Let a Thousand AIs Bloom

Home/News/'Let a Thousand AIs Bloom'/Full Text

ACM NEWS

'Let a Thousand AIs Bloom'    By Bennie Mols

Commissioned by CACM Staff   May 4, 2023

Data science and philosophy professor David Danks.

"Computer science students don’t need to become ethicists, and philosophy students don’t need being able to write code, but we need to teach them how to collaborate and understand each other."

Credit: DavidDanks.org

The field of artificial intelligence (AI) has been dominated by the deep learning approach in recent years, and there is some concern that focus may be limiting progress in the field. David Danks, a professor of data science and philosophy at the University of California, San Diego, advocates for more diversity in AI research or, as he puts it, "let a thousand AIs bloom." .... 

Bennie Mols interviewed Danks at the 2023 AAAS Annual Meeting in Washington, D.C.

What has led you to the conclusion there is too little diversity in the AI field?

We have seen enormous advances in the ability of AI, and in particular deep learning, to predict, classify, and generate what we might think of as the surface features of the world. These successes are marked by two fundamental features that don't always hold: having a measurement of what matters, and being able to define what counts as success. Deep learning can do amazing things, but what worries me is that it crowds everything else out.

Such as…

We have struggled to come up with AI systems that can discover the underlying structure of the world, things that show up in the data but are not defined by them. So one reason that we are struggling with developing more trustworthy and value-centered AI is because trust and values fundamentally are not things that we know how to give numerical expressions for.

Can you give an example?

It is difficult to figure out what counts as success for a self-driving car. Sure, we want to drive safely, but what counts as driving safely is very context-dependent. It depends on social norms, it depends on the weather, it depends on suddenly occurring situations on the road. As soon as there is an unusual context, self-driving cars can't reason their way out like a human driver can.

What is your proposal for increasing the diversity of AI research?

First, people need to realize that there are problems we are not considering because of the focus on deep learning. Deep learning is not good at symbolic reasoning, not good at planning, not good at reconciling conflicts between multiple agents that have different values. We need to let a thousand AIs bloom because we need different frameworks and we need to look at problems we are not considering because of the focus on deep learning.

Second, funding agencies in particular should be supporting the work that companies don't want to support. Right now, most companies are putting most effort into deep learning.

Third, I also think that there is an enormous opportunity for entrepreneurs to identify problems that deep learning is not going to solve and come up with new methods and new systems. If I were an entrepreneur, I would stay far away from deep learning, because I am not going to compete with the big tech companies.

How do you, as a philosopher, look at the recent hype around ChatGPT and similar large language models?

For me, the most interesting aspect is that ChatGPT is calling into question how deep a lot of our human conversation actually is. So much of human language seems to be highly predictable or ritualized; anybody who has ever taught classes for some years knows this. There are times when you just go in the classroom and you start talking on autopilot. I came to realize that my own speech is not nearly as profound as I might have thought it was.

Does ChatGPT have consequences for your way of teaching?

I put all the assignments for my classes through ChatGPT to see how it performs, and it did very badly. ChatGPT is particularly good at giving a Wikipedia-level summary of a topic, but it is bad at reasoning, drawing inferences, logical conclusions, and constructing good arguments. ChatGPT might actually push teachers to make better assignments by avoiding assignments that it can answer well. ... 

Bennie Mols is a science and technology writer based in Amsterdam, the Netherlands.

AI Addressing Dangerous Solar Storms

Am a long time follower of this very dangerous and inevitable natural event. 

Deep Learning to Address Impact of Solar Storms.

NASA AI model could help world prepare for impact of solar storms

DAGGER artificial intelligence model uses NASA data to produce swift predictions, within 30 minutes of detection.

By Julia Musto    https://www.nasa.gov/feature/goddard/2023/sun/nasa-enabled-ai-predictions-may-give-time-to-prepare-for-solar-storms

NASA said Thursday that a new computer model that combines artificial intelligence and agency satellite data could help prepare for dangerous space weather. 

The model, called DAGGER (Deep Learning Geomagnetic Perturbation), uses the technical tool to analyze spacecraft measurements of the solar wind and forecast where an impending solar storm will strike on Earth – with 30 minutes of advance warning. 

An international team of researchers at the Frontier Development Lab said the model can produce predictions in less than a second, with predictions updating every minute. 

The lab is a partnership that includes NASA, the U.S. Geological Survey and the Department of Energy.

