MIT to Reconcile Data Sharing with EU AI Regulations

Proposed regulations constraining AI proposed by the EU are over 100 pages long.  What is MIT's view?   Note some other  things also include  'subliminal behavior manipulation', I assume specifying exactly how this differs from clever advertising,  and 'Social Credit Scoring'. 

AI Weekly: MIT aims to reconcile data sharing with EU AI regulations  In VentureBeat By Kyle Wiggers, @Kyle_L_Wiggers

This week, the European Union (EU) unveiled regulations to govern the use of AI across the bloc’s 27 member states. The first-of-its-kind proposal spans more than 100 pages    and will take years to implement, but the ramifications are far-reaching. It imposes a ban — with some exceptions — on the use of biometric identification systems in public, including facial recognition. Other prohibited applications of AI include social credit scoring, the infliction of harm, and subliminal behavior manipulation.

The regulations are emblematic of an increased desire on the part of consumers for privacy-preserving, responsible implementations of AI and machine learning. A study by Capgemini found that customers and employees will reward organizations that practice ethical AI with greater loyalty, more business, and even a willingness to advocate for them — and in turn, punish those that don’t. And 87% of executives told Juniper in a recent survey that they believe organizations have a responsibility to adopt policies that minimize the negative impacts of AI  ..."

Many Simple Small Robots,

Robot work at Ga Tech, which we visited a number times.   Always liked the idea of small,  multiple and cooperative problem solving doing tasks.  Bur even without collaboration can have value.  Gathering data, especially can be a good example. 

Simple robots, smart algorithms   by Georgia Institute of Technology

Anyone with children knows that while controlling one child can be hard, controlling many at once can be nearly impossible. Getting swarms of robots to work collectively can be equally challenging, unless researchers carefully choreograph their interactions—like planes in formation—using increasingly sophisticated components and algorithms. But what can be reliably accomplished when the robots on hand are simple, inconsistent, and lack sophisticated programming for coordinated behavior?

A team of researchers led by Dana Randall, ADVANCE Professor of Computing and Daniel Goldman, Dunn Family Professor of Physics, both at Georgia Institute of Technology, sought to show that even the simplest of robots can still accomplish tasks well beyond the capabilities of one, or even a few, of them. The goal of accomplishing these tasks with what the team dubbed "dumb robots" (essentially mobile granular particles) exceeded their expectations, and the researchers report being able to remove all sensors, communication, memory and computation—and instead accomplishing a set of tasks through leveraging the robots' physical characteristics, a trait that the team terms "task embodiment."

The team's BOBbots, or "behaving, organizing, buzzing bots" that were named for granular physics pioneer Bob Behringer, are "about as dumb as they get," explains Randall. "Their cylindrical chassis have vibrating brushes underneath and loose magnets on their periphery, causing them to spend more time at locations with more neighbors." The experimental platform was supplemented by precise computer simulations led by Georgia Tech physics student Shengkai Li, as a way to study aspects of the system inconvenient to study in the lab.  ... ' 

How and Why to Share Scientific Code

Rarely done this consistently, but it is a useful approach to follow.   Add it to a review and follow up of results.   Plans for maintaining models.  

How and why to share scientific code

A simple guide to reproducible research without becoming a software engineer

By Nathan C. Frey

When you do an experiment, whether that’s in a lab or on a computer, you generate data that needs to be analyzed. If your analysis involves new methods, algorithms, or simulations, you probably wrote some code along the way. Scientific code is designed to be quick to write, easy for the writer to use, and never looked at again after the project is complete (maybe designed is a strong word).

For many scientists, packaging their code involves a lot of work and no reward. I want to share a few obvious benefits and some that are hopefully non-obvious. After that, I’ll give some tips for how to share your code as painlessly as possible without detouring into becoming a software engineer. If you want a simple example of what the finished product will look like, check out my repos for Python Topological Materials or Positive and Unlabeled Materials Machine Learning.

