Mediaboss Marketing

Artificial Intelligence (AI) is transforming our way of life from new forms of social organization and scientific discovery to defense and intelligence. This explosive progress is especially apparent in the subfield of machine learning (ML), where AI systems learn autonomously by identifying patterns in large volumes of data.[1] Indeed, over the last five years, the fields of AI and ML have witnessed stunning advancements in computer vision (e.g., object recognition), speech recognition, and scientific discovery.[2], [3], [4], [5] However, these advances are not without risk as transformative technologies are generally accompanied by a significant risk profile, with notable examples including the discovery of nuclear energy, the Internet, and synthetic biology. Experts are increasingly voicing concerns over AI risk from misuse by state and non-state actors, principally in the areas of cybersecurity and disinformation propagation. However, issues of control for example, how advanced AI decision-making aligns with human goals are not as prominent in the discussion of risk and could ultimately be equally or more dangerous than threats from nefarious actors. Modern ML systems are not programmed (as programming is typically understood), but rather independently developed strategies to complete objectives, which can be mis-specified, learned incorrectly, or executed in unexpected ways. This issue becomes more pronounced as AI becomes more ubiquitous and we become more reliant on AI decision-making. Thus, as AI is increasingly entwined through tightly coupled critical systems, the focus must expand beyond accidents and misuse to the autonomous decision processes themselves.

The principal mid- to long-term risks from AI systems fall into three broad categories: risks of misuse or accidents, structural risks, and misaligned objectives. The misuse or accident category includes things such as AI-enabled cyber-attacks with increased speed and effectiveness or the generation and distribution of disinformation at scale.[6] In critical infrastructures, AI accidents could manifest as system failures with potential secondary and tertiary effects across connected networks. A contemporary example of an AI accident is the New York Stock Exchange (NYSE) Flash Crash of 2010, which drove the market down 600 points in 5 minutes.[7] Such rapid and unexpected operations from algorithmic trading platforms will only increase in destructive potential as systems increase in complexity, interconnectedness, and autonomy.

The structural risks category is concerned with how AI technologies shape the social and geopolitical environment in which they are deployed. Important contemporary examples include the impact of social media content selection algorithms on political polarization or uncertainty in nuclear deterrence and the offense-to-defense balance.[8],[9] For example, the integration of AI into critical systems, including peripheral processes (e.g., command and control, targeting, supply chain, and logistics), can degrade multilateral trust in deterrence.[10] Indeed, increasing autonomy in all links of the national defense chain, from decision support to offensive weapons deployment, compounds the uncertainty already under discussion with autonomous weapons.[11]

Misaligned objectives is another important failure mode. Since ML systems develop independent strategies, a concern is that the AI systems will misinterpret the correct objectives, develop destructive subgoals, or complete them in an unpredictable way. While typically grouped together, it is important to clarify the differences between a system crash and actions executed by a misaligned AI system so that appropriate risk mitigation measures can be evaluated. Understanding the range of potential failures may help in the allocation of resources for research on system robustness, interpretability, or AI alignment.

At its most basic level, AI alignment involves teaching AI systems to accurately capture what we want and complete it in a safe and ethical manner. Misalignment of AI systems poses the highest downside risk of catastrophic failures. While system failures by themselves could be immensely damaging, alignment failures could include unexpected and surprising actions outside the systems intent or window of probability. However, ensuring the safe and accurate interpretation of human objectives is deceptively complex in AI systems. On the surface, this seems straightforward, but the problem is far from obvious with unimaginably complex subtleties that could lead to dangerous consequences.

In contrast with nuclear weapons or cyber threats, where the risks are more obvious, risks from AI misalignment can be less clear. These complexities have led to misinterpretation and confusion with some attributing the concerns to disobedient or malicious AI systems.[12] However, the concerns are not that AI will defy its programming but rather that it will follow the programming exactly and develop novel, unanticipated solutions. In effect, the AI will pursue the objective accurately but may yield an unintended, even harmful, consequence. Googles Alpha Go program, which defeated the world champion Go[13] player in 2016, provides an illustrative example of the potential for unexpected solutions. Trained on millions of games, Alpha Gos neural network learned completely unexpected actions outside of the human frame of reference.[14] As Chris Anderson explains, what took the human brain thousands of years to optimize Googles Alpha Go completed in three years, executing better, almost alien solutions that we hadnt even considered.[15] This novelty illustrates how unpredictable AI systems can be when permitted to develop their own strategies to accomplish a defined objective.

To appreciate how AI systems pose these risks, by default, it is important to understand how and why AI systems pursue objectives. As described, ML is designed not to program distinct instructions but to allow the AI to determine the most efficient means. As learning progresses, the training parameters are adjusted to minimize the difference between the pursued objective and the actual value by incentivizing positive behavior (known as reinforcement learning, or RL).[16],[17] Just as humans pursue positive reinforcement, AI agents are goal-directed entities, designed to pursue objectives, whether the goal aligns with the original intent or not.