NASA's Solar Dynamics Observatory captured this image of a solar flare on Oct. 2, 2014. The solar flare is the bright flash of light at top. A burst of solar material erupting out into space can be seen just to the right of it.

NASA's Solar Dynamics Observatory captured this image of a solar flare on Oct. 2, 2014. The solar flare is the bright flash of light at top. A burst of solar material erupting out into space can be seen just to the right of it. (Credits: NASA/SDO)

The scientists had used A.I. to look for links between the solar wind and geomagnetic interruptions, applying a method called "deep learning" that trains computers to recognize patterns based on previous examples.

The model was tested against previous geomagnetic storms from August 2011 and March 2015, with DAGGER accurately forecasting the storm's impacts. Previously, models had used A.I. to forecast for specific locations, but NASA said DAGGER is the first to combine A.I. with real measurements to generate frequent and precise predictions worldwide.

"With this AI, it is now possible to make rapid and accurate global predictions and inform decisions in the event of a solar storm, thereby minimizing – or even preventing – devastation to modern society," Vishal Upendran of the Inter-University Center for Astronomy and Astrophysics in India, who is the lead author of a paper about the DAGGER model published in the journal Space Weather, said in a statement.  ... '   (Images at the link)   (plan to look at nature and goals of the model) 

The Premature Obituary of Programming

Much yet to do here, but I would expect this would ultimately take over many such capabilities.    Especially when integrated with automated testing for threat weaknesses, which could be updated to changing contexts.  

The Premature Obituary of Programming  (abstract)     By Daniel M. Yellin

Communications of the ACM, February 2023, Vol. 66 No. 2, Pages 41-44  10.1145/3555367

Deep learning (DL) has arrived, not only for natural language, speech, and image processing but also for coding, which I refer to as deep programming (DP). DP is used to detect similar programs, find relevant code, translate programs from one language to another, discover software defects, and to synthesize programs from a natural language description. The advent of large transformer language models10 is now being applied to programs with encouraging results. Just like DL is enabled by the enormous amount of textual and image data available on the Internet, DP is enabled by the vast amount of code available in open source repositories such as GitHub, as well as the ability to reuse libraries via modern package managers such as npm and pip. Two trail-blazing transformer-based DP systems are OpenAI's Codex8 and Deepmind's AlphaCode.18 The former is used in the Github Copilot project14 and integrates with development environments to automatically suggest code to developers. The latter generates code to solve problems presented at coding competitions. Both achieve amazing results. Multiple efforts are under way to establish code repositories for benchmarking DP, such as CodeXGLUE19 and CodeNET.20

The advent of DP systems has led to a few sensational headlines declaring that in the not-too-distant future coding will be done by computers, not humans.1 As DL technologies get even better and more code is deposited into public repositories, programmers will be replaced by specification writers outlining what code they want in natural language and presto, the code appears. This Viewpoint argues that while DP will influence software engineering and programming, its effects will be more incremental than the current hype suggests. To get away from the hype, I provide a careful analysis of the problem. I also argue that for DP to broaden its influence, it needs to take a more multidisciplinary approach, incorporating techniques from software engineering, program synthesis, and symbolic reasoning, to name just a few. Note I do not argue with the premise that DL will be used to solve many problems that are solved today by traditional programming methods16 and that software engineering will evolve to make such systems robust.17 In this Viewpoint, I am addressing the orthogonal question of using DL to synthesize programs themselves. ... ' 

Microsoft Research Looks Back at Progress in AI

 Microsoft talks about some of its efforts. 

Research @ Microsoft 2022: A look back at a year of accelerating progress in AI

Published December 19, 2022

2022 has seen remarkable progress in foundational technologies that have helped to advance human knowledge and create new possibilities to address some of society’s most challenging problems. Significant advances in AI have also enabled Microsoft to bring new capabilities to customers through our products and services, including GitHub Copilot, an AI pair programmer capable of turning natural language prompts into code, and a preview of Microsoft Designer, a graphic design app that supports the creation of social media posts, invitations, posters, and one-of-a-kind images.

These offerings provide an early glimpse of how new AI capabilities, such as large language models, can enable people to interact with machines in increasingly powerful ways. They build on a significant, long-term commitment to fundamental research in computing and across the sciences, and the research community at Microsoft plays an integral role in advancing the state of the art in AI, while working closely with engineering teams and other partners to transform that progress into tangible benefits.

