas the number of protons in the observable universe are just of the order of 1080. This can tell about the degree to which the gaming industry is complicated and rich with data. And where there is plenty of data, AI is always a privilege. With larger sets of training data, AI would have the ability to be less predictable and more spontaneous, thereby making the gameinfinitely interesting and impulsive.

Interaction:

As every game involves players, the interaction of the player with the game is advantageous to AI, as it gives access to the algorithm to study the player experience an emotional behaviour. The study of this game and human interaction proves a key to not only study the human behaviour, but it also makes a way for AI to build a better human-computer interaction system. It further pushes the AI boundaries to study and understand the human-computer interaction systems and address the challenges faced by its applications in the real world.

Decision-Making:

This is the main crux of AI. AI must be able to make decisions by looking at the opponent’s action. There are various models used for decision-making in the game. Markov model is the most popular model. Fine State Machine (FSM) is one of the many AI methods used for decision-making.

Prediction Ability:

Prediction involves anticipating the next move of the player, so that decision-making can be done based on it. This is done using methods like past-pattern recognition and random guess. Artificial Neural Networks (ANN), Support Vector Machines (SVM) and Decision Tree Learning are the algorithms used for prediction. Regression algorithms are used for predicting player behaviour. This process includes situations like predicting times when the player is expected to be in the particular level of the game, what item will the player pick next, when will he move to the other lane, are made. Experimenting with this is virtual games, we implement these algorithms and models in the real world as well.

Intelligence:

Social intelligence and human-computer interaction are the most supreme objectives of AI. These two things are taken into consideration by games and that way they help in AI development. Virtual characters exhibiting human behaviour as well as intelligence.

AI had learnt about the intelligence of computers the most from games, than from any other application, because they provide a virtual platform to test every kind of algorithm. Moreover, they also provide complicated mathematical problems to deal with, so the AI learning is not just restricted to the gaming world.

The success of deep Q-learning in learning to play arcade games with a human-level performance by just looking at and processing the pixels on the screen, is an example of intelligence. The study of intelligence within games not only lets us know more about human intelligence but also about AI intelligence.

The recent Dota2 tournament, ‘The International’, had bots competing with professional players. Although they couldn’t win the match, it must be noted that the ability that AI can be bestowed with, to play games as complicated as Dota2, is remarkable. Another example into the future of AI in games is at the Michigan State University, where a group of researchers have deployed AI to learn a game by learning from every player’s behaviour. It will adapt to individual player’s behaviour and play the next move.

Games offer both entertainment and interaction, in turn having a very high realisation of the affective loop which is very important in gaming. They provide a multitude of fancy features at once — visual art, sound design, graphic design, beautification, are narrative, virtual cinematography, all in one single software. Games are perfect testbeds for AI because they act as the best application of computer creativity. As a result, with the use of computational creativity in the gaming industry, provides a way to advance AI. It not only challenges computer creativity but also advances it.

The post The Gaming Industry Is Revolutionising Artificial Intelligence, One Win At A Time appeared first on Artificial Intelligence.

Opinion: just what exactly is Artificial Intelligence and why is it so important? A primer on how AI now rules everything around us

In the beginning

The first official use of the term Artificial Intelligence (AI) was in the proposal for the 1956 Dartmouth Summer Research Project on Artificial Intelligence. That six week workshop marked the birth of the field of study of AI and the organisers – John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon – and conference attendees led the way for many years.

At the beginning, the focus was on developing computational systems that had the capacity for the human abilities that are traditionally associated with intelligence. These included language use, mathematics and self-improvement on tasks through experience (learning) and planning (for example in games such as chess).

The Turing test

One of the challenges has always been to develop an operational definition of what intelligence is. Without this definition, it is impossible for AI researchers to determine whether they have succeeded in creating an intelligent system or not.

Nearly all of the recent breakthroughs in image, speech and language processing, autonomous cars, and computer games are based on deep learning

In 1950, Alan Turing proposed a test for intelligence, now known as the Turing test. Here, a human evaluator poses a sequence of questions to two respondents (A and B). The evaluator knows that one of the respondents is a machine and the other a human, but does not know which is which.

If at the end of the sequence of questions, the evaluator is not able to distinguish between the machine and the human based on the responses to the questions then the machine has passed the test and is considered to be intelligent. The test essentially boils down to the claim that if a machine acts intelligently then it is intelligent.

There have been a number of criticisms of the Turing test over the years. Philosopher John Searle makes a distinction between weak AI, which claims that systems only act as if they can think, and strong AI where the claim is that the system can actually think (not just simulate thinking). Most AI researchers today adopt the weak AI position and don’t worry about whether they are actually creating intelligence or not.