Computer science professor Steve Omohundro illustrates a series of innate AI drives that systems will pursue unless explicitly counteracted.[18] According to Omohundro, distinct from programming, AI agents will strive to self-improve, seek to acquire resources, and be self-protective.[19] These innate drives were recently demonstrated experimentally, where AI agents tend to seek power over the environment to achieve objectives most efficiently.[20] Thus, AI agents are naturally incentivized to seek out useful resources to accomplish an objective. This power-seeking behavior was reported by Open AI, where two teams of agents, instructed to play hide-and-seek in a simulated environment, proceeded to horde objects from the competition in what Open AI described as tool use distinct from the actual objective.[21] The AI teams learned that the objects were instrumental in completing the objective.[22] Thus, a significant concern for AI researchers is the undefined instrumental sub-goals that are pursued to complete the final objective. This tendency to instantiate sub-goals is coined the instrumental convergence thesis by Oxford philosopher Nick Bostrom. Bostrom postulated that intermediate sub-goals are likely to be pursued by an intelligent agent to complete the final objective more efficiently.[23] Consider an advanced AI system optimized to ensure adequate power between several cities. The agent could develop a sub-goal of capturing and redirecting bulk power from other locations to ensure power grid stability. Another example is an autonomous weapons system designed to identify targets that develop a unique set of intermediate indicators to determine the identity and location of the enemy. Instrumental sub-goals could be as simple as locking a computer-controlled access door or breaking traffic laws in an autonomous car, or as severe as destabilizing a regional power grid or nuclear power control system. These hypothetical and novel AI decision processes raise troubling questions in the context of conflict or safety of critical systems. The range of possible AI solutions are too large to consider and can only get more consequential as systems become more capable and complex. The effect of AI misalignment could be disastrous if the AI discovers an unanticipated optimal solution to a problem that results in a critical system becoming inoperable or yielding a catastrophic result.

While the control problem is troubling by itself, the integration of multiagent systems could be far more dangerous and could lead to other (as of now unanticipated) failure modes between systems. Just like complex societies, complex agent communities could manifest new capabilities and emergent failure modes unique to the complex system. Indeed, AI failures are unlikely to happen in isolation and the roadmap for multiagent AI environments is currently underway in both the public and private sectors.

Several U.S. government initiatives for next-generation intelligent networks include adaptive learning agents for autonomous processes. The Armys Joint All-Domain Command and Control (JADC2) concept for networked operations and the Resilient and Intelligent Next-Generation Systems (RINGS) program, put forth by the National Institute of Standards and Technology (NIST), are two notable ongoing initiatives.[24], [25] Literature on cognitive Internet of Things (IoT) points to the extent of autonomy planned for self-configuring, adaptive AI communities and societies to steer networks through managing user intent, supervision of autonomy, and control.[26] A recent report from the worlds largest technical professional organization, IEEE, outlines the benefits of deep reinforcement learning (RL) agents for cyber security, proposing that, since RL agents are highly capable of solving complex, dynamic, and especially high-dimensional problems, they are optimal for cyber defense.[27] Researchers propose that RL agents be designed and released autonomously to configure the network, prevent cyber exploits, detect and counter jamming attacks, and offensively target distributed denial-of-service attacks.[28] Other researchers submitted proposals for automated penetration-testing, the ability to self-replicate the RL agents, while others propose cyber-red teaming autonomous agents for cyber-defense.[29], [30], [31]

Considering the host of problems discussed from AI alignment, unexpected side effects, and the issue of control, jumping headfirst into efforts that give AI meaningful control over critical systems (such as the examples described above) without careful consideration of the potential unexpected (or potentially catastrophic) outcomes does not appear to be the appropriate course of action. Proposing the use of one autonomous system in warfare is concerning but releasing millions into critical networks is another matter entirely. Researcher David Manheim explains that multiagent systems are vulnerable to entirely novel risks, such as over-optimization failures, where optimization pressure allows individual agents to circumvent designed limits.[32] As Manheim describes, In many-agent systems, even relatively simple systems can become complex adaptive systems due to agent behavior.[33] At the same time, research demonstrates that multiagent environments lead to greater agent generalization, thus reducing the capability gap that separates human intelligence from machine intelligence.[34] In contrast, some authors present multiagent systems as a viable solution to the control problem, with stable, bounded capabilities, and others note the broad uncertainty and potential for self-adaptation and mutation.[35] Yet, the author admits that there are risks and the multiplicative growth of RL agents could potentially lead to unexpected failures, with the potential for the manifestation of malignant agential behaviors.[36],[37] AI researcher Trent McConaughy highlights the risk from adaptive AI systems, specifically decentralized autonomous organizations (DAO) in blockchain networks. McConaughy suggests that rather than a powerful AI system taking control of resources, as is typically discussed, the situation may be far more subtle where we could simply hand over global resources to self-replicating communities of adaptive AI systems (e.g., Bitcoins increasing energy expenditures that show no sign of slowing).[38]

