Categories
AI

DeepMind is developing one algorithm to rule them all

The Transform Technology Summits start October 13th with Low-Code/No Code: Enabling Enterprise Agility. Register now!


DeepMind wants to enable neural networks to emulate algorithms to get the best of both worlds, and it’s using Google Maps as a testbed.

Classical algorithms are what have enabled software to eat the world, but the data they work with does not always reflect the real world. Deep learning is what powers some of the most iconic AI applications today, but deep learning models need retraining to be applied in domains they were not originally designed for.

DeepMind is trying to combine deep learning and algorithms, creating the one algorithm to rule them all: a deep learning model that can learn how to emulate any algorithm, generating an algorithm-equivalent model that can work with real-world data.

DeepMind has made headlines for some iconic feats in AI. After developing AlphaGo, a program that became the world champion at the game of Go in a five-game match after beating a human professional Go player, and AlphaFold, a solution to a 50-year-old grand challenge in biology, DeepMind has set its sights on another grand challenge: bridging deep learning, an AI technique, with classical computer science.

The birth of neural algorithmic reasoning

Charles Blundell and Petar Veličković both hold senior research positions at DeepMind. They share a background in classical computer science and a passion for applied innovation. When Veličković met Blundell at DeepMind, a line of research known as Neural Algorithmic Reasoning (NAR), was born, after the homonymous position paper recently published by the duo.

The key thesis is that algorithms possess fundamentally different qualities when compared to deep learning methods — something Blundell and Veličković elaborated upon in their introduction of NAR. This suggests that if deep learning methods were better able to mimic algorithms, then generalization of the sort seen with algorithms would become possible with deep learning.

Like all well-grounded research, NAR has a pedigree that goes back to the roots of the fields it touches upon, and branches out to collaborations with other researchers. Unlike much pie-in-the-sky research, NAR has some early results and applications to show.

We recently sat down to discuss the first principles and foundations of NAR with Veličković and Blundell, to be joined as well by MILA researcher Andreea Deac, who expanded on specifics, applications, and future directions. Areas of interest include the processing of graph-shaped data and pathfinding.

Pathfinding: There’s an algorithm for that

Deac interned at DeepMind and became interested in graph representation learning through the lens of drug discovery. Graph representation learning is an area Veličković is a leading expert in, and he believes it’s a great tool for processing graph-shaped data.

“If you squint hard enough, any kind of data can be fit into a graph representation. Images can be seen as graphs of pixels connected by proximity. Text can be seen as a sequence of objects linked together. More generally, things that truly come from nature that aren’t engineered to fit within a frame or within a sequence like humans would do it, are actually quite naturally represented as graph structures,” said Veličković.

Another real-world problem that lends itself well to graphs — and a standard one for DeepMind, which, like Google, is part of Alphabet — is pathfinding. In 2020, Google Maps was the most downloaded map and navigation app in the U.S. and is used by millions of people every day. One of its killer features, pathfinding, is powered by none other than DeepMind.

The popular app now showcases an approach that could revolutionize AI and software as the world knows them. Google Maps, features a real-world road network that assists in predicting travel times. Veličković noted DeepMind has also worked on a Google Maps application that applies graph networks to predict travel times. This is now serving queries in Google Maps worldwide, and the details are laid out in a recent publication.

Veličković said that although one of the most iconic graph algorithms, Dijkstra’s algorithm, could in theory help compute shortest paths or even estimate travel times, that’s not applicable in reality. To do that, you would have to take all the complexity of the real world — roadblocks, weather changes, traffic flows, bottlenecks, and whatnot — and convert that into an ‘abstractified’ graph of nodes and edges with some weights which correspond to the travel time.

This exemplifies some of the issues with algorithms. As Veličković put it, algorithms are very happy to give you perfect answers, regardless of whether or not the inputs themselves are perfect. So it’s going to give you the shortest path. But who, we asked Veličković, guarantees that this graph is an accurate representation of the real world scenario? He said:

“The algorithm might give you a great solution, but there’s really no way of saying whether this human-devised heuristic is actually the best way of looking at it, and that’s the first point of attack. Historically, we found that whenever there’s a lot of human feature engineering and lots of raw data that we need to work with for that human feature engineering, this is where deep learning shines. The initial success stories of deep learning were exactly in replacing handcrafted heuristics for, say, image recognition with deep neural networks.”

This is the starting point for NAR: replacing the human that maps the real world raw data to a graph input with a neural network. The network will map the complexity of the raw data to some ‘abstractified’ graph inputs that can then be used to run the algorithm over.

Veličković noted there is a good line of research that works exactly in this setting. This research even figured out a way to propagate gradients through the algorithm so you can actually get good neural network optimization in the setting and apply the algorithm.