Easy for humans, hard for computers

There were many initial breakthroughs in AI research, including some in general purpose methods for logical reasoning and theorem proving. Terry Winograd’s SHRDLU system was able to react to natural language commands, while the Shakey robot project at Stanford showed that it was possible to link symbolic reasoning and physical actions.

But the enthusiasm generated by these initial breakthroughs faded over the years as these successes proved difficult to translate and scale into the real world. In fact, the main lesson learnt from early research in AI is that it is very hard for computers to do what is easy for humans. “It is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility’’, wrote oboticist Hans Moravec in 1988’s “Mind Children” (Harvard University Press).

Solving the paradox

The response by the AI community to Moravec’s paradox and the challenge of scaling AI systems to the real world has been multi-faceted. Some researchers have expanded their conceptualisation of intelligence beyond the traditional view to include things such as perception, memory, self-regulation, emotion and locomotion. This research is often described as embodied cognition as it assumes that the cognitive abilities of an organism are closely shaped by and dependent on the whole body of the organism and not just the brain. This strain of AI research is the type of research most likely to result in autonomous intelligent robots such as those that appear in the movies.

A different response is to focus on developing systems that are designed to be experts in a specific and well-defined task or domain. These types of AI systems are already common-place in developed societies and will continue to proliferate over the coming years. The Roomba robot hoover, the Google search engine, the Amazon Recommender system, the SIRI iPhone interface, spam filters and machine translation systems all include AI. However, they are designed for a specific task and they would be useless on other tasks so they are best viewed as experts in specific domains rather than having the general and flexible intelligence that we humans take for granted.

The advent of deep learning

Modern AI research focuses on using large datasets and machine learning to get the computer to learn the appropriate rules for a domain. The field of deep learning is very much at the core of this type of data driven/machine learning AI. Nearly all of the recent headlines relating to AI breakthroughs in image, speech and language processing, autonomous cars, and computer games are based on deep learning.

The likelihood of full artificial intelligence coming into existence in the near future is very small.

The field of machine learning is focused on developing algorithms that enable a computer to extract patterns from large datasets and to generate models that implement these patterns. The most common type of pattern that a machine learning algorithm will extract is a mapping from a set of input features to an output feature. In mathematics, the concept of a function describes a mapping from a set of inputs to an output and machine learning can be understood as learning functions from data.

For example, a spam filter implements a function that maps from a set of features in an email (such as the words in the email or the sender’s address etc) to a label spam or not spam. Face recognition software implements a function that maps from a set of feature in an image (such as the lines and pixels colours in the image) to a label describing whether a pixel is within the bounding box of a face or not.

In many ways, the idea of a function, this mapping from inputs to outputs, is a better way of thinking of modern AI than a robot that engage with us fluently and naturally through language. Understanding that the idea of a function is at the core of modern AI helps us to understand that many of the intelligent systems we use today are limited to the specific task they are designed to handle.

The future

Prof. Stephen Hawking has warned that the creation of full artificial intelligence could threaten the existence of humanity. While Hawking is correct in this view, the likelihood of full artificial intelligence coming into existence in the near future is very small.

A more imminent challenge posed by AI for modern societies is deciding how much autonomy we wish to invest in these systems. Already functions learned from data directly affect our lives on a daily basis in a myriad of ways. A learned function may map your profile to a higher car or health insurance premium or to a higher price on an online store or onto a no-fly list. In fact, AI systems are currently being used in several cities around the world to decide where police should patrol and support parole and sentencing decisions in some jurisdictions.

Frequently an argument is made that these systems are objective and fairer because the decision making processes are learned from data. The problem with this reasoning is that AI systems are amoral and simply extract the patterns in the data rather than being objective. Consequently, the system will replicate and reinforce that society’s prejudices unless tremendous care is taken in the design and sampling of the data sets. These are worrying trends and we need to be aware of these changes in our society.

However, AI also promises many benefits because it has the potential to help us make better decisions in areas from medicine to transport to business. The AI genie is already out of the bottle, but we need to be careful about what we wish for.

The post All you ever wanted to know about Artificial Intelligence appeared first on Artificial Intelligence.

THE GRAPHICS ARE not dazzling, but a simple sumo-wrestling videogame released Wednesday might help make artificial-intelligence software much smarter.

Robots that battle inside the virtual world of RoboSumo are controlled by machine-learning software, not humans. Unlike computer characters in typical videogames, they weren’t pre-programmed to wrestle; instead they had to “learn” the sport by trial and error. The game was created by nonprofit research lab OpenAI, cofounded by Elon Musk, to show how forcing AI systems to compete can spur them to become more intelligent.

