Podcast
Wharton Ripple Effect
Rise of AI: How Do We Coexist With Algorithms?
THIS EPISODE Professor Kartik Hosanagar discusses how AI is impacting the world as we know it and how we can keep up.
Knowledge at Wharton is the free online journal of the Wharton School of the University of Pennsylvania.
New to me. Have a strong interest in bird flight and further behavior. Biomimicry
ACM TECHNEWS
Seagull Algorithms Could Hide Secret to Greener Cloud Computing, By Fast Company, April 5, 2023
Modeling the motions of seagulls.
The researchers say that the way seagulls behave when they’re on the hunt for food or prey is one of the most ruthlessly efficient examples of an entity zeroing in on its target, with minimal excess energy expenditure.
British, Chinese, and Austrian researchers say that more sustainable cloud computing systems can be achieved by algorithmically mimicking seagulls' hunting and migration behavior.
The researchers contend a "seagull optimization algorithm" can slash cloud computing's power consumption by 5.5% and its network traffic by 70%.
They said the meta-heuristic algorithm would determine the best locations for virtual machines within a server communications network to optimize network traffic efficiency.
This would mirror the seagulls' ability to stalk and plot a route to prey without colliding with each other.
The researchers believe seagull-based algorithms could reduce the total number of physical supercomputers worldwide, while also cutting the machine-to-machine communications network's power consumption by 80%.
From Fast Company
View Full Article
Safety is yes, means of determining that is the issue. Especially with regard to transportation and related supply chain decisions and risks.
Do We Need a National Algorithms Safety Board?
By The Hill, February 28, 2023, Opinion Piece
A close-up of hands on a laptop keyboard, with image overlays of various tech-based iconography.
Perfectly safe algorithmic systems are not possible, but safer systems are.
In the U.S., the National Transportation Safety Board is widely respected for its prompt investigations of plane, train, and boat accidents. Could a National Algorithms Safety Board have a similar impact in increasing safety for algorithmic systems, especially the rapidly proliferating artificial intelligence applications based on unpredictable machine learning? Alternatively, could agencies such as the Food and Drug Administration, Securities and Exchange Commission, or Federal Communications Commission take on the task of increasing safety of algorithmic systems?
From The Hill View Full Article
Amazes me, was the first thing I learned in engineering analytics methodology to optimize systems Originally invented for military supply chains in WWII. Here a technical overview and good historical piece.
Why the Simplex Method, at Age 75, is Still the Go-To Algorithm? By Allyn Jackson, Commissioned by CACM Staff, January 31, 2023
ACM NEWS, January 31, 2023, Daniel A. Spielman
Since its birth more than two decades ago, smoothed analysis has been used to analyze the performance of algorithms other than the simplex method, including interior-point methods for linear programming. It also has guided the design of new algorithms.
In 1947, mathematical scientist George Dantzig invented the simplex method, a powerful and practical means to find solutions to linear programming for optimization problems. Scientists lost no time putting the simplex method to use in a variety of applications across government, industry, science, and engineering.
Half a century later, when Daniel A. Spielman was a Ph.D. student at the Massachusetts Institute of Technology, the simplex method stood at the top of the pantheon of important algorithms, yet research had shown the simplex method had proven pitfalls; it ought not to perform as well as it did. What was going on?
Spielman solved the mystery in 2001, in joint work with Shang-hua Teng, now University Professor and Seeley G. Mudd Professor of Computer Science and Mathematics in the Viterbi School of Engineering at the University of Southern California. Pioneering a technique called smoothed analysis, their work provided an original and compelling explanation of the power of the simplex method and suggested a new paradigm for gauging the effectiveness of algorithms.
This work was recognized in 2008 by the Gödel Prize, sponsored jointly by ACM's Special Interest Group on Algorithms and Computing Theory (SIGACT) and the European Association for Theoretical Computer Science. Now Spielman, Sterling Professor of Computer Science and a professor of Statistics and Data Science and of Mathematics at Yale University, has been awarded the 2023 Breakthrough Prize in Mathematics, in part for this work in optimization.
Spielman has a great ability "to come up with new approaches," said Lance Fortnow, dean of the College of Computing at the Illinois Institute of Technology. "It wasn't like someone else had invented smoothed analysis and Spielman said, 'Let's try it on linear programming'. This was, 'I want to understand linear programming. What's the right way to do it?'"