But, he went on to add, there are a few limitations with the setting, even if you’re able to propagate gradients through it. First of all, the algorithm may not be everything you need to compute the final solution. Some parts of the problem may require shortest path solutions, but maybe at some point you also need to run a flow algorithm and you are not sure how to combine results in the best way, because the problem is very raw and noisy:

“Forcing your representations to rely on the output of this one algorithm might be too limiting. Dijkstra requires one scalar value per every edge in the graph, which means that you’re compressing all that amazing complexity of the real world into just one number for every road segment.

That’s potentially problematic because, if you don’t have enough data to estimate that scalar correctly, you’re doomed. The algorithm is not going to give you correct solutions once again, because you just haven’t seen enough raw data to estimate that scalar correctly.”

Breaking the algorithmic bottleneck

This is what Veličković and Blundell refer to as the algorithmic bottleneck. It happens because we’re placing all of our bets on this one number for every edge, which is a very low dimensional representation of the problem. The way to break bottlenecks, the DeepMind duo suggests, is by using vectors, rather than numbers. In other words, staying highly dimensional.

Deep neural networks are deriving a lot of their success from staying highly dimensional and applying high-dimensional regularization techniques. This means that even if you predict badly some parts of the internal vector in the neural network, the other parts of that vector can still step in and compensate. The problem is that algorithms were designed to run on low dimensional representations, not high dimensional inputs.

The key idea in NAR is to replace the algorithm with a neural network, typically a graph neural network (GNN). The GNN takes a high dimensional input, and does one step of reasoning to produce another high dimensional input. Then once enough steps of that reasoning neural network are completed, final outputs are predicted based on this.

Of course, this high dimensional GNN neural executor needs to actually imitate the algorithm at hand. For this reason, NAR also includes an abstract pipeline where that neural network is pre-trained, typically using lots of synthetic data.

Blundell and Veličković’s previous work, such as neural execution of graph algorithms, pointer graph networks, and persistent message passing, dealt with that part: how to get reliable and robust neural networks that simulate algorithms in an abstract space.

What Blundell and Veličković had not done was to examine whether this can actually be used in a real-world problem. This is where Deac’s work comes in. The approach is based on a combination of neural execution of graph algorithms, pointer graph networks, and persistent message passing. All of this, Deac noted, was plugged into a reinforcement learning (RL) framework. In RL, problems are framed as states and actions, and the goal is to estimate how good each state is. This is what has driven RL applications in games, such as AlphaStar or AlphaGo, Deac noted:

“We define some value over the states. If you think of a chess game, this value is – how good is this move? And what we want is to make plans, make decisions so that this value is maximized. If we know the environment dynamics, [then how do we know] how we can move from one state to another, and what sorts of rewards do we get from a state for a specific action?”

The way to do this is via a so-called value iteration algorithm, which can provide estimates of how good each state is. The algorithm iterates on the value estimates and gradually improves them, until they converge to the ground truth values. This is the algorithm the NAR team is trying to imitate, using synthetic data and simple graphs to estimate meta-values.

The difference is that the team wanted to move from using single numerical values to using multiple-value high-dimensional vectors. Initially, the output might not be that good, but you don’t need that much data to get off the ground because the algorithm can deal with imperfect estimates when you don’t know the environment’s dynamics.

The key, as Deac explained, is iteration, which leads to convergence. In this case, a shortest path algorithm was of interest. The algorithm can be learned and then it will be plugged in. But the idea is that the RL framework should be usable for any algorithm, or combination of algorithms. That is an important step for machine learning, which sometimes struggles as it is re-applied from domain to domain

High performance in low data regimes

The approach Deac and the DeepMind duo worked on is called eXecuted Latent Value Iteration Network, or XLVIN. First, the value iteration algorithm is learned in the abstract space. Then a RL agent is plugged in. The team compared it against an almost identical architecture, with one crucial difference: rather than using their algorithmic component, that architecture predicts the values and runs value iteration on them directly.

Veličković said that this second agent actually managed in some cases to catch up, when fed with a lot more interactions with different environments. But in very low data regimes, the team’s RL architecture does better. This is important, as for environments such as Atari, a classical benchmark in AI also used in XLVIN, more data means more simulation budget, which is not always feasible.

XLVIN empirically validated a strong theoretical link between dynamic programing algorithms and the computations of a GNN. This means, Veličković said, that most polynomial time heuristics can be interpreted as dynamic programming, which in turn means that graph networks might be the right inductive bias for this kind of computation.

Previous theoretical work described a best case scenario, where there are settings of weights for a GNN that’ll make it nicely behave as that dynamic programming algorithm. But it doesn’t necessarily tell you what’s the best way to get there and how to make it work with the specific data you have or extrapolate, and these are issues that are important for algorithms, Veličković noted.

This led the duo to extend their work to models like pointer graph networks and persistent message passing, which moved one step further. They imitate the iterative computation of dynamic programming, but they also try to include some data structures which are crucial to how algorithms operate nowadays and incorporate some aspects of persistent reasoning.