The benefits of sharing scientific code

Encourage reproducibility. As soon as a method has more than one step (click the big red button) or a data analysis pipeline is more complex than “we divided all the numbers by this number,” it becomes unlikely that other scientists will be able to really explore what you did. If you developed a set of instructions to process or generate your data, you wrote a program, whether you wrote it down in code or not. It’s much more natural to share that program than to only write out what you did in your paper.  ... " 

Overview of Graph Databases

Overview of Graph DBs and their uses.  From the CACM.  Below the intro, more at the link. 

Understanding NoSQL Database Types: Graph Databases     By Alex Williams  in CACM

While originating as a subset of NoSQL or "Not Only SQL," graph databases represent a sharp closing of SQL and NoSQL demarcations. Graph technologies are exploding in its market size as more companies and developers take up their hybrid flexibility offerings. Those offerings: Intuitivity plus scalability with a high connection and robust data pattern.

While I won't go into depth on the formation of the 'SQL vs NoSQL' debate, you could quite accurately say that SQL represents data stored in rows and tables, while high-growth NoSQL is data stores arranged via nested documents as columnar schemas or key-value pairs. One is relational, the other not so much.

Graph databases are formed from nodes, properties, and relationships—all in a very interlinked data structure. And yet it supports advanced, rich querying with scalability. In this model, relationships matter just as much as the data itself. In a sense, it combines the querying power of relational databases with the intuitive flexibility of columnar non-relational databases—supporting agile development while also letting you gain deep insights.

Why use graph databases: The benefits

The graph model is a general-purpose data technology. While many know it for its social media implementations—this 'emerging shape', as it's known amongst data scientists due to being a non-typical dataset, has become most popular with social media companies for performing social network analyses, and for creating social graphs via companies like Facebook and Twitter who are particularly focused on the Six Degrees of Separation concept—graph databases are actually found in a large variety of industries, ranging from finance to healthcare, to emergency-response networks.

The principal benefit of graph databases is using its ability to assign values to links or connections. If your data has connections, whether for offline machine learning systems or online mobile applications, implementing this emerging shape will likely be beneficial.

In short: Build high-fidelity, highly interconnected networks made of bite-sized, scalable patterns (ie. great for CI/CD dev) that can together service, query, and manage sophisticated problem domains.  ... ' 

Perspective Brain Hack

Good thought, but not always.

Perspective Taking: A Brain Hack That Can Help You Make Better Decisions

Mar 22, 2021 Opinion North America

Supports K@W's Innovation Content

In business and in life, many of our interactions benefit from perspective taking, or our ability to put ourselves in someone else’s shoes. In a unique corporate partnership with SEB, a leading Swedish corporate bank, The Wharton Neuroscience Initiative, explores the neural basis of perspective taking and its effects on collaboration and business outcomes. In this piece, Wharton marketing professor and neuroscientist Michael Platt, Vera Ludwig, Elizabeth Johnson and Per Hugander shed light on the neural basis of perspective taking and why it may lead to more innovation and better business outcomes. 

According to Martin Lorentzon, co-founder of Spotify and Tradedoubler, “The value of your company is equal to the sum of the problems you are able to solve.” But how can we build this problem-solving capability into our organizations? Neuroscience suggests that one key strategy may be taking the perspective of others. Not only does this crucial skill provide us with additional information about complex situations, it also activates brain regions linked with creativity and innovation.

Indeed, many frameworks and tools for solving tough and complex problems are centered around the ability to take the perspective of others. Innovation frameworks start with taking the customer’s perspective; collaboration and negotiation frameworks are centered around understanding others’ viewpoints; and dialogue models recommend postponing judgment in order to take different perspectives for solving numerous challenges from business issues to marital problems.

When considering how often perspective taking appears in the problem-solving literature, it is surprising that so few leaders invest time and effort in developing this skill. Even though organizations frequently use the aforementioned tools and frameworks, including the well-known approach of Design Thinking, the results may be suboptimal if individuals are not skilled in perspective taking itself.... " 

New Small Drone Rules in US

How full will the skies ultimately become,  with more drone-like and autonomous devices?  Expect it. Note remote ID requirements.