In 2022, Microsoft Research established AI4Science, a global organization applying the latest advances in AI and machine learning toward fundamentally transforming science; added to and expanded the capabilities of the company’s family of foundation models; worked to make these models and technologies more adaptable, collaborative, and efficient; further developed approaches to ensure that AI is used responsibly and in alignment with human needs; and pursued different approaches to AI, such as causal machine learning and reinforcement learning.

We shared our advances across AI and many other disciplines during our second annual Microsoft Research Summit, where members of our research community gathered virtually with their counterparts across industry and academia to discuss how emerging technologies are being explored and deployed to bring the greatest possible benefits to humanity.  

Plenary sessions at the event focused on the transformational impact of deep learning on the way we practice science, research that empowers medical practitioners and reduces inequities in healthcare, and emerging foundations for planet-scale computing. Further tracks and sessions over three days provided deeper dives into the future of the cloud; efficient large-scale AI; amplifying human productivity and creativity; delivering precision healthcare; building user trust through privacy, identity, and responsible AI; and enabling a resilient and sustainable world.

BLOG

Microsoft Climate Research Initiative (MCRI) 

In June, the Microsoft Climate Research Initiative (MCRI) announced its first phase of collaborations among multidisciplinary researchers working together to accelerate cutting-edge research and transformative innovation in climate science and technology.

PUBLICATION

New Future of Work Report 2022 

In May, researchers across Microsoft published the New Future of Work Report 2022, which summarizes important recent research developments related to hybrid work. It highlights themes that have emerged in the findings of the past year and resurfaces older research that has become newly relevant.

In this blog post, we look back at some of the key achievements and notable work in AI and highlight other advances across our diverse, multidisciplinary, and global organization.

Advancing AI foundations and accelerating progress

Over the past year, the research community at Microsoft made significant contributions to the rapidly evolving landscape of powerful large-scale AI models. Microsoft Research and the Microsoft Turing team unveiled a new Turing Universal Language Representation model capable of performing both English and multilingual understanding tasks. In computer vision, advancements for the Project Florence-VL (Florence-Vision and Language) team spanned still imagery and video: its GIT model was the first to surpass human performance on the image captioning benchmark TextCaps; LAVENDER showed strong performance in video question answering, text-to-video retrieval, and video captioning; and GLIP and GLIPv2 combined localization and vision-language understanding. The group also introduced NUWA-Infinity, a model capable of converting text, images, and video into high-resolution images or long-duration video. Meanwhile, the Visual Computing Group scaled up its Transformer-based general-purpose computer vision architecture, Swin Transformer, achieving applicability across more vision tasks than ever before.   .... ' 

Network Pruning Can Skew Deep Learning Models

Making Better Models

Network Pruning Can Skew Deep Learning Models

North Carolina State University News

Matt Shipman, November 2, 2022

Computer science researchers at North Carolina State (NC State), Syracuse, and Carnegie Mellon universities have shown that neural network pruning can undermine the performance of deep learning models at identifying certain groups. The researchers cited disparities in gradient norms across groups, and in Hessian norms linked to inaccuracies of a group's data, as factors impacting performance. This implies network pruning can compound existing accuracy deficiencies. NC State's Jung-Eun Kim said the team has demonstrated a remedial mathematical method "to equalize the groups that the deep learning model is using to categorize data samples." Tests of the mitigation technique found basically restored a deep learning model to pre-pruning levels of accuracy.  ... '

Full Article

Fast and Scalable Search of Images

Image Search has many useful applications, notably health care

Harvard Researchers Propose a Self-Supervised Deep Learning Algorithm for Fast and Scalable Search of Whole-Slide Images

By Mahmoud Ghorbel- October 14, 2022  in MarktechPost

The necessity for accurate and economical gigapixel image analysis has risen as whole-slide imaging has become more widely used. Deep learning is at the forefront of computer vision, showing considerable advancements in visual comprehension over earlier approaches. However, whole-slide images (WSI) include billions of pixels and are plagued by many sorts of artifacts as well as significant morphological variation. All of these work against the traditional usage of deep learning. These challenges must be overcome for the clinical translation of deep learning solutions to become a reality.

Most computational pathology approaches use supervised deep learning with slide- or case-level labels to address classification or ranking issues. For many applications, an image search engine that uses the detailed, spatially resolved information in pathology images is far more potent. However, scalability poses a significant obstacle to the widespread and effective use of histology whole-slide image search and retrieval systems. Compared to other image databases, this presents a challenging issue for the WSI retrieval system since it must effectively search an increasing number of slides, each of which may contain billions of pixels and be several gigabytes. Since WSIs are too large to process computationally, most methods either divide them into smaller image patches or concentrate on patch or region of interest (ROI) retrieval that is specialized for specific purposes. Recently, a new article published in the journal nature biomedical engineering suggests a search pipeline called self-supervised image search for histology (SISH) to overcome the problems listed above.