So, rather than just being able to support a simple data structure on top of an existing set of nodes, is it possible to create additional memory additional nodes? Many algorithms rely on initializing additional memory aside from the memory required to store their inputs. So, DeepMind’s research developed models that enable more and more alignment with computation while still following that same GNN blueprint.

RL, Blundell noted, is basically a graph algorithm. It’s a dynamic programming update and it’s closely related to a shortest path algorithm — it’s like an online shortest path algorithm. It’s not surprising that if you’re trying to find the shortest possible path in a graph, and then you want to represent your problem as a graph, that maybe there’s a good relationship there.

Dynamic programming is a great way to think about solving any kind of problem, Blundell went on to add. You can’t always do it, but when you can, it works really, really well. And that’s potentially one of the deep connections between graph algorithms, reinforcement learning and graph networks.

One algorithm to rule them all

In their most recently published work, Reasoning-Modulated Representations, Blundell and Veličković show that they are able to use the algorithmic reasoning blueprints to support unsupervised learning and self-supervised learning. Often, Veličković said, unsupervised learning is all about, “Hey, let’s take a massive quantity of data and try to extract the most meaningful properties out of it.”

But that’s not always all the information you have. You might have some knowledge about how your data came to be. For example, if you’re working with estimating some representations from physics simulations, like a bunch of balls bouncing around or N-body systems, you don’t just see a bunch of snapshots of that system. You know that it has to obey certain laws of physics.

We think the neural algorithmic reasoning blueprint is an excellent way to take those laws of physics, package them up into a neural network that you can splice-in as part of your unsupervised architecture and, as a result, get better representations. And we’re starting to see some really nice results on a variety of environments that this blueprint is actually promising.”

As far as the future of this research goes, DeepMind’s duo wants to extend Deac’s work, and apply it as widely as possible in reinforcement learning, which is an area of really high interest for DeepMind and beyond. There’s algorithms ‘left, right and center inside the reinforcement learning pipeline,’ as Veličković put it.

Blundell on his part reiterated that there aren’t that many algorithms out there. So the question is, can we learn all of them? If you can have a single network that is able to execute any one of the algorithms that you know already, then if you get that network to plug those algorithms together, you start to form really quite complicated processing pipelines, or programs. And if it’s all done with gradients going through it, then you start to learn programs:

“If you really take this to its limit, then you start to really learn algorithms that learn. That becomes very interesting because one of the limitations in deep learning is the algorithms that we have to learn. There hasn’t been that much change in the best optimizers we use or how we update the weights in a neural network during training for quite a long time.

There’s been a little research over different architectures and so forth. But they haven’t always found the next breakthrough. The question is, is this a different way of looking at that, where we can start to find new learning algorithms?

Learning algorithms are just algorithms, and maybe what’s missing from them is this whole basis that we have for other algorithms that we’re using. So we need a slightly more universal algorithm executor to use as the basis for better methods for ML.”

Deac also noted she would like to pursue a network which tries multiple algorithms — all algorithms, if possible. She and some of her MILA colleagues have taken some steps in that direction. They are doing some transfer learning, chaining a couple of algorithms together and seeing if they can transfer between one algorithm, making it easier to learn a separate related algorithm, she said.

Or in other words, as Veličković framed what everyone seems to view as the holy grail of this research: “One algorithm to rule them all.”

VentureBeat

VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative technology and transact.

Our site delivers essential information on data technologies and strategies to guide you as you lead your organizations. We invite you to become a member of our community, to access:

  • up-to-date information on the subjects of interest to you
  • our newsletters
  • gated thought-leader content and discounted access to our prized events, such as Transform 2021: Learn More
  • networking features, and more

Become a member

Repost: Original Source and Author Link

Categories
AI

LinkedIn says it reduced bias in its connection suggestion algorithm

All the sessions from Transform 2021 are available on-demand now. Watch now.


In a blog post today, LinkedIn revealed that it recently completed internal audits aimed at improving People You May Know (PYMK), an AI-powered feature on the platform that suggests other members for users to connect with. LinkedIn claims the changes “level the playing field” for those who have fewer connections and spend less time building their online networks, making PYMK ostensibly useful for more people.

PYMK was the first AI-powered recommender feature at LinkedIn. Appearing on the My Network page, it provides connection suggestions based on commonalities between users and other LinkedIn members, as well as contacts users have imported from email and smartphone address books. Specifically, PYMK draws on shared connections and profile information and experiences, as well as things like employment at a company or in an industry and educational background.

PYMK worked well enough for most users, according to LinkedIn, but it gave some members a “very large” number of connection requests, creating a feedback loop that decreased the likelihood other, less-well-connected members would be ranked highly in PYMK suggestions. Frequently active members on LinkedIn tended to have greater representation in the data used to train the algorithms powering PYMK, leading it to become increasingly biased toward optimizing for frequent users at the expense of infrequent users.