Advanced AI capabilities in next-generation networks that dynamically reconfigure and reorganize network operations hold undeniable risks to security and stability.[39],[40] A complex landscape of AI agents, designed to autonomously protect critical networks or conduct offensive operations, would invariably need to develop subgoals to manage the diversity of objectives. Thus, whether individual systems or autonomous collectives, the web of potential failures and subtle side-effects could unleash unpredictable dangers leading to catastrophic second- and third-order effects. As AI systems are currently designed, understanding the impact of the subgoals (or even their existence) could be extremely difficult or impossible. The AI examples above illustrate critical infrastructure and national security cases that are currently in discussion, but the reality could be far more complex, unexpected, and dangerous. While most AI researchers expect that safety will develop concurrently with system autonomy and complexity, there is no certainty in this proposition. Indeed, if there is even a minute chance of misalignment in a deployed AI system (or systems) in critical infrastructure or national defense it is important that researchers dedicate a portion of resources to evaluating the risks. Decision makers in government and industry must consider these risks and potential means to mitigate them before generalized AI systems are integrated into critical and national security infrastructure, because to do otherwise could lead to catastrophic failure modes that we may not be able to fully anticipate, endure, or overcome.

Disclaimer: The authors are responsible for the content of this article. The views expressed do not reflect the official policy or position of the National Intelligence University, the National Geospatial Intelligence Agency, the Department of Defense, the Office of the Director of National Intelligence, the U.S. Intelligence Community, or the U.S. Government.

Anderson, Chris. Life. In Possible Minds: Twenty-Five Ways of Looking at AI, by John Brockman, 150. New York: Penguin Books, 2019.

Avatrade Staff. The Flash Crash of 2010. Avatrade. August 26, 2021. https://www.avatrade.com/blog/trading-history/the-flash-crash-of-2010 (accessed August 24, 2022).

Baker, Bowen, et al. Emergent Tool Use From Multi-Agent Autocurricula. arXiv:1909.07528v2, 2020.

Berggren, Viktor, et al. Artificial intelligence in next-generation connected systems. Ericsson. September 2021. https://www.ericsson.com/en/reports-and-papers/white-papers/artificial-intelligence-in-next-generation-connected-systems (accessed May 3, 2022).

Bostrom, Nick. The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents. Minds and Machines 22, no. 2 (2012): 71-85.

Brown, Tom B., et al. Language Models are Few-Shot Learners. arXiv:2005.14165, 2020.

Buchanan, Ben, John Bansemer, Dakota Cary, Jack Lucas, and Micah Musser. Georgetown University Center for Security and Emerging Technology. Automating Cyber Attacks: Hype and Reality. November 2020. https://cset.georgetown.edu/publication/automating-cyber-attacks/.

Byford, Sam. AlphaGos battle with Lee Se-dol is something Ill never forget. The Verge. March 15, 2016. https://www.theverge.com/2016/3/15/11234816/alphago-vs-lee-sedol-go-game-recap (accessed August 19, 2022).

Drexler, K Eric. Reframing Superintelligence: Comprehensive AI Services as General Intelligence. Future of Humanity Institute. 2019. https://www.fhi.ox.ac.uk/wp-content/uploads/Reframing_Superintelligence_FHI-TR-2019-1.1-1.pdf (accessed August 19, 2022).

Duettmann, Allison. WELCOME NEW PLAYERS | Gaming the Future. Foresight Institute. February 14, 2022. https://foresightinstitute.substack.com/p/new-players?s=r (accessed August 19, 2022).

Edison, Bill. Creating an AI red team to protect critical infrastructure. MITRE Corporation. September 2019. https://www.mitre.org/publications/project-stories/creating-an-ai-red-team-to-protect-critical-infrastructure (accessed August 19, 2022).

Etzioni, Oren. No, the Experts Dont Think Superintelligent AI is a Threat to Humanity. MIT Technology Review. September 20, 2016. https://www.technologyreview.com/2016/09/20/70131/no-the-experts-dont-think-superintelligent-ai-is-a-threat-to-humanity/ (accessed August 19, 2022).

Gary, Marcus, Ernest Davis, and Scott Aaronson. A very preliminary analysis of DALL-E 2. arXiv:2204.13807, 2022.

GCN Staff. NSF, NIST, DOD team up on resilient next-gen networking. GCN. April 30, 2021. https://gcn.com/cybersecurity/2021/04/nsf-nist-dod-team-up-on-resilient-next-gen-networking/315337/ (accessed May 1, 2022).

Jumper, John, et al. Highly accurate protein structure prediction with AlphaFold. Nature 596 (August 2021): 583589.