Regardless of the repository size, SISH searches for and retrieves WSIs quickly. It uses a tree data structure for quick searching and only needs slide-level annotations for training. It transforms WSIs into useful discrete latent representations and then ranks them according to an algorithm based on uncertainty. To decrease storage and labeling costs, SISH specifically draws on a collection of preprocessed mosaics from WSIs without pixel-wise or ROI-level labels by relying on indices discovered by self-supervised learning and pretrained embeddings. The proposed approach leverages discrete latent codes from a Vector Quantized-Variational AutoEncoder (VQ-VAE) in addition to the guided search  ... ' 

Deep Learning Could Bring 3D Concert Experience Home

 New Realism in concerts

DEEP LEARNING COULD BRING THE CONCERT EXPERIENCE HOME  in IEEE 

The century-old quest for truly realistic sound production is finally paying off

QI “PETER” LIYIN DING  JOREL OLAN

OW THAT RECORDED sound has become ubiquitous, we hardly think about it. From our smartphones, smart speakers, TVs, radios, disc players, and car sound systems, it’s an enduring and enjoyable presence in our lives. In 2017, a survey by the polling firm Nielsen suggested that some 90 percent of the U.S. population listens to music regularly and that, on average, they do so 32 hours per week.

Behind this free-flowing pleasure are enormous industries applying technology to the long-standing goal of reproducing sound with the greatest possible realism. From Edison’s phonograph and the horn speakers of the 1880s, successive generations of engineers in pursuit of this ideal invented and exploited countless technologies: triode vacuum tubes, dynamic loudspeakers, magnetic phonograph cartridges, solid-state amplifier circuits in scores of different topologies, electrostatic speakers, optical discs, stereo, and surround sound. And over the past five decades, digital technologies, like audio compression and streaming, have transformed the music industry.

And yet even now, after 150 years of development, the sound we hear from even a high-end audio system falls far short of what we hear when we are physically present at a live music performance. At such an event, we are in a natural sound field and can readily perceive that the sounds of different instruments come from different locations, even when the sound field is criss-crossed with mixed sound from multiple instruments. There’s a reason why people pay considerable sums to hear live music: It is more enjoyable, exciting, and can generate a bigger emotional impact.

To hear the author's 3D Soundstage audio for yourself, grab your headphones and head over to 3dsoundstage.com/ieee.

Today, researchers, companies, and entrepreneurs, including ourselves, are closing in at last on recorded audio that truly re-creates a natural sound field. The group includes big companies, such as Apple and Sony, as well as smaller firms, such as Creative. Netflix recently disclosed a partnership with Sennheiser under which the network has begun using a new system, Ambeo 2-Channel Spatial Audio, to heighten the sonic realism of such TV shows as “Stranger Things” and “The Witcher.”

There are now at least half a dozen different approaches to producing highly realistic audio. We use the term “soundstage” to distinguish our work from other audio formats, such as the ones referred to as spatial audio or immersive audio. These can represent sound with more spatial effect than ordinary stereo, but they do not typically include the detailed sound-source location cues that are needed to reproduce a truly convincing sound field.

We believe that soundstage is the future of music recording and reproduction. But before such a sweeping revolution can occur, it will be necessary to overcome an enormous obstacle: that of conveniently and inexpensively converting the countless hours of existing recordings, regardless of whether they’re mono, stereo, or multichannel surround sound (5.1, 7.1, and so on). No one knows exactly how many songs have been recorded, but according to the entertainment-metadata concern Gracenote, more than 200 million recorded songs are available now on planet Earth. Given that the average duration of a song is about 3 minutes, this is the equivalent of about 1,100 years of music. ... 

Watershed Models with Deep Learning

Unusual application of interest.