“A common problem when optimizing an AI model for connections is that it often creates a strong ‘rich getting richer’ effect, where the most active members on the platform build a great network, but less active members lose out,” Albert Cui, senior product manager of AI and machine learning at LinkedIn, told VentureBeat via email. “It’s important for us to make PYMK as equitable as possible because we have seen that members’ networks, and their strength, can have a direct impact on professional opportunities. In order to positively impact members’ professional networks, we must acknowledge and remove any barriers to equity.”

Biased algorithms

This isn’t the first time LinkedIn has discovered bias in the recommendation algorithms powering its platform’s features. Years ago, the company found that the AI it used to match job candidates with opportunities was ranking candidates partly on the basis of how likely they were to apply for a position or respond to a recruiter. The system wound up referring more men than women for open roles simply because men are often more aggressive at seeking out new opportunities. To counter this, LinkedIn built an adversarial algorithm designed to ensure that the recommendation system includes a representative distribution of users across gender before referring the matches curated by the original system.

In 2016, a report in the Seattle Times suggested LinkedIn’s search algorithm might be giving biased results, too — along gender lines. According to the publication, searches for the 100 most common male names in the U.S. triggered no prompts asking if users meant predominantly female names, but similar searches of popular female first names paired with placeholder last names brought up LinkedIn’s suggestion to change “Andrea Jones” to “Andrew Jones,” “Danielle” to “Daniel,” “Michaela” to “Michael,” and “Alexa” to “Alex,” for example. LinkedIn denied at the time that its search algorithm was biased but later rolled out an update so any user who searches for a full name if they meant to look up a different name wouldn’t be prompted with suggestions.

Recent history has shown that social media recommendation algorithms are particularly prone to bias, intentional or not. A May 2020 Wall Street Journal article brought to light an internal Facebook study that found the majority of people who join extremist groups do so because of the company’s recommendation algorithms. In April 2019, Bloomberg reported that videos made by far-right creators were among YouTube’s most-watched content. And in a recent report by Media Matters for America, the media monitoring group presents evidence that TikTok’s recommendation algorithm is pushing users toward accounts with far-right views supposedly prohibited on the platform.

Correcting for imbalance

To address the problems with PYMK, LinkedIn researchers used a post-processing technique that reranked PYMK candidates to decrement the score of recipients who’d already had many unanswered invitations. These were mostly “ubiquitously popular” members or celebrities, who often received more invites than they could respond to due to their prominence or networks. LinkedIn thought that this would decrease the number of invitations sent to candidates suggested by PYMK and therefore overall activity. However, while connection requests sent by LinkedIn members indeed decreased 1%, sessions from the people receiving invitations increased by 1% because members with fewer invitations were now receiving more and invitations were less likely to be lost in influencers’ inboxes.

As a part of its ongoing Fairness Toolkit work, LinkedIn also developed and tested methods to rerank members according to theories of equality of opportunity and equalized odds. In PYMK, qualified IMs and FMs are now given equal representation in recommendations, resulting in more invites sent (a 5.44% increase) and connections made (a 4.8% increase) to infrequent members without majorly impacting frequent members.

“One thing that interested us about this work was that some of the results were counterintuitive to what we expected. We anticipated a decrease in some engagement metrics for PYMK as a result of these changes. However, we actually saw net engagement increases after making these adjustments,” Cui continued. “Interestingly, this was similar to what we saw a few years ago when we changed our Feed ranking system to also optimize for creators, and not just for viewers. In both of these instances, we found that prioritizing metrics other than those typically associated with ‘virality’ actually led to longer-term engagement wins and a better overall experience.”

All told, LinkedIn says it reduced the number of overloaded recipients — i.e., members who received too many invitations in the past week — on the platform by 50%. The company also introduced other product changes, such as a Follow button to ensure members could still hear from popular accounts. “We’ve been encouraged by the positive results of the changes we’ve made to the PYMK algorithms so far and are looking forward to continuing to use [our internal tools] to measure fairness to groups along the lines of other attributes beyond frequency of platform visits, such as age, race, and gender,” Cui said.

VentureBeat

VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative technology and transact.

Our site delivers essential information on data technologies and strategies to guide you as you lead your organizations. We invite you to become a member of our community, to access:

  • up-to-date information on the subjects of interest to you
  • our newsletters
  • gated thought-leader content and discounted access to our prized events, such as Transform 2021: Learn More
  • networking features, and more

Become a member

Repost: Original Source and Author Link

Categories
AI

AI Weekly: TikTok’s algorithm licensing signals China’s play for AI dominance

Join AI & data leaders at Transform 2021 on July 12th for the AI/ML Automation Technology Summit. Register today.