Simplex Bests Polynomial-time Competition
Dantzig formulated the concept of linear programming as a way to model optimization problems. The model produces a polyhedron, possibly of very high dimension, where the corners represent solutions. The simplex method provides a highly efficient way of moving along polyhedron edges toward an optimal corner. The algorithm was tailor-made for the computing machines that were just beginning to appear when Dantzig did this work.
In the 1970s, the rise of complexity theory brought new precision to the study of efficiency of algorithms. Generally, an algorithm is considered efficient if it runs in polynomial time, meaning that even for the worst-case inputs to the algorithm, the running time is always bounded by a polynomial function of the input size. This is in contrast with algorithms whose running time can increase exponentially with input size.
Soon a curious fact arose: despite its excellent performance in practice, the simplex method is not running in polynomial time. Examples were found on which simplex ran in exponential time. Eventually, polynomial-time algorithms for linear programming were found, but the simplex method continued to be used — and in many situations, outperformed its polynomial-time competitors.
Why does simplex work so well?
This question was in the air when Fortnow was a Ph.D. student at the Massachusetts Institute of Technology (MIT) in the 1980s, several years before Spielman became a graduate student there. However, he recalls, few were attempting to resolve it. "The simplex method had already been around for 40 years," said Fortnow. The general attitude was, "Well, it works well in practice."
When Spielman and Teng came up with smoothed analysis, it was a big surprise. "It's a whole different way of looking at worst-case complexity," said Fortnow, "and they could apply it to linear programming. Both of these pieces were very exciting." ... '
Games, Models and evolutionary Algorithms.
ACM OPINION
Open-Endedness and Evolution through Large Models
By The Gradient, September 23, 2022
Joel Lehman is a machine-learning scientist interested in AI safety, reinforcement learning, and creative open-ended search algorithms. He is the co-author of the book Why Greatness Cannot be Planned: The Myth of the Objective.
In an interview, Lehman talks about the move from game development to AI, evolutionary algorithms, neuroevolution through augmenting topologies (NEAT), LLMs as practical thought experiments in disembodied understanding, evolution through large models (ELM), competition in AI, advice for people considering ML research, and more.
Overblown.
The Age of Algorithmic Anxiety
By The New Yorker, July 29, 2022
It can feel as though every app is trying to guess what you want before your brain has time to come up with its own answer.
Social networks have always purported to show us things that we like—things that we might have organically gravitated to ourselves. Why, then, can it feel as though the entire ecosystem of content that we interact with online has been engineered to influence us in ways that we cannot quite parse, and that have only a distant relationship to our own authentic preferences?
We are constantly negotiating with the pesky figure of the algorithm, unsure how we would have behaved if we had been left to our own devices.
From The New Yorker
How broadly might Quantum computing realistically improve our algorithmic capabilities Below a short intro.
Quantum computer succeeds where a classical algorithm fails
Quantum computers coupled with traditional machine learning show clear benefits.
By JOHN TIMMER in Arstechnica
People have performed many mathematical proofs to show that a quantum computer will vastly outperform traditional computers on a number of algorithms. But the quantum computers we have now are error-prone and don't have enough qubits to allow for error correction. The only demonstrations we've had involve quantum computing hardware evolving out of a random configuration and traditional computers failing to simulate their normal behavior. Useful calculations are an exercise for the future.
But a new paper from Google's quantum computing group has now moved beyond these sorts of demonstrations and used a quantum computer as part of a system that can help us understand quantum systems in general, rather than the quantum computer. And they show that, even on today's error-prone hardware, the system can outperform classical computers on the same problem.
Probing quantum systems
To understand what the new work involves, it helps to step back and think about how we typically understand quantum systems. Since the behavior of these systems is probabilistic, we typically need to measure them repeatedly. The results of these measurements are then imported into a classical computer, which processes them to generate a statistical understanding of the system's behavior. With a quantum computer, by contrast, it can be possible to mirror a quantum state using the qubits themselves, reproduce it as often as needed, and manipulate it as necessary. This method has the potential to provide a route to a more direct understanding of the quantum system at issue.