New Rules Allowing Small Drones to Fly Over People in U.S. Take Effect

Reuters, David Shepardson, April 21, 2021

Final rules from the U.S. Federal Aviation Administration that permit small drones to fly over people and at night took effect April 21. The rules also allow drones to fly over moving vehicles in some instances. To address security concerns, remote identification technology (Remote ID) will be required in most cases so drones can be identified from the ground. Drone manufacturers have been given 18 months to begin production of drones with Remote ID, and an additional year has been granted to operators to provide Remote ID. The rules do not require drones to be connected to the Internet to transmit location data, but they must use radio frequency (RF) broadcasting to transmit remote ID messages. U.S. Transportation Secretary Pete Buttigieg called the rules "an important first step in safely and securely managing the growing use of drones in our airspace."

More on Amazon's Palm Scanning System

Amazon Bringing Palm-Scanning Payment System to Whole Foods Stores

CNBC, Annie Palmer, April 21, 2021

Amazon's palm-scanning payment system will be rolled out to a Whole Foods store in Seattle's Capitol Hill neighborhood before expanding to seven other Whole Foods stores in the area in the coming months. About a dozen Amazon physical stores already offer the Amazon One payment system, which allows shoppers who have linked a credit card to their palm print to pay for items by holding their palm over a scanning device. Amazon says the palm-scanning system is "highly secure" and more private than facial recognition and other biometric systems. The company says thousands of people have signed up to use the system at the Amazon stores.... '

Apple Supplier Targeted with Ransomware

Even Apple suppliers are vulnerable.  Note in particular theft of corporate data and plans.

 Apple Targeted in $50-Million Ransomware Hack of Supplier Quanta

Bloomberg, Kartikay Mehrotra, April 21, 2021

Taiwan-based Apple contract manufacturer Quanta Computer suffered a ransomware attack apparently by Russian operator REvil, which claimed to have stolen the blueprints of Apple's latest products. A user on the cybercrime forum XSS posted Sunday that REvil was about to declare its "largest attack ever," according to an anonymous source. REvil named Quanta its latest victim on its "Happy Blog" site, claiming it had waited to publicize the breach until Apple's latest product launch because Quanta had refused to pay its ransom demands. By the time the launch ended, REvil had posted schematics for a new laptop, including the workings of what seems to be a Macbook designed as recently as March. ... '

EU to Constrain Certain AI Uses?

To be expected reaction from EU regulators.

Facial Recognition, Other 'Risky' AI Set for Constraints in EU

Bloomberg, Natalia Drozdiak, April 21, 2021

The European Commission has proposed new rules constraining the use of facial recognition and other artificial intelligence applications, and threatening fines for companies that fail to comply. The rules would apply to companies that, among other things, exploit vulnerable groups, deploy subliminal techniques, or score people’s social behavior. The use of real-time remote biometric identification systems by law enforcement also would be prohibited unless used specifically to prevent a terror attack, find missing children, or for other public security emergencies. Other high-risk applications, including for self-driving cars and in employment or asylum decisions, would have to undergo checks of their systems before deployment. The proposed rules need to be approved by the European Parliament and by individual member-states before they could become law.   ... " 

AI Centered Product Design

Had not heard of the concept as stated, worth a look:

How is AI-Centered Product Design Different?

By Amanda Linden  in TowardsDataScience

People who are interested in AI often ask me what an AI designer is, and I’ve attempted to answer that question in this article. I wanted to go a step further, by helping designers and product teams understand how designing AI-based product experiences is different from traditional product design. Here is what I’ve learned over the last two years of managing AI design & innovation teams, about how AI-first product thinking is evolving the traditional product design process. ....  " 

Related earlier article:  

Wharton: Planning for AI Risk Governance

Useful thoughts on the concept of governance of AI in the paper linked to below.

How Can Financial Institutions Prepare for AI Risks?