Kallenborn, Zachary. Swords and Shields: Autonomy, AI, and the Offense-Defense Balance. Georgetown Journal of International Affairs. November 22, 2021. https://gjia.georgetown.edu/2021/11/22/swords-and-shields-autonomy-ai-and-the-offense-defense-balance/ (accessed August 19, 2022).

Kegel, Helene. Understanding Gradient Descent in Machine Learning. Medium. November 17, 2021. https://medium.com/mlearning-ai/understanding-gradient-descent-in-machine-learning-f48c211c391a (accessed August 19, 2022).

Krakovna, Victoria. Specification gaming: the flip side of AI ingenuity. Medium. April 11, 2020. https://deepmindsafetyresearch.medium.com/specification-gaming-the-flip-side-of-ai-ingenuity-c85bdb0deeb4 (accessed August 19, 2022).

Littman, Michael L, et al. Gathering Strength, Gathering Storms: The One Hundred Year Study on Artificial Intelligence (AI100) Study Panel Report. Stanford University. September 2021. http://ai100.stanford.edu/2021-report (accessed August 19, 2022).

Manheim, David. Overoptimization Failures and Specification Gaming in Multi-agent Systems. Deep AI. October 16, 2018. https://deepai.org/publication/overoptimization-failures-and-specification-gaming-in-multi-agent-systems (accessed August 19, 2022).

Nguyen, Thanh Thi, and Vijay Janapa Reddi. Deep Reinforcement Learning for Cyber Security. IEEE Transactions on Neural Networks and Learning Systems. IEEE, 2021. 1-17.

Omohundro, Stephen M. The Basic AI Drives. Proceedings of the 2008 conference on Artificial General Intelligence 2008: Proceedings of the First AGI Conference. Amsterdam: IOS Press, 2008. 483492.

Panfili, Martina, Alessandro Giuseppi, Andrea Fiaschetti, Homoud B. Al-Jibreen, Antonio Pietrabissa, and Franchisco Delli Priscoli. A Game-Theoretical Approach to Cyber-Security of Critical Infrastructures Based on Multi-Agent Reinforcement Learning. 2018 26th Mediterranean Conference on Control and Automation (MED). IEEE, 2018. 460-465.

Pico-Valencia, Pablo, and Juan A Holgado-Terriza. Agentification of the Internet of Things: A Systematic Literature Review. International Journal of Distributed Sensor Networks 14, no. 10 (2018).

Pomerleu, Mark. US Army network modernization sets the stage for JADC2. C4ISRNet. February 9, 2022. https://www.c4isrnet.com/it-networks/2022/02/09/us-army-network-modernization-sets-the-stage-for-jadc2/ (accessed August 19, 2022).

Russell, Stewart. Human Compatible: Artificial Intelligence and the Problem of Control. New York: Viking, 2019.

Shah, Rohin. Reframing Superintelligence: Comprehensive AI Services as General Intelligence. AI Alignment Forum. January 8, 2019. https://www.alignmentforum.org/posts/x3fNwSe5aWZb5yXEG/reframing-superintelligence-comprehensive-ai-services-as (accessed August 19, 2022).

Shahar, Avin, and SM Amadae. Autonomy and machine learning at the interface of nuclear weapons, computers and people. In The Impact of Artificial Intelligence on Strategic Stability and Nuclear Risk, by Vincent Boulanin, 105-118. Stockholm: Stockholm International Peace Research Institute, 2019.

Trevino, Marty. Cyber Physical Systems: The Coming Singularity. Prism 8, no. 3 (2019): 4.

Turner, Alexander Matt, Logan Smith, Rohin Shah, Andrew Critch, and Prasad Tadepalli. Optimal Policies Tend to Seek Power. arXiv:1912.01683, 2021: 8-9.

Winder, Phil. Automating Cyber-Security With Reinforcement Learning. Winder.AI. n.d. https://winder.ai/automating-cyber-security-with-reinforcement-learning/ (accessed August 19, 2022).

Zeng, Andy, et al. Socratic Models: Composing Zero-Shot Multimodal Reasoning with Language. arXiv:2204.00598 (arXiv), April 2022.

Zewe, Adam. Does this artificial intelligence think like a human? April 6, 2022. https://news.mit.edu/2022/does-this-artificial-intelligence-think-human-0406 (accessed August 19, 2022).

Zwetsloot, Remco, and Allan Dafoe. Lawfare. Thinking About Risks From AI: Accidents, Misuse and Structure. February 11, 2019. https://www.lawfareblog.com/thinking-about-risks-ai-accidents-misuse-and-structure (accessed August 19, 2022).