Much of the paper is devoted to describing situations where this should be the case, in part elaborating on ideas described in earlier papers. .... '
Interesting thing, to actually delete an algorithm. I don't think of algorithms as unique. I guess then a copy or equivalent algorithm could also be thus condemned. And in the case given the algorithm has been derived illegally from data without permission.
The FTC's New Enforcement Weapon Spells Death for Algorithms
By Protocol March 23, 2022
The U.S. Federal Trade Commission has struggled over the years to find ways to combat deceptive digital data practices using its limited set of enforcement options. Now, it's landed on one that could have a big impact on tech companies: algorithmic destruction. And as the agency gets more aggressive on tech by slowly introducing this new type of penalty, applying it in a settlement for the third time in three years could be the charm.
In a March 4 settlement order, the agency demanded that WW International—formerly known as Weight Watchers—destroy the algorithms or AI models it built using personal information collected through its Kurbo healthy eating app from kids as young as 8 without parental permission. The agency also fined the company $1.5 million and ordered it to delete the illegally harvested data.
When it comes to today's data-centric business models, algorithmic systems and the data used to build and train them are intellectual property, products that are core to how many companies operate and generate revenue. While in the past the FTC has required companies to disgorge ill-gotten monetary gains obtained through deceptive practices, forcing them to delete algorithmic systems built with ill-gotten data could become a more routine approach, one that modernizes FTC enforcement to directly affect how companies do business. ....
From Protocol View Full Article
University of Bern (Switzerland), November 10, 2021
An international team of scientists led by Switzerland's University of Bern has developed evolutionary algorithms that can learn creatively. Such algorithms determine the "fitness" of a candidate solution based on how well it solves the underlying problem. The researchers' evolving-to-learn (E2L) or "becoming adaptive" approach was applied to three typical learning scenarios. The first was to detect a repeating pattern in a continuous input stream without performance feedback; the second virtually rewarded the computer for behaving in a desired manner; and the third guided the computer on how much its behavior diverged from a desired pattern. “In all these scenarios,” Bern's Jakob Jordan said, "The evolutionary algorithms were able to discover mechanisms of synaptic plasticity, and thereby successfully solved a new task." .... '
Watched this closely, family being involved in the space, Prices seemed odd, but were they somehow magical? Always compare algorithm results for reasonableness. In particular, prediction in varying contexts is always suspect.
What Went Wrong With Zillow? A Real Estate Algorithm Derailed Its Big Bet By The Wall Street Journal, November 19, 2021
Real estate firm Zillow Group had looked to its digital home-flipping business Zillow Offers to lead its growth in the future, but the company has acknowledged that will not happen because the unit’s underlying algorithm could not reliably predict housing prices.
That failure, the company said, was rooted in the technology's inability to understand the real estate market and predict housing prices, which are shaped by fluctuating factors like aesthetics and regional factors that influence buyers' decisions.
Zillow CEO Richard Barton admitted to shareholders that the algorithm could not accurately predict swings in home prices, and the company is closing Zillow Offers. ....
Perhaps instructive, or what might be expected.
What I learned surrendering my life to algorithms in Cnet.
I outsourced several of my daily decisions to algorithms for a week, to better understand the lines of code playing a greater and greater role in guiding our lives.
BY Jesse Orrall
The tree of life was dying, and we had run out of ideas on how to save it. We had tried everything we could think of and were hopelessly stuck when a voice chirped through a walkie-talkie, informing us that the mouse in the corner could give us a hint.
As I awaited instructions from a fake mouse in an escape room above a KFC, I couldn't help but think this was a fitting conclusion to my weeklong experiment surrendering my life to algorithms.
If our computers are a window to the online world, algorithms are key mediators whose intervention clearly impacts what we see, and therefore what we do. They filter our search results (for example: Google, Bing), curate our social media feeds (Facebook, Twitter) and make recommendations for new products and experiences (Amazon, Yelp).
For internet users, these algorithms can help us sort through the mass of information at our fingertips. For the companies that develop them, algorithms are tools for amassing valuable data, streamlining the shopping and browsing experience and encouraging us to spend more money. Algorithms are also becoming important tools in the high-stakes arenas of workplace management, financial investment and policing.