Apr 13, 2021 Analytics Wharton Research North America

Artificial intelligence (AI) technologies hold big promise for the financial services industry, but they also bring risks that must be addressed with the right governance approaches, according to a white paper  by a group of academics and executives from the financial services and technology industries, published by Wharton AI for Business. 

Wharton is the academic partner of the group, which calls itself Artificial Intelligence/Machine Learning Risk & Security, or AIRS. Based in New York City, the AIRS working group was formed in 2019, and includes about 40 academics and industry practitioners. ..." 

Book: End-to-end Encrypted Messaging

Continue to make my way through this excellent and detailed book.    Rolf Oppliger's site describes this book and others he has written.   Plus his company's ongoing research and work.   Order it below.

End-to-End Encrypted Messaging Hardcover – April 30, 2020   by Rolf Oppliger  (Author)

This exciting resource introduces the core technologies that are used for Internet messaging. The book explains how Signal protocol, the cryptographic protocol that currently dominates the field of end to end encryption (E2EE) messaging, is implemented and addresses privacy issues related to E2EE messengers. The Signal protocol and its application in WhatsApp is explored in depth, as well as the different E2EE messengers that have been made available in the last decade are also presented, including SnapChat. It addresses the notion of self-destructing messages (as originally introduced by SnapChat) and the use of metadata to perform traffic analysis.

Uses of AI for Small Business

 Fairly obvious, but the stats show you that small business understands the needs and possibilities.

The Growing Importance of AI for Small Businesses

AI has become enormously important for small businesses in recent years.  Posted By Gaurav Sharma

Artificial intelligence (AI) is no more confined to big businesses. As the technology matures and becomes more affordable, it has found a place in startups and small businesses as well.

A survey of 1,467 CEOs of small and mid-sized businesses (SMBs) found that out of all the current technologies, AI has had the maximum impact on their business.   ... " 

CPG Costs Rise in Pandemic

Seems like a need for analytical modeling of costs.  My former employer needs to step up to the  problem.

Price hikes on the horizon for P&G as material costs rise

In Reuters: 

Procter & Gamble Co (PG.N) said on Tuesday it would raise prices of certain products in the United States to offset rising costs that were already weighing on its fourth quarter, after reporting a better-than-expected quarterly result.

The Cincinnati-based company joins a growing list of consumer product makers hiking prices this year as they battle increasing costs for everything from transport to pulp and resin or edible oils and nuts.

P&G said since it gave its initial guidance for fiscal 2021 last year, costs had risen by $400 million, including after tax costs of $125 million for commodities that will largely hit the fourth quarter and $200 million in higher freight costs.... " 

Palm Swipe Scanning at Whole Foods

 We experimented with a number of ID/payment methods.   From Iris to thumbprint to Face.   But not palm swipe scanning, installed at Wholefoods now. Contact-less.

Amazon to let Whole Foods shoppers pay with a swipe of their palm   By Jeffrey Dastin

 (Reuters) - Amazon.com Inc said it is rolling out biometric technology at its Whole Foods stores around Seattle starting on Wednesday, letting shoppers pay for items with a scan of their palm.

The move shows how Amazon is bringing some of the technology already in use at its namesake brick-and-mortar Go and Books stores to the grocery chain it acquired in 2017.

The system, called Amazon One, lets customers associate a credit card with their palm print. It offers a contact-less alternative to cash and card payments, Amazon said.  .... " 

Need for Continuous and Dynamic Threat Modeling

Well done post from Cisco, with useful explanatory visuals. Strongly agree. Using and applying specific risk models.

The Need for Continuous and Dynamic Threat Modeling  From Cisco

By Sujata Ramamoorthy

This blog is co-authored by Mohammad Iqbal and is part four of a four-part series about DevSecOps.

The trend towards accelerated application development, and regular updates to an architecture through an agile methodology, reduces the efficacy and effectiveness of point-in-time threat modeling. This recognition led us to explore and strategize ways to continuously, and dynamically, threat model an application architecture during runtime.