Deep learning improves stream discharge-based estimates of subsurface permeability, allowing scientists to create more accurate watershed models. Deep Learning Uses Stream Discharge to Estimate Watershed Subsurface Permeability

U.S. Department of Energy

September 26, 2022

Deep learning can calculate a watershed's subsurface permeability from stream discharge data more accurately than conventional methods. Scientists from the U.S. Department of Energy's Pacific Northwest National Laboratory, Oak Ridge National Laboratory, and Los Alamos National Laboratory taught deep neural networks (DNNs) to estimate subsurface permeability from stream discharge hydrographs. The researchers trained the DNNs to map relationships between soil and geologic layer permeabilities and simulated stream discharge acquired via an integrated surface-subsurface hydrologic watershed model; this returned more accurate permeability than inverse modeling. The networks then estimated the permeability of an actual watershed using observed stream discharge from the study site, accurately predicting stream flows. The enhanced parameter estimation promises to reduce uncertainty in predictive

Algorithm Learns to Correct 3D Printing Errors

 Seems a very useful capability. 

Algorithm Learns to Correct 3D Printing Errors

University of Cambridge (U.K.)

August 16, 2022

An algorithm developed by researchers at the U.K.'s University of Cambridge can be added to new or existing three-dimensional (3D) printers to enable real-time error detection and correction across different materials and printing systems. The Deep Learning computer vision model was trained with about 950,000 images from the production of 192 printed objects, with labels detailing the printer's speed, nozzle temperature, and flow rate, among other settings. Said University of Cambridge's Sebastian Pattinson, "Once trained, the algorithm can figure out just by looking at an image which setting is correct and which is wrong; is a particular setting too high or too low, for example, and then apply the appropriate correction. And the cool thing is that printers that use this approach could be continuously gathering data, so the algorithm could be continually improving as well." ... 

Thinking Causal AI

Good thoughts on the topic: 

Use Causal AI to Go Beyond Correlation-Based Prediction

Gartner, By Leinar Ramos | August 10, 2022   Intro below

This is a short introduction to a research note we published recently on Causal AI, which is accessible here: Innovation Insight: Causal AI.

Correlation is not causation

“Correlation is not causation” is often mentioned, but rarely given the importance it deserves on AI. Correlations are how we see variables moving together in the data, but these relationships are not always causal. 

We can only say that A causes B when an intervention that changes A would also change B as a result (whilst keeping everything else constant). For example, forcing a rooster to crow won’t make the sun rise, even if the two events are correlated.

In other words, correlations are the data we see, whereas causal relationships are the underlying cause-and-effect relationships that generate this data (see image below). Crucially, the data we typically work with exists in a complex web of correlations that obscure the causal relationships we care about.

An image illustrating the distinction between correlations, which are the relationships we directly observe in the data, and causation, which is the underlying set of cause-and-effect relationships that generate the data 

Despite their notable success, statistical models, including those in advanced deep learning (DL) systems, use surface-level correlations to make predictions. The current DL paradigm doesn’t drive models to uncover underlying cause-and-effect relationships but simply to maximize predictive accuracy.

Now, it is worth asking: What is the problem of using correlations for prediction? After all, in order to predict, we just need enough predictive power in the data, regardless of whether it comes from causal relationships or statistical correlations. For instance, hearing a rooster crow is useful to predict sunrises.

The core problem lies with the brittleness of the predictions. For correlation-based predictions to remain valid, the process that generated the data needs to remain the same (e.g., the roosters need to keep crowing before sunrise).

There are two fundamental challenges with this correlation-based approach:

Problem #1: We want to intervene in the world

Prediction is rarely the end goal. We often want to intervene in the world to achieve a specific outcome. Anytime we ask a question of the form “How much can we change Y by doing X?”, we are asking a causal question about a potential intervention. An example would be: “What would happen to customer churn if we increased a loyalty incentive?”

And the problem with correlation-based predictive models, like Deep Learning, is that our actions are likely to change the data-generation process and therefore the statistical correlations we see in the data, rendering correlation-based predictions useless to estimate the effect of interventions. 

For instance, when we use a churn model (prediction) to decide whether or not to give a customer a loyalty incentive (intervention), the incentive affects the data that generated the prediction (we hope the incentive makes the customer stay). In this case, causality really matters, and we can’t simply use correlations to answer questions on what would happen if we took an action (we need to run controlled experiments or use causal techniques to estimate the effects)  .... ' 

Deep Learning from Molten Salts

Interesting results from local university. 

Deep Learning Method Worth Its Salt

UC (University of Cincinnati) News

Michael Miller, July 22, 2022

A multi-institutional team of researchers led by the University of Cincinnati's Yu Shi has developed a novel technique for modeling the thermodynamic properties of molten salts via deep learning artificial intelligence. Shi said the researchers trained a neural network on data produced by quantum simulations, which they used to estimate the free energy of molten sodium chloride. The research, according to Shi, offers a reliable way of studying the conversion of dissolved gas to vapor in molten salts, helping to understand how impurities and solutes affect corrosion. He added that the method also could help scientists analyze the emission of potentially toxic gas into the atmosphere, which will be useful for fourth-generation molten salt nuclear reactors.  ...