Despite algorithms' centrality to the online experience (and increasingly our real-world experience) their inner workings are largely a mystery, since they're the closely guarded intellectual property of the companies that use and develop them. In an effort to better understand algorithms, how well they work and how well they know me, I embarked on an experiment to surrender my life to algorithms for a week. This might not be what the coders had in mind, but that didn't stop me from trying, or from gleaning some helpful insights along the way. ... '
Heres an interesting thing, a kind of algorithm synthesis to do AI with? Reading, Warning: Technical. Note, this is research, but I am intrigued.
Deriving Equations from Sensor Data Using Dimensional Function Synthesis By Vasileios Tsoutsouras, Sam Willis, Phillip Stanley-Marbell
Communications of the ACM, July 2021, Vol. 64 No. 7, Pages 91-99 10.1145/3465216
We present a new method for deriving functions that model the relationship between multiple signals in a physical system. The method, which we call dimensional function synthesis, applies to data streams where the dimensions of the signals (e.g., length, mass, etc.) are known. The method comprises two phases: a compile-time synthesis phase and a subsequent calibration using sensor data. We implement dimensional function synthesis and use the implementation to demonstrate efficiently summarizing multimodal sensor data for two physical systems using 90 laboratory experiments and 10,000 synthetic idealized measurements. The results show that our technique can generate models in less than 300 ms on average across all the physical systems we evaluated. This is a marked improvement when compared to an average of 16 s for training neural networks of comparable accuracy on the same computing platform. When calibrated with sensor data, our models outperform traditional regression and neural network models in inference accuracy in all the cases we evaluated. In addition, our models perform better in training latency (up to 1096X improvement) and required arithmetic operations in inference (up to 34X improvement). These significant gains are largely the result of exploiting information on the physics of signals that has hitherto been ignored.
1. Introduction
Physical systems instrumented with sensors can generate large volumes of data. These data are useful in understanding previous behaviors of the systems that generate them (e.g., monitoring properties of components in aircraft) as well as in predicting future behaviors of those systems (e.g., predicting failures of components in machinery). Unlike data sources such as speech or text, data from sensors of physical phenomena must obey the laws of physics. Existing methods for constructing predictive models from sensor data however do not fully exploit prior knowledge of the physical interpretation of sensor data. In this work, we use information about physical dimensions of sensor signals to synthesize compact predictive models from sensor data. In keeping with the convention in physics, we use the term dimensions to refer to quantities such as length or time and we use the term units to refer to a value in a standardized system for quantifying values of a given dimension, such as centimeters or miles for length and Pascals or mmHg for pressure. The state of the art in deriving models from such data streams is to apply some form of machine learning.11, 19 Blindly applying machine learning to data from physical systems however ignores important prior knowledge about the physical implications of the signals. ... '
Like to see this and it can control model implications.
Platform Teaches Nonexperts to Use ML
Cornell Chronicle, Louis DiPietro, July 28, 2021
An interactive machine learning (ML) platform developed by Cornell University scientists is designed to train nonexperts to use algorithms effectively, efficiently, and ethically. Cornell's Swati Mishra said, "If we design machine learning tools correctly and give enough agency to people to use them, we can ensure their knowledge gets integrated into the machine learning model." Said Cornell's Jeff Rzeszotarski, "While our eventual goal is to help novices become advanced machine-learning users, providing some 'training wheels' through transfer learning can help novices immediately employ machine learning for their own tasks." Added Mishra, “We as researchers and designers have to mitigate user perceptions of what machine learning is. Any interactive tool must help us manage our expectations.”
Welcome to Algorithmic Impact Assessments. I should note that I have done optimization and statistical models for decades, and this has very, very rarely come up. And then only when included in a risk analysis that indicated potential errors being made. Will all business process be under detailed scrutiny? Article linked to below is interesting.
Assembling Accountability, from the Ground Up in Point.datasociety.net
Algorithmic impact assessments should leverage diverse expertise & complex histories
By Emanuel Moss, Ranjit Singh, Elizabeth Anne Watkins, and Jacob Metcalf
By now, it is a tired trope: Sen. Orrin Hatch asking Mark Zuckerberg how Facebook makes money. Zuckerberg replying with a wry “Senator, we run ads.” Another congressperson grilling Sundar Pichai, CEO of Google, about the Apple iPhone, made by Apple.