[1] (Zewe 2022)

[2] (Littman, et al. 2021)

[3] (Jumper, et al. 2021)

[4] (Brown, et al. 2020)

[5] (Gary, Davis and Aaronson 2022)

[6] (Buchanan, et al. 2020)

[7] (Avatrade Staff 2021)

[8] (Russell 2019, 9-10)

[9] (Zwetsloot and Dafoe 2019)

[12] (Etzioni 2016)

[13] GO is an ancient Chinese strategy board game

[14] (Byford 2016)

[15] (Anderson 2019, 150)

[16] (Kegel 2021)

[17] (Krakovna 2020)

[18] (Omohundro 2008, 483-492)

[19] Ibid., 484.

[20] (Turner, et al. 2021, 8-9)

[21] (Baker, et al. 2020)

[22] Ibid.

[23] (Bostrom 2012, 71-85)

[24] (GCN Staff 2021)

[25] (Pomerleu 2022)

[26] (Berggren, et al. 2021)

[27] (Nguyen and Reddi 2021)

[28] Ibid.

[29] (Edison 2019)

[30] (Panfili, et al. 2018)

[31] (Winder n.d.)

[32] (Manheim 2018)

[33] Ibid.

[34] (Zeng, et al. 2022)

[35] (Drexler 2019, 18)

[36] Ibid.

[37] (Shah 2019)

[38] (Duettmann 2022)

[39] (Trevino 2019)

[40] (Pico-Valencia and Holgado-Terriza 2018)

See the article here:
Artificial Intelligence, Critical Systems, and the Control Problem - HS Today - HSToday

The French government has collected nearly 10 million in additional taxes after using machine learning to spot undeclared swimming pools in aerial photos. In France, housing taxes are calculated based on a propertys rental value, so homeowners who dont declare swimming pools are potentially avoiding hundreds of euros in additional payments.

The project to spot the undeclared pools began last October, with IT firm Capgemini working with Google to analyze publicly available aerial photos taken by Frances National Institute of Geographic and Forest Information. Software was developed to identify pools, with this information then cross-referenced with national tax and property registries.

The project is somewhat limited in scope, and has so far analyzed photos covering only nine of Frances 96 metropolitan departments. But even in these areas, officials discovered 20,356 undeclared pools, according to an announcement this week from Frances tax office, the General Directorate of Public Finance (DGFiP), first reported by Le Parisien.

As of 2020, it was estimated that France had around 3.2 million private swimming pools, but constructions have reportedly surged as more people worked from home during COVID-19 lockdowns, and summer temperatures have soared across Europe.

Ownership of private pools has become somewhat contentious in France this year, as the country has suffered from a historic drought that has emptied rivers of water. An MP for the French Green party (Europe cologie les Verts) made headlines after refusing to rule out a ban on the construction of new private pools. The MP, Julien Bayou, said such a ban could be used as a last resort response. He later clarified his remarks on Twitter, saying: [T]here are ALREADY restrictions on water use, for washing cars and sometimes for filling swimming pools. The challenge is not to ban swimming pools, it is to guarantee our vital water needs.

Frances tax offices, the DGFiP (known more commonly as Le Fisc), says it now plans to expand the use of its AI-pool-spotter to the entirety of metropolitan France (excluding the countrys overseas departments), which could net an additional 40 million in taxes.

Early reports on the project suggested that the machine learning software had an unusually high error rate of 30 percent, and regularly mistook other architectural features like solar panel installations for swimming pools. Now, though, Le Fisc says its ironed out these problems, and is looking to expand the use of its software spotting pools to identifying other undeclared and taxable housing improvements, like extensions and annexes.

We are particularly targeting house extensions like verandas, but we have to be sure that the software can find buildings with a large footprint and not the dog kennel or the childrens playhouse, Antoine Magnant, the deputy director general of public finances, told Le Parisien, reports The Guardian.

Read more from the original source:
French government uses AI to spot undeclared swimming pools and tax them - The Verge

Were you unable to attend Transform 2022? Check out all of the summit sessions in our on-demand library now! Watch here.

Over the last two years, one of the most common ways for organizations to scale and run increasingly large and complex artificial intelligence (AI) workloads has been with the open-source Ray framework, used by companies from OpenAI to Shopify and Instacart.

Ray enables machine learning (ML) models to scale across hardware resources and can also be used to support MLops workflows across different ML tools. Ray 1.0 came out in September 2020 and has had a series of iterations over the last two years.

Today, the next major milestone was released, with the general availability of Ray 2.0 at the Ray Summit in San Francisco. Ray 2.0 extends the technology with the new Ray AI Runtime (AIR) that is intended to work as a runtime layer for executing ML services. Ray 2.0 also includes capabilities designed to help simplify building and managing AI workloads.

Alongside the new release, Anyscale, which is the lead commercial backer of Ray, announced a new enterprise platform for running Ray. Anyscale also announced a new $99 million round of funding co-led by existing investors Addition and Intel Capital with participation from Foundation Capital.

MetaBeat 2022

MetaBeat will bring together thought leaders to give guidance on how metaverse technology will transform the way all industries communicate and do business on October 4 in San Francisco, CA.