Deploying Deep Learning

 Most interesting  article...

Lessons From Deploying Deep Learning to Production

By The Gradient  May 18, 2022

Peter Gao is the co-founder and CEO of Aquarium, and he previously worked on machine learning for self-driving cars, education, and social media.

In an article, Peter Gao discusses his experiences deploying deep learning into production environments. "There are a lot more people using deep learning in production applications nowadays compared to when I was doing research at Berkeley, but many problems that they face are the same ones I grappled with in 2016 at Cruise," Gao writes.

"I used to think that machine learning was about the models. Actually, machine learning in production is about pipelines. One of the best predictors of success is the ability to effectively iterate on your model pipeline. That doesn't just mean iterating quickly, but also iterating intelligently...otherwise you end up with a pipeline that produces bad models very quickly."

"The vast majority of problems with model performance can be solved with data, but there are certain issues that can only be solved with changes to the model code." -Peter Gao

From The Gradient

View Full Article     

Designing Societally Beneficial Reinforcement Learning Systems

Integrating game playing.

Designing Societally Beneficial Reinforcement Learning Systems  in Bir Berkeluy

Nathan Lambert, Aaron Snoswell, Sarah Dean, Thomas Krendl Gilbert, Tom Zick,     Apr 29, 2022

Deep reinforcement learning (DRL) is transitioning from a research field focused on game playing to a technology with real-world applications. Notable examples include DeepMind’s work on controlling a nuclear reactor or on improving Youtube video compression, or Tesla attempting to use a method inspired by MuZero for autonomous vehicle behavior planning. But the exciting potential for real world applications of RL should also come with a healthy dose of caution - for example RL policies are well known to be vulnerable to exploitation, and methods for safe and robust policy development are an active area of research.

At the same time as the emergence of powerful RL systems in the real world, the public and researchers are expressing an increased appetite for fair, aligned, and safe machine learning systems. The focus of these research efforts to date has been to account for shortcomings of datasets or supervised learning practices that can harm individuals. However the unique ability of RL systems to leverage temporal feedback in learning complicates the types of risks and safety concerns that can arise.

This post expands on our recent whitepaper and research paper, where we aim to illustrate the different modalities harms can take when augmented with the temporal axis of RL. To combat these novel societal risks, we also propose a new kind of documentation for dynamic Machine Learning systems which aims to assess and monitor these risks both before and after deployment.

What’s Special About RL? A Taxonomy of Feedback

Reinforcement learning systems are often spotlighted for their ability to act in an environment, rather than passively make predictions. Other supervised machine learning systems, such as computer vision, consume data and return a prediction that can be used by some decision making rule. In contrast, the appeal of RL is in its ability to not only (a) directly model the impact of actions, but also to (b) improve policy performance automatically. These key properties of acting upon an environment, and learning within that environment can be understood as by considering the different types of feedback that come into play when an RL agent acts within an environment. We classify these feedback forms in a taxonomy of (1) Control, (2) Behavioral, and (3) Exogenous feedback. The first two notions of feedback, Control and Behavioral, are directly within the formal mathematical definition of an RL agent while Exogenous feedback is induced as the agent interacts with the broader world.

1. Control Feedback

First is control feedback - in the control systems engineering sense - where the action taken depends on the current measurements of the state of the system. RL agents choose actions based on an observed state according to a policy, which generates environmental feedback. For example, a thermostat turns on a furnace according to the current temperature measurement. Control feedback gives an agent the ability to react to unforeseen events (e.g. a sudden snap of cold weather) autonomously.  ....' 

For Better Understanding of Deep Learning

Very good extensive piece.

A Deeper Understanding of Deep Learning   By Don Monroe

Communications of the ACM, June 2022, Vol. 65 No. 6, Pages 19-20   10.1145/3530686

Deep learning should not work as well as it seems to: according to traditional statistics and machine learning, any analysis that has too many adjustable parameters will overfit noisy training data, and then fail when faced with novel test data. In clear violation of this principle, modern neural networks often use vastly more parameters than data points, but they nonetheless generalize to new data quite well.