Today, thanks to a robust DevOps environment, developers can deploy a complex architecture within a public cloud such as Amazon Web Services (AWS) or Google Cloud Platform without requiring support from a network or database administrator. A single developer can develop code, deploy an infrastructure through code into a public cloud, construct security groups through code, and deploy an application on the resulting environment all through a continuous integration/continuous delivery (CI/CD) pipeline. While this enables deployment velocity, it also eliminates multiple checks and balances. At Cisco, we recognized the risks introduced by such practices and decided to explore strategies to continuously evaluate how an architecture evolves in production runtime to guard against architecture drift.

Dynamic threat modeling must begin with a solid baseline threat model that is done in real-time. This can in turn be monitored for architecture drift. Our approach to obtain such a real-time view is to use dynamic techniques to allow security and ops teams to threat model live environments instead of diagraming on paper or whiteboards alone.

How Does Dynamic Threat Modeling Work?

Threat modeling is the practice of identifying data flows through systems and various constructs within an architecture that exhibit a security gap or vulnerabilities. A crucial element that enables the practice of threat modeling is generating the right kind of visual representation of a given architecture in an accurate manner. This approach can differ based on context and from one team to another.  At Cisco, we instead focused on elements and features that need to exist to allow a team to dynamically perform a threat modeling exercise. These elements include the ability:  .... " 

GE Working on Detecting COVID with a Smartphone

More detail, here from GE on their work on detecting COVID with a smartphone.

GE Scientists Developing Technology to Add COVID-19 Virus Detector to Your Mobile Device

Sensing Materials

Awarded National Institutes of Health (NIH) grant to develop tiny sensors smaller than your fingertip that can detect the presence of COVID-19 virus nano-particles on screens, tables and other surfaces

Multi-disciplinary team from GE Research will draw from years of development and commercial success with physical, environmental, gas and biosensors for industrial  monitoring

The Team’s work has been featured in journals Nature Electronics 2020 and Lab on a Chip 2021

NISKAYUNA, New York, April 8, 2021 – Building on a suite of successful sensing technologies that have resulted in field demonstrations and a commercial launch for industrial monitoring, GE Research has been awarded a 24-month NIH grant (U01AA029324) of the RADx-rad program to develop miniature sensors that can detect the presence of the COVID-19 virus nano-particles on an array of different surfaces.

“One of the first lines of defense against any virus is avoiding exposure, which is easier said than done when you can’t see it,” said Radislav Potyrailo, a principal scientist at GE Research and principal investigator on the NIH project. “Through our project with the NIH, we are developing a sensor small enough to embed in a mobile device that could detect the presence of the COVID-19 virus.”

Potyrailo added, “We all come into contact with different surfaces during any given day, from computer screens and conference tables to kiosks at the airport and of course, credit card machines at stores while running errands.  While everyone does a great job keeping these surfaces clean, we want to add an extra layer of safety by being able to detect the presence of the virus.”  ... '

MBRL Tuning for Partially Understood Environments

Below is very technical,  but I do like some of the background statements such as 'solving tasks in a partially understood environment ...'.    And the idea of optimizing agents to resolve elements of understanding.  (Which exemplifies the situations we are often in).    So I am not saying I understand this yet, but working through it now for broader application.  As part of my broader study of practical reinforcement learning.

The Importance of Hyperparameter Optimization for Model-based Reinforcement Learning

Nathan Lambert, Baohe Zhang, Raghu Rajan, André Biedenkapp    Apr 19, 2021  From BAIR  Berkeley

Model-based reinforcement learning (MBRL) is a variant of the iterative learning framework, reinforcement learning, that includes a structured component of the system that is solely optimized to model the environment dynamics. Learning a model is broadly motivated from biology, optimal control, and more – it is grounded in natural human intuition of planning before acting. This intuitive grounding, however, results in a more complicated learning process. In this post, we discuss how model-based reinforcement learning is more susceptible to parameter tuning and how AutoML can help in finding very well performing parameter settings and schedules. Below, left is the expected behavior of an agent maximizing velocity on a “Half Cheetah” robotic task, and to the right is what our paper with hyperparameter tuning finds.