This week, TikTok parent company ByteDance began licensing parts of its AI technologies to third parties through a new division called BytePlus. The Financial Times reports that customers can pay fees to use recommendation algorithms, real-time filters and effects, automated translations, and computer vision tech akin to what’s found in TikTok, which has over 65.9 million users in the U.S. and is expected to see double-digit growth this year.

Readers will recall that in August 2020, when the Trump Administration said it would prohibit transactions with ByteDance, the Chinese government implemented restrictions to prevent companies from selling their algorithms without obtaining approval from officials. The jury’s out on whether ByteDance’s move runs afoul of those policies, but it comes amid Chinese regulators’ renewed clampdown on big tech. Companies like Alibaba and Tencent face fines and rules aimed at reining in specific business practices, including the listing of stock on overseas exchanges — even if the unit selling shares is incorporated outside China.

China’s crackdown on national tech giants has wiped a combined $823 billion off their value since February, with Tencent, Alibaba, and Kuaishou emerging as the biggest losers, according to Bloomberg. But as evidenced by ByteDance, China-based firms — particularly those focused on AI — are looking defiantly beyond the country’s borders for growth opportunities. For example, video surveillance company Hikvision, which is owned by the Chinese state, claims it now receives nearly 30% of its 50 billion yuan ($7 billion) in revenue from overseas. Facial recognition giant Megvii says 4.9% of its revenue came from outside China in the first six months of 2019. And in 2017, ByteDance said it expected to earn over half of its revenue from international users within five years.

While blacklisting by the U.S. Commerce Department cut off one potential avenue for Chinese AI companies, the sector continues to grow at a rapid clip in China. As of March 2019, the number of AI firms in the country reached 1,189, second only to the U.S. at 2,000. Investments in Chinese AI startups topped investments in American AI startups in 2018, the same year China filed 2.5 times more patents in AI technologies than the U.S. And China overtook the U.S. in 2019 for the number of most-cited AI research papers, according to the Allen Institute for AI.

The Biden Administration has clearly expressed an intention to reinvigorate the U.S. industry, in part through the National Artificial Intelligence Research Resource Task Force, which will be responsible for developing a roadmap to democratize access to AI research tools. In the coming weeks, President Joe Biden aims to establish the National AI Advisory Committee to provide recommendations and advice on a range of AI topics. And in February, the White House said it would bump non-defense-related AI investment to $2 billion annually by 2022, while President Biden has proposed increasing the amount of federal R&D spending to $300 billion over four years.

But with Chinese AI companies eyeing international markets, U.S. firms are likely to find it increasingly difficult to compete. China graduates as many as 3 times the number of computer scientists as the U.S., and the country’s AI Innovation Action Plan for Colleges and Universities called for the establishment of 50 new AI institutions in 2020. Moreover, China continues to invest heavily in the hardware needed to train AI models, particularly supercomputers, and now has over 200 of the world’s fastest supercomputers.

For competitors, the results have been stark. At a recent Stanford University-led international challenge for machine reading comprehension, Chinese teams won three of the top five spots, including first place. And researchers at the government-funded Beijing Academy of Artificial Intelligence (BAAI) in June announced the release of a multimodal AI model 10 times larger than San Francisco-based OpenAI’s 175-billion-parameter GPT- 3.

So how might the U.S. make up lost ground? Last July, the President’s Council of Advisors on Science and Technology (PCAST) released a report outlining what it believes must happen for the U.S. to advance “industries of the future,” including AI. PCAST recommended driving opportunities for AI education and training, creating incentive programs at universities, and increasing investments in AI educators, scientists, and technologists at all levels. Former Google CEO Eric Schmidt has also urged lawmakers to ramp up funding in the AI space while incentivizing public-private partnerships to develop AI applications across government agencies.

With a number of China’s largest AI companies pursuing what some would consider ethically questionable technologies, like facial recognition, it’s not clear whether the U.S. will ever corner every AI market — nor that it would wish to. But there’s clearly much to be done to promote the development of AI in the U.S., which the government seems — possibly — poised to do.

For AI coverage, send news tips to Kyle Wiggers — and be sure to subscribe to the AI Weekly newsletter and bookmark our AI channel, The Machine.

Thanks for reading,

Kyle Wiggers

AI Staff Writer

VentureBeat

VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative technology and transact.

Our site delivers essential information on data technologies and strategies to guide you as you lead your organizations. We invite you to become a member of our community, to access:

  • up-to-date information on the subjects of interest to you
  • our newsletters
  • gated thought-leader content and discounted access to our prized events, such as Transform 2021: Learn More
  • networking features, and more

Become a member

Repost: Original Source and Author Link

Categories
AI

Researchers develop algorithm to identify well-liked brands

Where does your enterprise stand on the AI adoption curve? Take our AI survey to find out.