The shaky theoretical basis for generalization has been noted for many years. One proposal was that neural networks implicitly perform some sort of regularization—a statistical tool that penalizes the use of extra parameters. Yet efforts to formally characterize such an "implicit bias" toward smoother solutions have failed, said Roi Livni, an advanced lecturer in the department of electrical engineering of Israel's Tel Aviv University. "It might be that it's like a needle in a haystack, and if we look further, in the end we will find it. But it also might be that the needle is not there."

A Profusion of Parameters

Recent research has clarified that learning systems operate in an entirely different regime when they are highly overparameterized, such that more parameters let them generalize better. Moreover, this property is shared not just by neural networks but by more comprehensible methods, which makes more systematic analysis possible.

"People were kind of aware that there were two regimes," said Mikhail Belkin, a professor in the Haliciolu Data Science Institute of the University of California, San Diego. However, "I think the clean separation definitely was not understood" prior to work he and colleagues published in 2019. "What you do in practice," such as forced regularization or early stopping of training, "mixes them up." ... ' 

Tracking Animals with Deep Learning

 Unusual Application

Time to Get Social: Tracking Animals with Deep Learning

EPFL (Switzerland), April 22, 2022

Researchers at Switzerland's École polytechnique fédérale de Lausanne (EPFL) enhanced their DeepLabCut software to enable high-performance tracking of animals in videos. The open source software uses deep learning to teach computers "pose estimation" without requiring physical or virtual markers on the animals. The update allows DeepLabCut to track the movements of social animals, like mice or fish, which can confuse the computer because they look too similar or obscure each other. With the help of researchers at the Massachusetts Institute of Technology and Harvard University, the EPFL team developed four datasets for benchmarking multi-animal pose estimation networks. The researchers also developed a multi-task neural network that can predict keypoints, limbs, and animal identity from single frames, and an assembly algorithm to help identify animals with varied body shapes.... ' 

AI Strips out Noise

Note AI methods in use for noice removal. 

Deep Learning, Supervised Learning. Other applications in noise reduction?

AI Strips Out City Noise to Improve Earthquake Monitoring Systems

New Scientist, Chris Stokel-Walker, April 13, 2022

Stanford University's Gregory Baroza and colleagues used a deep learning algorithm to eliminate city noise from earthquake monitors, in an attempt to fine-tune the ability to locate where tremors originate. The researchers trained the artificial intelligence on 80,000 samples of urban noise and 33,751 samples of earthquake signals to distinguish between the two. Running audio through the neural network enhanced the signal-to-noise ratio by an average of 15 decibels, triple the average of previous denoising methods. Rice University's Maarten de Hoop said one shortcoming of the approach was the network's training via supervised learning using human-labeled data sampled from one area; he said this makes the technique less likely to be effective when presented with noise from somewhere else. ... '

Deep Learning is a Bad Idea for Security

Deep Learning Is a Bad Idea for Security

By Fortune  February 28, 2022

Zulfikar Ramzan is the chief scientist and CEO of Aura Labs.

Ramzan holds a controversial belief regarding artificial intelligence (AI) and security: "I think deep learning is a completely terrible idea for security.  Ramzan's statement flies in the face of conventional wisdom that neural networks and their ability to discover patterns in massive amounts of data will revolutionize the cybersecurity industry.

While he concedes that exceptions may exist and that deep-learning experts may be doing innovative work applying AI to security problems, deep learning still has some major issues that need to be addressed before he feels comfortable enough for AI to be used more widely in the security industry.

From Fortune  

Details on Microsoft Singularity for AI Workflow

Worth a look, I will.  Note indication of scheduling AI workloads.

Microsoft goes public with details on its 'Singularity' AI infrastructure service

Microsoft's Azure and Research teams are working together on the 'Singularity' AI infrastructure service.

By Mary Jo Foley

Posted in All About Microsoft on February 21, 2022 | Topic: AI & Robotics

Microsoft's Azure and Research teams are working together to build a new AI infrastructure service, codenamed "Singularity." The Singularity team is working to build what Microsoft describes in some of its job postings for the group as "a new AI platform service ground-up from scratch that will become a major driver for AI, both inside Microsoft and outside."

A group of those working on the project have published a paper entitled "Singularity: Planet-Scale, Preemptible and Elastic Scheduling of AI Workloads,"  which provides technical details about the Singularity effort. The Singularity service is about providing data scientists and AI practitioners with a way to build, scale, experiment and iterate on their models on a Microsoft-provided distributed infrastructure service built specifically for AI.