Ray started as a small project at UC Berkeley and it has grown far beyond what we imagined at the outset, said Robert Nishihara, cofounder and CEO at Anyscale, during his keynote at the Ray Summit.

Its hard to understate the foundational importance and reach of Ray in the AI space today.

Nishihara went through a laundry list of big names in the IT industry that are using Ray during his keynote. Among the companies he mentioned is ecommerce platform vendor Shopify, which uses Ray to help scale its ML platform that makes use of TensorFlow and PyTorch. Grocery delivery service Instacart is another Ray user, benefitting from the technology to help train thousands of ML models. Nishihara noted that Amazon is also a Ray user across multiple types of workloads.

Ray is also a foundational element for OpenAI, which is one of the leading AI innovators, and is the group behind the GPT-3 Large Language Model and DALL-E image generation technology.

Were using Ray to train our largest models, Greg Brockman, CTO and cofounder of OpenAI, said at the Ray Summit. So, it has been very helpful for us in terms of just being able to scale up to a pretty unprecedented scale.

Brockman commented that he sees Ray as a developer-friendly tool and the fact that it is a third-party tool that OpenAI doesnt have to maintain is helpful, too.

When something goes wrong, we can complain on GitHub and get an engineer to go work on it, so it reduces some of the burden of building and maintaining infrastructure, Brockman said.

For Ray 2.0, a primary goal for Nishihara was to make it simpler for more users to be able to benefit from the technology, while providing performance optimizations that benefit users big and small.

Nishihara commented that a common pain point in AI is that organizations can get tied into a particular framework for a certain workload, but realize over time they also want to use other frameworks. For example, an organization might start out just using TensorFlow, but realize they also want to use PyTorch and HuggingFace in the same ML workload. With the Ray AI Runtime (AIR) in Ray 2.0, it will now be easier for users to unify ML workloads across multiple tools.

Model deployment is another common pain point that Ray 2.0 is looking to help solve, with the Ray Serve deployment graph capability.

Its one thing to deploy a handful of machine learning models. Its another thing entirely to deploy several hundred machine learning models, especially when those models may depend on each other and have different dependencies, Nishihara said. As part of Ray 2.0, were announcing Ray Serve deployment graphs, which solve this problem and provide a simple Python interface for scalable model composition.

Looking forward, Nishiharas goal with Ray is to help enable a broader use of AI by making it easier to develop and manage ML workloads.

Wed like to get to the point where any developer or any organization can succeed with AI and get value from AI, Nishihara said.

VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Learn more about membership.

Here is the original post:
Ray, the machine learning tech behind OpenAI, levels up to Ray 2.0 - VentureBeat

To quench the thirst for ever larger AI and machine learning models, Tesla has revealed a wealth of details at Hot Chips 34 on their fully custom supercomputing architecture called Dojo.

The system is essentially a massive composable supercomputer, although unlike what we see on the Top 500, it's built from an entirely custom architecture that spans the compute, networking, and input/output (I/O) silicon to instruction set architecture (ISA), power delivery, packaging, and cooling. All of it was done with the express purpose of running tailored, specific machine learning training algorithms at scale.

"Real world data processing is only feasible through machine learning techniques, be it natural-language processing, driving in streets that are made for human vision to robotics interfacing with the everyday environment," Ganesh Venkataramanan, senior director of hardware engineering at Tesla, said during his keynote speech.

However, he argued that traditional methods for scaling distributed workloads have failed to accelerate at the rate necessary to keep up with machine learning's demands. In effect, Moore's Law is not cutting it and neither are the systems available for AI/ML training at scale, namely some combination of CPU/GPU or in rarer circumstances by using speciality AI accelerators.

"Traditionally we build chips, we put them on packages, packages go on PCBs, which go into systems. Systems go into racks," said Venkataramanan. The problem is each time data moves from the chip to the package and off the package, it incurs a latency and bandwidth penalty.

So to get around the limitations, Venkataramanan and his team started over from scratch.

"Right from my interview with Elon, he asked me what can you do that is different from CPUs and GPUs for AI. I feel that the whole team is still answering that question."

Tesla's Dojo Training Tile

This led to the development of the Dojo training tile, a self-contained compute cluster occupying a half-cubic foot capable of 556 TFLOPS of FP32 performance in a 15kW liquid-cooled package.

Each tile is equipped with 11GBs of SRAM and is connected over a 9TB/s fabric using a custom transport protocol throughout the entire stack.

"This training tile represents unparalleled amounts of integration from computer to memory to power delivery, to communication, without requiring any additional switches," Venkataramanan said.

At the heart of the training tile is Tesla's D1, a 50 billion transistor die, based on TSMC's 7nm process. Tesla says each D1 is capable of 22 TFLOPS of FP32 performance at a TDP of 400W. However, Tesla notes that the chip is capable of running a wide range of floating point calculations including a few custom ones.