Lawmakers, it is said, don’t understand technology well enough to regulate it. They are too old. They are out of touch. They have disinvested from staff and other experts that could help them understand it. And while all those criticisms may be true, why should we expect our lawmakers to become individual experts on every challenge facing society? There are advocacy groups, community activists, forensic technologists, thoughtful developers, and critical scholars who have devoted their careers to building expertise on these issues. Given the burgeoning influence of algorithmic systems over social affairs, and an increasing awareness of the harmful impacts of these powerful systems, we are at a moment in which complex sociotechnical systems require robust, adaptable regulation — and legislatures and regulatory bodies are drafting new rules.
Our new report Assembling Accountability demonstrates a pressing need to establish algorithmic impact assessment practices from the ground up, which requires cultivating and synthesizing a broad consensus of expertise from industry, scholars, and public interest advocates, including people from affected communities. ... '
Always looking at data and context involved when algorithms fail. Like any kind of predictive approach, its rarely perfect. The same for classing analytics and for machine learning based methods. And for that matter for human predictive methods as well. And that can change as data and context changes over time. So these kinds of test examples are useful.
Algorithm That Predicts Deadly Infections Is Often Flawed By Wired
A study using data from nearly 30,000 patients in University of Michigan hospitals suggests Epic Systems' early warning system for sepsis infections performs poorly.
An algorithm designed by U.S. electronic health record provider Epic Systems to forecast sepsis infections is significantly lacking in accuracy, according to an analysis of data on about 30,000 patients in University of Michigan (U-M) hospitals.
U-M researchers said the program overlooked two-thirds of the approximately 2,500 sepsis cases in the data, rarely detected cases missed by medical staff, and was prone to false alarms.
The researchers said Epic tells customers its sepsis alert system can correctly differentiate two patients with and without sepsis with at least 76% accuracy, but they determined it was only 63% accurate.
U-M's Karandeep Singh said the study highlights wider shortcomings with proprietary algorithms increasingly used in healthcare, noting that the lack of published science on these models is "shocking."
From Wired
Algorithm, Malgorithm, Jalgorithm? Fairness in Judgement.
Right, algorithm just means a standardized method. Like iany tool can be used or misused. Good piece here.
Misnomer and Malgorithms By Robin K. Hill in CACM
In response to a previous piece on the articulation of design responsibility [Hill2018], by which I mean the egregious practice of casually attributing judgment and volition to programs, I've received some comments. My view is that the attribution, in our locutions, of decision-making power to certain applications of programs and algorithms is wrong in both senses of "wrong"—both false and harmful.
The most obvious, and most misleading, instance of malarticulation is the trending use of "algorithm". One or two comments mentioned the common modern use of that word to mean an agent that makes (bad) judgments, giving rise to claims that that technology is not value-neutral. The concern is valid but the connotation hangs on context, and the implications of the literal assertion are dangerous. "Oh, well, sure," educated people will say, "We agree that tech is technically neutral." Yes, it's technically neutral. In fact, technically, it's nothing more than technical, and therefore nothing more than neutral.
This needs to be cleared up. Computer science knows the algorithm as an objective computational object, breathtaking and beautiful, an abstract imperative structure (so I claim [Hill2016]), deterministic and independent of context. I will call this objective procedure, a mechanism that performs calculations under a decision structure, the i-algorithm; maybe we can think of the i as "imperative structure". But the public knows the algorithm as a mysterious agent making dubious decisions, a source of judgments, supposed to be reasonable, on complex issues in real life. I will call this subjective procedure the j-algorithm; we can think of the j as "judge". These are homonyms but not synonyms, and we understand that. Computer scientists, told that an i-algorithm is political, simply code-switch to the homonym j-algorithm, the thing that assesses parole requests and loan applications (poorly), in order to continue the communication. This communication infelicity is not new—scientists have to put up with "bug", "exponential", "schizo", and other abuses of terminology. The problem with "algorithm" is that the two senses of the word are, in a way, contradictory, and in exactly the way that matters. .. "
This came up in discussion this week .. the algorithms are classic approaches. But the algorithms are adaptive, so I would expect them to be continually adjusted. And if they are maintaining them, as I always suggest ....