MBRL

Model-based reinforcement learning (MBRL) is an iterative framework for solving tasks in a partially understood environment. There is an agent that repeatedly tries to solve a problem, accumulating state and action data. With that data, the agent creates a structured learning tool – a dynamics model – to reason about the world. With the dynamics model, the agent decides how to act by predicting into the future. With those actions, the agent collects more data, improves said model, and hopefully improves future actions.  ... " 

Database of World Management Practices

This came to my attention.   Could have been useful, depending on how precise the practices were detailed, in past modeling efforts.   This effort under way for 18 years! Now if we could have working process models of each organization type to work from.   Likely much harder.  But what if we could build up from basic models?  Forecast of future practices mentioned also of interest. Note offering measurement as well.  Has to be good data here. 

The World Management Survey at 18: Lessons and the Way Forward   by Daniela Scur, Raffaella Sadun, John Van Reenen, Renata Lemos, and Nicholas Bloom  in HBSWK

With a dataset of 13,000 firms and 4,000 schools and hospitals spanning more than 35 countries, the World Management Survey provides a systematic measure of management practices used in organizations. This paper gives an overview of lessons learned and a management policy toolkit for policymakers.

Author Abstract

Understanding how differences in management “best practices” affect organizational outcomes has been a focus of both theoretical and empirical work in the fields of management, sociology, economics and public policy. The World Management Survey (WMS) project was born almost two decades ago with the main goal of developing a new systematic measure of management practices being used in organizations. The WMS has contributed to a body of knowledge around how managerial structures, not just managerial talent, relates to organizational performance. Over 18 years of research, a set of consistent patterns have emerged and spurred new questions. We will present a brief overview of what we have learned in terms of measuring and understanding management practices and condense the implications of these findings for policy. We end with an outline of what we see as the path forward for both research and policy implications of this research program. ... '

  Paper: https://www.nber.org/papers/w28524

Facebook Research Allocates Ad Funding (without AI)

Struck me because we did something similar and also quite different for advertising fund allocation very early on.    We the results of integer optimization models with montecarlo simulation, constrained to advertising agreements.   I like the fact that AI is never mentioned!  But in much later spins on this,  what could be called AI was used to check for accuracy and regulatory compliance to contractual agreements.  Nice.

Auto-placement of ad campaigns using multi-armed bandits  in Facebook Research

By: Vashist Avadhanula, Riccardo Colini Baldeschi, Stefano Leonardi, Karthik Abinav Sankararaman, Okke Schrijvers

What the research is:

We look at the problem of allocating the budget of an advertiser across multiple surfaces optimally when both the demand and the value are unknown. Consider an advertiser who uses the Facebook platform to advertise a product. They have a daily budget that they would like to spend on our platform. Advertisers want to reach users where they spend time, so they spread their budget over multiple platforms, like Facebook, Instagram, and others. They want an algorithm to help bid on their behalf on the different platforms and are increasingly relying on automation products to help them achieve it.

In this research, we model the problem of placement optimization as a stochastic bandit problem. In this problem, the algorithm is participating in k different auctions, one for each platform, and needs to decide the correct bid for each of the auctions. The algorithm is given a total budget B (e.g., the daily budget) and a time horizon T over which this budget should be spent. At each time-step, the algorithm should decide the bid it will associate with each of the k platform, which will be input into the auctions for the next set of requests on each of the platforms. At the end of a round (i.e., a sequence of requests), the algorithm sees the total reward it obtained (e.g., number of clicks) and the total budget that was consumed in the process, on each of the different platforms. Based on just this history, the algorithm should decide the next set of bid multipliers it needs to place. The goal of the algorithm is to maximize the total advertiser value with the given budget across the k platforms.   ... "    

Full paper:  https://arxiv.org/abs/2103.10246