Authors listed on the newly published paper include Azure Chief Technical Officer Mark Russinovich; Partner Architect Rimma Nehme, who worked on Azure Cosmos DB until moving to Azure to work on AI and deep learning in 2019; and Technical Fellow Dharma Shukla. From that paper:

"At the heart of Singularity is a novel, workload-aware scheduler that can transparently preempt and elastically scale deep learning workloads to drive high utilization without impacting their correctness or performance, across a global fleet of accelerators (e.g., GPUs, FPGAs)."  (Details at her ZDnet site linked to above) .... ' 

Deep Learning and Quantitative Finance

Of special interest, plan to attend:  

March 16 Talk, "Deep Learning for Sequences in Quantitative Finance" with David Kriegman

Register now (and more info)    for the next free ACM TechTalk, "Deep Learning for Sequences in Quantitative Finance," presented on Wednesday, March 16 at 12:00 PM ET/17:00 PM UTC by David Kriegman, Professor of Computer Science & Engineering at the University of California, San Diego. Andrew Rosenfeld, Senior Vice President at Two Sigma, will moderate the questions and answers session following the talk.

Leave your comments and questions with our speaker now and any time before the live event on ACM's Discourse Page. And check out the page after the webcast for extended discussion with your peers in the computing community, as well as further resources on deep learning, machine learning, and more.

(If you'd like to attend but can't make it to the virtual event, you still need to register to receive a recording of the TechTalk when it becomes available.)

Note: You can stream this and all ACM TechTalks on your mobile device, including smartphones and tablets.

The quantitative investment process can be viewed as one that takes in raw data at one end and executes trades that buy and sell financial instruments at the other end. The process naturally decomposes into steps of feature extraction, forecasting the returns of individual instruments, portfolio allocation to decide quantities to trade, and trading execution. Many of the steps in this process are readily expressed as machine learning problems that can be addressed using deep learning sequence methods. This talk will provide an overview of this pipeline and deep learning for sequences. No background knowledge in finance or deep learning is required. .... ' 

Deep Learning Can’t be Trusted

Deep Learning Can’t be Trusted,  Brain Modelling Pioneer Says Stephen Grossberg explains why his ART model is better  in Spectrum  IEEE

By  KATHY PRETZ During the past 20 years, deep learning has come to dominate artificial intelligence research and applications through a series of useful commercial applications. But underneath the dazzle are some deep-rooted problems that threaten the technology’s ascension.

The inability of a typical deep learning program to perform well on more than one task, for example, severely limits application of the technology to specific tasks in rigidly controlled environments. More seriously, it has been claimed that deep learning is untrustworthy because it is not explainable—and unsuitable for some applications because it can experience catastrophic forgetting. Said more plainly, if the algorithm does work, it may be impossible to fully understand why. And while the tool is slowly learning a new database, an arbitrary part of its learned memories can suddenly collapse. It might therefore be risky to use deep learning on any life-or-death application, such as a medical one.

Now, in a new book, IEEE Fellow Stephen Grossberg argues that an entirely different approach is needed. Conscious Mind, Resonant Brain: How Each Brain Makes a Mind describes an alternative model for both biological and artificial intelligence based on cognitive and neural research Grossberg has been conducting for decades. He calls his model Adaptive Resonance Theory (ART).

Grossberg—an endowed professor of cognitive and neural systems, and of mathematics and statistics, psychological and brain sciences, and biomedical engineering at Boston University—based ART on his theories about how the brain processes information.

“Our brains learn to recognize and predict objects and events in a changing world that is filled with unexpected events,” he says.

Based on that dynamic, ART uses supervised and unsupervised learning methods to solve such problems as pattern recognition and prediction. Algorithms using the theory have been included in large-scale applications such as classifying sonar and radar signals, detecting sleep apnea, recommending movies, and computer-vision-based driver-assistance software.

ART can be used with confidence because it is explainable and does not experience catastrophic forgetting, Grossberg says. He adds that ART solves what he has called the stability-plasticity dilemma: How a brain or other learning system can autonomously learn quickly (plasticity) without experiencing catastrophic forgetting (stability).

An illustration of a brain over a blue and red checkered pattern.  

Grossberg, who formulated ART in 1976, is a pioneer in modelling how brains become intelligent. He is the founder and director of Boston University’s Center for Adaptive Systems and the founding director of the Center of Excellence for Learning in Education, Science, and Technology. Both centers have sought to understand how the brain adapts and learns, and to develop technological applications based on their findings. .... ;