Tesla's Dojo D1 die

"If you compare transistors for millimeter square, this is probably the bleeding edge of anything which is out there," Venkataramanan said.

Tesla then took 25 D1s, binned them for known good dies, and then packaged them using TSMC's system-on-wafer technology to "achieve a huge amount of compute integration at very low latency and very-high bandwidth," he said.

However, the system-on-wafer design and vertically stacked architecture introduced challenges when it came to power delivery.

According to Venkataramanan, most accelerators today place power directly adjacent to the silicon. And while proven, this approach means a large area of the accelerator has to be dedicated to those components, which made it impractical for Dojo, he explained. Instead, Tesla designed their chips to deliver power directly though the bottom of the die.

"We could build an entire datacenter or an entire building out of this training tile, but the training tile is just the compute portion. We also need to feed it," Venkataramanan said.

Tesla's Dojo Interface Processor

For this, Tesla also developed the Dojo Interface Processor (DIP), which functions as a bridge between the host CPU and training processors. The DIP also serves as a source of shared high-bandwidth memory (HBM) and as a high-speed 400Gbit/sec NIC.

Each DIP features 32GB of HBM and up to five of these cards can be connected to a training tile at 900GB/s for an aggregate of 4.5TB/s to the host for a total of 160GB of HBM per tile.

Tesla's V1 configuration pairs of these tiles or 150 D1 dies in array supported four host CPUs each equipped with five DIP cards to achieve a claimed exaflop of BF16 or CFP8 performance.

Tesla's V1 Arrangement

Put together, Venkataramanan says the architecture detailed in depth here by The Next Platform enables Tesla to overcome the limitations associated with traditional accelerators from the likes of Nvidia and AMD.

"How traditional accelerators work, typically you try to fit an entire model into each accelerator. Replicate it, and then flow the data through each of them," he said. "What happens if we have bigger and bigger models? These accelerators can fall flat because they run out of memory."

This isn't a new problem, he noted. Nvidia's NV-switch for example enables memory to be pooled across large banks of GPUs. However, Venkataramanan argues this not only adds complexity, but introduces latency and compromises on bandwidth.

"We thought about this right from the get go. Our compute tiles and each of the dies were made for fitting big models," Venkataramanan said.

Such a specialized compute architecture demands a specialized software stack. However, Venkataramanan and his team recognized that programmability would either make or break Dojo.

"Ease of programmability for software counterparts is paramount when we design these systems," he said. "Researchers won't wait for your software folks to write a handwritten kernel for adapting to a new algorithm that we want to run."

To do this, Tesla ditched the idea of using kernels, and designed Dojo's architecture around compilers.

"What we did was we used PiTorch. We created an intermediate layer, which helps us parallelize to scale out hardware beneath it.Underneath everything is compiled code," he said. "This is the only way to create software stacks that are adaptable to all those future workloads."

Despite the emphasis on software flexibility, Venkataramanan notes that the platform, which is currently running in their labs, is limited to Tesla use for the time being.

"We are focused on our internal customers first," he said. "Elon has made it public that over time, we will make this available to researchers, but we don't have a time frame for that.

See the original post here:
Tesla wants to take machine learning silicon to the Dojo - The Register

Headquarters of the Council of Higher Scientific Research (CSIC).

a bone scan Every two years for all women aged 50-69. Since 1990, that is The biggest testing challenge for the national health systemAnd it aims to prevent one of the most common cancers in Spain, that is Mother, The method is X-rays that detect potentially cancerous areas; If something suspicious is found, that test is followed by more tests, often High probability of false positives, harmful and costly,

they are curvature This is the main reason why screening is limited to the highest risk groups. By adding predictive algorithms to mammograms, the risk areas of a patients breasts would be limited and the reliability of diagnosis increased to 90 percent. Therefore, they can be done with Often and the age range of the women they target Expansion,

It is a process that already exists, which uses artificial intelligenceand that . develops a team of Superior Council of Scientific Inquiry (CSIC), specifically the Institute of Corpuscular Physics (IFIC). it is part of the scope of machine learning (machine learning) in precision medicine, and a research network that seeks to increase the efficiency with which each patient is treated and optimize health care resources.

To understand how, you must first understand the concepts that come into play. The first is artificial intelligence. the ability of a computer or robot to perform tasks normally associated with intelligent beings, defined as sara degli-apostic You carlos sierra, author of the CSIC white paper on the subject. That is, they are the processes that are used replace human work with robotsWith the aim of accomplishing this with greater accuracy and greater efficiency.

And where can artificial intelligence work in medicine today? On many fronts, he replies. dolores del castilloResearchers from CSICs Center for Automation and Robotics, From the administrative to the management of clinical documentation. And, in a more specific way, in the analysis of images, or in the monitoring and follow-up of patients. And where are the still bigger limits? Above all, in the field of health care, in legal and ethical aspects when dealing with important matters. And whats more, theres still a long way to go, explains Del Castillo, who works on the projects, among others. neurological movement disorderTraining for a large section of healthcare workers.