Measuring sentiment can provide a snapshot of how customers feel about companies, products, or services. It’s important for organizations to be aware: 86% of people say that authenticity is a key factor when deciding what brands they like and support. In an Edelman survey, 8% of consumers [SHOULD THAT BE 80%?] said that they need to be able to trust a brand in order to buy products from them.

While sentiment analysis technology has been around for a while, researchers at the University of Maryland’s Robert H. Smith School of Business claim to have improved upon prior methods with a new system that leverages machine learning. They say that their algorithm, which sorts through social media posts to understand how people perceive brands, can comb through more data and better measure favorability.

Sentiment analysis isn’t a perfect science, but social media provides rich signals that can be used to help shape brand strategies. According to statistics, 46% of people have opted to use social media in the past to extend their complaints to a particular company.

“There is a vast amount of social media data available to help brands better understand their customers, but it has been underutilized in part because the methods used to monitor and analyze the data have been flawed,” Wendy W. Moe, University of Maryland associate dean of master’s programs, who created the algorithm with colleague Kunpeng Zhang, said in a statement. “Our research addresses some of the shortcomings and provides a tool for companies to more accurately gauge how consumers perceive their brands.”

Algorithmic analysis

Zhang’s and Moe’s method sifts through data from posts on a brand’s page, including how many users have expressed positive or negative sentiments, “liked” something, or shared something. It predicts how people will feel about that brand in the future, scaling to billions of pages of user-brand interaction data and millions of users.

The algorithm specifically looks at users’ interactions with brands to measure favorability — whether people view that brand in a positive or negative way. And it takes into account biases, inferring favorability and measuring social media users’ positivity based on their comments in the user-brand interaction data.

Zhang and Moe say that brands can apply the algorithm to a range of platforms, such as  Facebook, Twitter, and Instagram, as long as the platforms provide user-brand interaction data and allow users to comment, share, and like content. The algorithm importantly doesn’t use private information, like user demographics, relying instead on user-brand publicly available interaction data.

“A brand needs to monitor the health of their brand dynamically,” Zhang said in a statement. “Then they can change marketing strategy to impact their brand favorability or better respond to competitors. They can better see their current location in the market in terms of their brand favorability. That can guide a brand to change marketing [practices].”

Zhang’s and Moe’s research is detailed in the paper “Measuring Brand Favorability Using Large-Scale Social Media Data,” which will be published in the forthcoming issue of the journal Information Systems Research.

VentureBeat

VentureBeat’s mission is to be a digital town square for technical decision-makers to gain knowledge about transformative technology and transact.

Our site delivers essential information on data technologies and strategies to guide you as you lead your organizations. We invite you to become a member of our community, to access:

  • up-to-date information on the subjects of interest to you
  • our newsletters
  • gated thought-leader content and discounted access to our prized events, such as Transform 2021: Learn More
  • networking features, and more

Become a member

Repost: Original Source and Author Link

Categories
Tech News

Google’s algorithm misidentified an engineer as a serial killer

Google’s algorithmic failures can dreadful consequences, from directing racist search terms to the White House in Google Maps to labeling Black people as gorillas in Google Photos.

This week, the Silicon Valley giant added another algorithmic screw-up to the list: misidentifying a software engineer as a serial killer.

The victim of this latest botch was Hristo Georgiev, an engineer based in Switzerland. Georgiev discovered that a Google search of his name returned a photo of him linked to a Wikipedia entry on a notorious murderer.

“My first reaction was that somebody was trying to pull off some sort of an elaborate prank on me, but after opening the Wikipedia article itself, it turned out that there’s no photo of me there whatsoever,” said Georgiev in a blog post.

[Read: Why entrepreneurship in emerging markets matters]

Georgiev believes the error was caused by Google‘s knowledge graph, which generates infoboxes next to search results.

He suspects the algorithm matched his picture to the Wikipedia entry because the now-dead killer shared his name.

Georgiev is far from the first victim of the knowledge graph misfiring. The algorithm has previously generated infoboxes that falsely registered actor Paul Campbell as deceased and listed the California Republican Party’s ideology as “Nazism”.

In Georgiev’s case, the issue was swiftly resolved. After reporting the bug to Google, the company removed his image from the killer’s infobox. Georgiev gave credit to the HackerNews community for accelerating the response.

Other victims, however, may not be so lucky. If they never find the error — or struggle to resolve it — the misinformation could have troubling consequences.

I certainly wouldn’t want a potential employer, client, or partner to see my face next to an article about a serial killer.

Greetings Humanoids! Did you know we have a newsletter all about AI? You can subscribe to it right here.



Repost: Original Source and Author Link

Categories
Tech News

‘Bat-sense’ algorithm could be used to monitor people and property without cameras

A “bat-sense” algorithm that generates images from sounds could be used to catch burglars and monitor patients without using CCTV, the technique’s inventors say.

The machine-learning algorithm developed at Glasgow University uses reflected echoes to produce 3D pictures of the surrounding environment.