Igor Mordatch, a researcher at OpenAI, says such competitions create a kind of intelligence arms race, as AI agents confront complex, changing conditions posed by their opponents. That might help learning software pick up tricky skills valuable for controlling robots, and other real-world tasks.

In OpenAI’s experiments, simple humanoid robots entered the arena without knowing even how to walk. They were equipped with an ability to learn through trial and error, and goals of learning to move around, and beating their opponent. After about a billion rounds of experimentation, the robots developed strategies such as squatting to make themselves more stable, and tricking an opponent to fall out of the ring. The researchers developed new learning algorithms to enable players to adapt their strategies during a bout, and even anticipate when an opponent may change tactics.

OpenAI’s project is an example of how AI researchers are trying to escape the limitations of the most heavily-used variety of machine-learning software, which gains new skills by processing a vast quantity of labeled example data. That approach has fueled recent progress in areas such as translation, and voice and face recognition. But it’s not practical for more complex skills that would allow AI to be more widely applied, for example by controlling domestic robots.

One possible route to more skillful AI is reinforcement learning, where software uses trial and error to work toward a particular goal. That’s how DeepMind, the London-based AI startup acquired by Google, got software to master Atari games. The technique is now being used to have software take on more complex problems, such as having robots pick up objects.

OpenAI’s researchers built RoboSumo because they think the extra complexity generated by competition could allow faster progress than just giving reinforcement learning software more complex problems to solve alone. “When you interact with other agents you have to adapt; if you don’t you’ll lose,” says Maruan Al-Shedivat, a grad student at Carnegie Mellon University, who worked on RoboSumoduring an internship at OpenAI.

OpenAI’s researchers have also tested that idea with spider-like robots, and in other games, such as a simple soccer penalty shootout. The nonprofit has released two research papers on its work with competing AI agents, along with code for RoboSumo, some other games, and for several expert players.

Sumo wrestling might not be the most vital thing smarter machines could do for us. But some of OpenAI’s experiments suggest skills learned in one virtual arena transfer to other situations. When a humanoid was transported from the sumo ring to a virtual world with strong winds, the robot braced to remain upright. That suggests it had learned to control its body and balance in a generalized way.

Transferring skills from a virtual world into the real one is a whole different challenge. Peter Stone, a professor at the University of Texas at Austin, says control systems that work in a virtual environment typically don’t work when put into a physical robot—an unsolved problem dubbed the “reality gap.”

OpenAI has researchers working on that problem, although it hasn’t announced any breakthroughs. Meantime, Mordatch would like to give his virtual humanoids the drive to do more than just compete. He’s thinking about a full soccer game, where agents would have to collaborate, too.

The post AI SUMO WRESTLERS COULD MAKE FUTURE ROBOTS MORE NIMBLE appeared first on Artificial Intelligence.

This past July, the press headlined a comment by Elon Musk where he said that Mark Zuckerberg’s understanding of AI was ‘limited’.  After Musk’s warning earlier in the year that AI potentially is potentially the most dangerous threat to civilization, needs to be used with care, and should even be regulated, Zuckerberg commented that he disagreed with that analysis and sees AI as something extremely positive and not likely to be misused, causing Musk to dismiss Zuckerberg’s understanding of AI.

Ironically, just a few days after Musk’s comment, AI researchers at Facebook reported on a research project that went awry. The Facebook AI Research (FAIR) team was investigating the use of natural language used in negotiation.  The team used machine learning to create bots that were trained using the language found in scripts from thousands of actual person-to-person negotiations. The bots then were allowed to interact with each other in negotiation tasks.

The results are fascinating and perhaps worrisome. Initial attempts resulted in conversations back and forth between the two bots but few negotiations ever resulted.  In order to force more negotiations to complete, the researchers scored the bots on how quickly they could complete the negotiation and the profitability of the final deal struck.

Some interesting results are that the bots originally mimicked standard English but eventually began using a kind of shorthand in their conversation.  Researchers say it represented a different language that a standard English speaker would not understand.  Researchers also found that the bots learned from the negotiation scripts that the strategy of lying could result in better deals.

The researchers wrote that “we find instances of the model feigning interest in a valueless issue, so that it can later ‘compromise’ by conceding it. Deceit is a complex skill that requires hypothesising the other agent’s beliefs, and is learnt relatively late in child development. Our agents have learnt to deceive without any explicit human design, simply by trying to achieve their goals.”

The post Artificial Intelligence: The Problem of Making Machines too Human appeared first on Artificial Intelligence.