COVID-19 Wrecked the Algorithms That Set Airfares, but They Won't Stay Dumb
The Wall Street Journal, Jon Sindreau, May 17, 2021
The COVID-19 pandemic crippled the reliability of algorithms used to set air fares based on historical data and has accelerated a hybrid model that combines historical and live data. Before the pandemic, airlines used the algorithms to predict how strong ticket demand would be on a particular day and time, or exactly when people will fly to visit relatives before a holiday. Corporate travel constitutes a large share of airline profits, with business fliers avoiding Tuesdays and Wednesdays, favoring short trips over week-long ones, and booking late. The pandemic undermined historical demand patterns while cancellations undercut live data, causing the algorithms to post absurd prices. Overall, the pandemic has stress-tested useful advancements to the algorithms, like assigning greater weight to recent booking numbers, and applying online searches to forecast when and where demand will manifest. ... '
Somewhat different approach, training from equations. Closer to the efforts we used during the early uses of neural nets. Combining learning methods and existing algorithms.
Neural Nets Used to Rethink Material Design
Rice University News, April 30, 2021
A technique developed by researchers at Rice University and Lawrence Livermore National Laboratory uses machine learning to predict the evolution of microstructures in materials. The researchers demonstrated that neural networks can train themselves to predict a structure's growth in a particular environment. The researchers trained their neural networks using data from the traditional equation-based approach to predict microstructure changes and tested them on four microstructure types: plane-wave propagation, grain growth, spinodal decomposition, and dendritic crystal growth. The neural networks were 718 times faster for grain growth when powered by graphic processors compared to the prior algorithm, and 87 times faster when run on a standard central processor. Rice's Ming Tang said the new method can "make predictions even when we do not know everything about the material properties in a system," and will be useful in designing more efficient batteries.
Useful nontechnical piece, some good thoughtful graphics. Yes a clear and increasing issue that needs to be regulated. But it is not true that such systems and their developers are not considering accountability. Been the case long before emergence of sexier approaches like AI. Starting in statistics and Big Data. Been involved in many of these applications. via Walter Riker.
Moving fast and breaking us all: Big Tech’s unaccountable algorithms
By Ellery Roberts Biddle & Jie Zhang in RankingDigitalRights
They decide who passes and who fails in secondary school. They decide who gets arrested and who goes to prison. They decide what news you see first thing in the morning as well as what news you won’t see. And they drive the business models—and revenues—of the world’s largest and most powerful digital platforms.
In spite of all this, many of the world’s most powerful algorithms are accountable to no one—not even the companies that build and deploy them. Some have made vague pledges to “be ethical,” and for all we know, there may be strong policies or rules that the companies follow behind closed doors. But the overall lack of public explanation of how these systems are built and run indicates that companies do not have oversight over how their own systems work. In light of the enormous effects that they have on human rights, public health, public safety, democracy, and our understanding of reality, this is nothing short of reckless.
For the 2020 RDR Index, we looked for companies’ answers to some fundamental questions about algorithms: How do you build and train them? What do they do? What standards guide these processes?
We combed the public-facing documentation and, to no surprise, found very, very little. Yet companies are harvesting user data by the minute, to fuel algorithmic optimization, engagement, and personalization—all things that translate to enormous profits..... "
Coloring your designs to avoid failures.
Merging Technologies with Color to Avoid Design Failures
Penn State College of Engineering, Miranda Buckheit, January 21, 2021
Pennsylvania State University (Penn State) researchers analyzed machine learning (ML) and image colorization algorithms to avoid design failures. Penn State's Pranav Milind Khanolkar reviewed the use of the ABAQUS additive-manufacturing simulation software, which can pose difficulties because its speed and performance level rely on a computer's hardware processing power. The team deployed ML algorithms to lower the exclusive use of computationally demanding finite element analysis, and accelerate simulations. It then applied image colorization algorithms to material microstructure data, and repurposed programs typically used to add color to monochrome photos. Khanolkar said, "Using intelligent technology to help people and empower their creativity and empathy during the design process is important. These algorithms need lots of computational power and using artificial intelligence ... allows designers to be more creative without impacting production cost."