The researchers say smartphones and laptops running the algorithm could detect intruders and monitor care home patients.

[Read: 3 new technologies ecommerce brands can use to connect better with customers]

Study lead author Dr Alex Turpin said two things set the tech apart from other systems:

Firstly, it requires data from just a single input — the microphone or the antenna — to create three-dimensional images. Secondly, we believe that the algorithm we’ve developed could turn any device with either of those pieces of kit into an echolocation device.

The system analyses sounds emitted by speakers or radio waves pulsed from small antennas. The algorithm measures how long it takes for these signals to bounce around a room and return to the sensor.

It then analyzes the signal to calculate the shape, size, and layout of the room, as well as pick out the presence of objects or people. Finally, the data is converted into 3D images that are displayed as a video feed.

Credit: University of Glasgow
Categories
Tech News

Instagram apologizes after algorithm promotes harmful diet content to people with eating disorders

Instagram has apologized for a “mistake” that led its algorithm to promote harmful diet content to people with eating disorders.

The social network had been automatically recommending terms including “appetite suppressant,” which campaigners feared could lead vulnerable people to relapse.

Facebook, which owns Instagram, told the BBC that the suggestions were triggered by a new search functionality:

As part of this new feature, when you tap on the search bar, we’ll suggest topics you may want to search for. Those suggestions, as well as the search results themselves, are limited to general interests, and weight loss should not have been one of them.

The company said the harmful terms have now been removed and the issue with the search feature has been resolved.

[Read: The biggest tech trends of 2021, according to 3 founders]

Content that promotes disorders is banned from Instagram, while posts advertising weight-loss products are supposed to be hidden from users known to be under 18.

However, the error shows that policies alone can’t control the platform’s algorithms.

Greetings Humanoids! Did you know we have a newsletter all about AI? You can subscribe to it right here.



Repost: Original Source and Author Link

Categories
Tech News

People trust the algorithm more than each other

Our daily lives are run by algorithms. Whether we’re shopping online, deciding what to watch, booking a flight, or just trying to get across town, artificial intelligence is involved. It’s safe to say we rely on algorithms, but do we actually trust them?

Up front: Yes. We do. A trio of researchers from the University of Georgia recently conducted a study to determine whether humans are more likely to trust an answer they believe was generated by an algorithm or crowd-sourced from humans.

The results indicated that humans were more likely to trust algorithms when problems become to complex for them to trust their own answers.

Background: We all know that, to some degree or another, we’re beholden to the algorithm. We tend to trust that Spotify and Netflix know how to entertain us. So it’s not surprising that humans would choose answers based on the sole distinction that they’ve been labeled as being computer-generated.

But the interesting part isn’t that we trust machines, it’s that we trust them when we probably shouldn’t.

How it works: The researchers tapped 1,500 participants for the study. Participants were asked to look at a series of images and determine how many people were in each image. As the number of people in the image increased, humans gained less confidence in their answers and were offered the ability to align their responses with either crowd-sourced answers from a group of thousands of people, or answers they were told had been generated by an algorithm.

Per the study:

In three preregistered online experiments, we found that people rely more on algorithmic advice relative to social influence as tasks become more difficult. All three experiments focused on an intellective task with a correct answer and found that subjects relied more on algorithmic advice as difficulty increased. This effect persisted even after controlling for the quality of the advice, the numeracy and accuracy of the subjects, and whether subjects were exposed to only one source of advice, or both sources.

The problem here is that AI isn’t very well suited for a task such as counting the number of humans in an image. It may sound like a problem built for a computer – it’s math-based, after all – but the fact of the matter is that AI often struggles to identify objects in images especially when there aren’t clear lines of separation between objects of the same type.

Quick take: The research indicates the general public is probably a little confused about what AI can do. Algorithms are getting stronger and AI has become an important facet of our everyday lives, but it’s never a good sign when the average person seems to believe a given answer is better just because they think it was generated by an algorithm.



Repost: Original Source and Author Link

Categories
Tech News

Texas A&M drops “race” from student risk algorithm following Markup investigation

A major public university has paused its use of risk scores following a Markup investigation that found several universities using race as a factor in predicting student success. Our investigation also found that the software, Navigate, created by EAB and used by more than 500 schools across the country, was disproportionately labeling Black and other minority students “high risk”—a practice experts said ends up pushing Black kids out of math and science into “easier” majors.

Following our report, Texas A&M University announced it will stop including such risk scores on adviser dashboards and asked EAB to create new models that do not include race as a variable.

“We are committed to the success of all Texas A&M students,” Tim Scott, Texas A&M’s associate provost for academic affairs and student success, wrote in an email to The Markup. “Any decisions made about our students’ success will be done in a way that is fair and equitable to all students.”

The response from other schools has been mixed.