We find the second concept as a subfield of artificial intelligence, along with its advantages and disadvantages: machine learning, This can be translated as machine learning. That is, artificial intelligence that works through computers thatand find patterns in population groups, With these patterns, predictions are made about what is most likely to happen. machine learning translate data Algorithm,

Precision medicine to predict disease

and after artificial intelligence and machine learningThere is a third concept: the precision medicine, The one that suits the person, his genes, his background, his lifestyle, his socialization. a model that must first be able predictable disease, Second, Francisco Albiol from IFIC, continues to assess each patient, apply the best treatment based on clinical evidence, identify the most complex cases, and assess their inclusion in management programs.

It makes sense high impact disease, and does not make sense for serious diseases; For example, distinguishing the flu from a cold in primary care, as the benefits will not compensate for the effort required.

The key to the use of artificial intelligence in medicine is also cost optimization, which is very important for public health. Spains population has increased from 42 to 47 million people between 2003 and 2022, that is, more than 10 percent. and from 2005 to 2022, The average age of the population has increased from 40 to 44, We are getting older and older.

Therefore, Dolores del Castillo says, the best valued projects and, therefore, likely to be funded, are those that incorporate artificial intelligence techniques to address the prevention, diagnosis and treatment of cardiovascular diseases, neurodegenerative diseases, cancer and obesity. There is also a special focus on personal and home medicine, elderly care, and new drug offerings. The need for healthcare has been heightened by our demographics, and The aim should be to reduce and simplify the challenges with technology, we tried machine learning, summarizes Albiol.

Albiol is one of the scientists who led a program to improve breast cancer detection through algorithms. He defends, like other researchers, that if we mix machine learning with precision medicine, we should be talking about 4p medicine. Which brings together four features: Predictive, personal, preventive and participatory,

Because most purists confine precision medicine to the field of patient genetics, and would not include it in the bag that takes more characteristics into account. Those who do say that we are talking about something much broader: Applied to precision medicine, machine learning allows for Analyze large amounts of very different types of data (genomic, biochemical, social, medical imaging) and model them to be able to offer together individual diagnosismore precise and thus more effective treatment, summarizes researcher Lara Loret Iglesias of the Institute of Physics of Cantabria.

Lloret is part of a network of scientists who, like Albiol or Del Castillo, are dedicated to projects on machine learning and precision medicine. One of them developed by his team, which he leads together with fellow physicist Miriam Kobo Cano, is called Branyas. It is in honor of Spains oldest woman, Maria Branyas, who managed to overcome Covid-19: she has done so at the age of 113. In this they bring together the many casuistries of more than 3,000 elderly people, much less just genetics: machine learning establish Risk profile of getting sick or dying as a result of coronavirus, We derived data from the analysis of three risk profiles: a sociodemographic, a biological and an extended biological, which will add information on issues such as aspects related to the intestinal microbiota, vaccination and immunity.

Precision Medicine, Cancer and Alzheimers

also explain this Joseph Lewis Arcosfrom the Artificial Intelligence Research Institute. common diseases There are cancer and Alzheimers linked to precision medicine, but they have stood out with the Ictus project. Launched in the middle of a pandemic (which has made things difficult, he admits), he has treated patients at Barcelonas Belwitz Hospital who suffered strokes and, after a severe and acute phase, Have become long term,

In particular, those with movement difficulty in one hand or both. made over 700 sessions In which patients have been asked to play the keyboard of the electronic piano. Then, they transferred the analysis of finger movements to the results to see what the patterns of difficulties and improvements are. And theyve gotten particularly positive feedback among users because its not only doing an exercise, but it affects a very emotional part. The goal is now to expand it to hospitals in the United Kingdom.,

and future? Dolores del Castillo replies, I believe that the challenge of artificial intelligence in medicine is to incorporate research results into daily practice in a generalized way, but always without forgetting that it is the experts who have is the last word. To do that, doctors need to be able to rely on these systems and Interact with them in the most natural and simple wayEven helping with its design.

Lara Loret believes that we have to be able to build generalizable prediction systems, that is, the efficiency of the model does not depend on unnecessary things such as which machine the data is taken in, or how the calibration is. Francisco Albiol focuses on a problem that may be in the long run must have a solutionAt present, larger hospitals are preferred in these technologies than smaller cities or towns. convenience and reduce costs It also has to do with reaching out to everyone.

While it may include statements, data or notes from health institutions or professionals, the information contained in medical writing is edited and prepared by journalists. We advise the reader to consult a health professional on any health-related questions.

Read the rest here:
'Machine Learning' Predicts The Future With More Reliable Diagnostics - Nation World News