Maryclare Griffin, a statistics professor at the University of Massachusetts Amherst, another school featured in the story, said her institution appears to have taken down the option to view student risk scores for some Navigate users. One other professor at the school told The Markup that they were still able to view student risk scores.

UMass Amherst spokesperson Mary Dettloff would not confirm whether the school had made changes to its Navigate system and declined to answer other questions for this story.

The University of Houston, one of the four schools from which The Markup obtained data showing racial disparities in the risk scores, has not made any changes to its use of EAB’s algorithms, Shawn Lindsey, a spokesperson for the university, said.

The other schools mentioned in the original story—the University of Wisconsin–Milwaukee, South Dakota State University, Texas Tech University, and Kansas State University—did not respond to questions for this story.

The Markup obtained data from public universities showing that the algorithms embedded in educational research company EAB’s Navigate software assigned Black students high risk scores at double to quadruple the rate of their White peers. The risk scores purport to predict how likely a student is to drop out of school if that student remains within his or her selected major.

At nearly all the schools The Markup examined, the EAB algorithms used by the schools explicitly factored students’ race into their predictive models. And in several cases, the schools used race as a “high impact predictor” of success, meaning it was one of the variables with the most influence over students’ risk scores.

“EAB is deeply committed to equity and student success. Our partner schools hold differing views on the value of including demographic data in their risk models. That is why we are engaging our partner institutions to proactively review the use of demographic data,” EAB spokesperson John Michaels wrote in an email to The Markup. “Our goal has always been to give schools a clear understanding of the data that informs their customized models. We want to ensure that each institution can use the predictive analytics and broader platform as it is intended—to provide the best support for their students.”

EAB has marketed its advising software as a tool for cash-strapped universities to better direct their resources to the students who need help the most and, in the process, boost retention and avoid the additional cost of recruiting students to take the place of those who drop out.

But at the schools The Markup examined, we found that faculty and advisers who had access to EAB’s student risk scores were rarely, if ever, told how the scores were calculated or trained on how to interpret and use them. And in several cases, including at Texas A&M University, administrators were unaware that race was being used as a variable.

Instead, the software provided advisers a first impression of whether a student was at high-, moderate-, or low-risk of dropping out within his or her selected major, and then, through a function called Major Explorer, they were shown how that student’s risk might decrease if the student were to switch into a different, “less risky” field of study.

Experts said that design feature, coupled with the racial disparities in risk scores, was likely to perpetuate historic racism in higher education and result in students of color, particularly Black students, being encouraged to leave science, math, and engineering programs.

Iris Palmer, a senior adviser for higher education and workforce policy at New America, has studied the predictive analytics systems universities use to boost retention and has written a guide for schools to follow when considering whether to implement such systems.

“I don’t think taking race explicitly out of the algorithm solves the problem or makes the situation better necessarily,” she said. “Algorithms can predict race based on all sorts of other things that go into the algorithm,” such as combinations of data like zip code, high school name, and family income.

There is potential value in using predictive analytics to identify the students most in need of support, Palmer said, if schools actually train staff members on how the algorithms work and if the software explains, in a concise and understandable manner, which factors lead to each student being assigned a particular risk score. “And that’s a big if.”

Schools “need to do due diligence around disparate impact and why you’re seeing disparate impact on your campus,” she said. Had schools been doing that before signing multiyear contracts with EAB, “they wouldn’t have been caught unawares.”

This article by Todd Feathers was originally published on The Markup and was republished under the Creative Commons Attribution-NonCommercial-NoDerivatives license.

Repost: Original Source and Author Link

Categories
Tech News

Instagram’s algorithm pushes users towards COVID-19 misinformation, study finds

Instagram‘s algorithm is recommending COVID-19 and anti-vaccination misinformation to potentially millions of users, according to a new report.

The Center for Countering Digital Hate (CCDH) used test accounts to investigate the recommendations on Instagram’s Explore page and new Suggested Post feature.

They found that the tools encourage users to view misinformation and then push those who engage with the posts towards other extremist content:

If a user follows anti-vaxxers, they are fed QAnon conspiracism and antisemitic hate; if they engage with conspiracies, they are fed electoral and anti-vaxx misinformation.

The researchers generated the recommendations by creating 15 Instagram profiles and following different lists of accounts, from health authorities to anti-vaxxers.

[Read: How do you build a pet-friendly gadget? We asked experts and animal owners]

They logged into the Instagram accounts every day and recorded the recommendations they received.

As the Suggested Posts feature doesn’t trigger for new accounts that haven’t interacted with posts, the users scrolled through their feeds and the Explore section, and liked random posts to generate suggested content.

They then screenshot the recommendations they got between 14 September to 16 November 2020.

In total, Instagram recommended 104 posts containing misinformation. More than half of them were about COVID-19, while a fifth were about vaccines, and a tenth about the US election.

Credit: CCDH