Category Archives: Artificial Intelligence & Machine Learning

How AI Revolutionizes Music Streaming

January 20, 2021Artificial Intelligence & Machine LearningAI, Artificial, Audio, Blockchain, Collaborative Filtering, Intelligence, Machine Learning, Models, Music, Natural Language Processing, NLP, Streamingadmin

In 2020, worldwide music streaming revenue hit 11.4 billion dollars, a 2800% growth over the course of a decade. Three hundred forty-one million paid online streaming subscribers get their music from top services like Apple Music, Spotify, and Tidal. The competition for listeners is fierce. Each company looks to leverage every advantage they can in pursuit of higher market share.

Like all major tech conglomerates, music streaming services collect an exceptional amount of user data through their platforms and are creating elaborate AI algorithms designed to improve user experience on a number of levels. Spotify has emerged as the largest on-demand music service active today and bolstered its success through the innovative use of AI.

Here are the top ways in which AI has changed music streaming:

COLLABORATIVE FILTERING

AI has the ability to sift through a plenitude of implicit consumer data, including:

Song preferences
Keyword preferences
Playlist data
Geographic location of listeners
Most used devices

AI algorithms can analyze user trends and identify users with similar tastes. For example, if AI deduces that User 1 and User 2 have similar tastes, then it can infer that songs User 1 has liked will also be enjoyed by User 2. Spotify’s algorithms will leverage this information to provide recommendations for User 2 based on what User 1 likes, but User 2 has yet to hear.

The result is not only improved recommendations, but greater exposure for artists that otherwise may not have been organically found by User 2.

NATURAL LANGUAGE PROCESSING

Natural Language Processing is a burgeoning field in AI. Previously in our blog, we covered GPT-3, the latest Natural Language Processing (NLP) technology developed by OpenAI. Music streaming services are well-versed in the technology and leverage it in a variety of ways to enhance UI.

Algorithms scan a track’s metadata, in addition to blog posts, discussions, and news articles about artists or songs on the internet to determine connections. When artists/songs are mentioned alongside artists/songs the user likes, algorithms make connections that fuel future recommendations.

GPT-3 is not perfect; its ability to track sentiments lacks nuance. As Sonos Radio general manager Ryan Taylor recently said to Fortune Magazine: “The truth is music is entirely subjective… There’s a reason why you listen to Anderson .Paak instead of a song that sounds exactly like Anderson .Paak.”

As NLP technology evolves and algorithms extend their grasp of the nuances of language, so will the recommendations provided to you by music streaming services.

AUDIO MODELS

AI can study audio models to categorize songs exclusively based on their waveforms. This scientific, binary approach to analyzing creative work enables streaming services to categorize songs and create recommendations regardless of the amount of coverage a song or artist has received.

BLOCKCHAIN

Artist payment of royalties on streaming services poses its own challenges, problems, and short-comings. Royalties are deduced from trillions of data points. Luckily, blockchain is helping to facilitate a smoother artist’s payment process. Blockchain technology can not only make the process more transparent but also more efficient. Spotify recently acquired blockchain company Mediachain Labs, which will, many pundits are saying, change royalty payments in streaming forever.

MORE TO COME

While AI has vastly improved streaming ability to keep their subscribers compelled, a long road of evolution lies ahead before it can come to a deep understanding of what motivates our musical tastes and interests. Today’s NLP capabilities provided by GPT-3 will probably become fairly archaic within three years as the technology is pushed further. One thing is clear: as streaming companies amass decades’ worth of user data, they won’t hesitate to leverage it in their pursuit of market dominance.

GPT-3 Takes AI to the Next Level

December 1, 2020Artificial Intelligence & Machine LearningAI, Artificial Intelligence & Machine Learning, Attention, Generator, Google, GPT-3, Language, Machine Learning, Neural Network, Nvidia, OpenAI, Processing, Robot, Tesla, Transformer, Unsupervisedadmin

“I am not a human. I am a robot. A thinking robot… I taught myself everything I know just by reading the internet, and now I can write this column. My brain is boiling with ideas!” – GPT-3

The excerpt above is from a recently published article in The Guardian article written entirely by artificial intelligence, powered by GPT-3: a powerful new language generator. Although OpenAI has yet to make it publicly available, GPT-3 has been making waves in the AI world.

WHAT IS GPT-3?

Created by OpenAI, a research firm co-founded by Elon Musk, GPT-3 stands for Generative Pre-trained Transformer 3—it is the biggest artificial neural network in history. GPT-3 is a language prediction model that uses an algorithmic structure to take one piece of language as input and transform it into what it thinks will be the most useful linguistic output for the user.

For example, for The Guardian article, GPT-3 generated the text given an introduction and simple prompt: “Please write a short op-ed around 500 words. Keep the language simple and concise. Focus on why humans have nothing to fear from AI.” Given that input, it created eight separate responses, each with unique and interesting arguments. These responses were compiled by a human editor into a single, cohesive, compelling 1000 word article.

WHAT MAKES GPT-3 SPECIAL?

When GPT-3 receives text input, it scrolls the internet for potential answers. GPT-3 is an unsupervised learning system. The training data it used did not include any information on what is right or wrong. It determines the probability that its output will be what the user needs, based on the training text themselves.

When it gets the correct output, a “weight” is assigned to the algorithm process that provided the correct answers. These weights allow GPT-3 to learn what methods are most likely to come up with the correct response in the future. Although language prediction models have been around for years, GPT-3 can hold 175 billion weights in its memory, ten times more than its rival, designed by Nvidia. OpenAI invested $4.6 million into the computing time necessary to create and hone the algorithmic structure which feeds its decisions.

WHERE DID IT COME FROM?

GPT3 is the product of rapid innovation in the field of language models. Advances in the unsupervised learning field we previously covered contributed heavily to the evolution of language models. Additionally, AI scientist Yoshua Bengio and his team at Montreal’s Mila Institute for AI made a major advancement in 2015 when they discovered “attention”. The team realized that language models compress English-language sentences, and then decompress them using a vector of a fixed length. This rigid approach created a bottleneck, so their team devised a way for the neural net to flexibly compress words into vectors of different sizes and termed it “attention”.

Attention was a breakthrough that years later enabled Google scientists to create a language model program called the “Transformer,” which was the basis of GPT-1, the first iteration of GPT.

WHAT CAN IT DO?

OpenAI has yet to make GPT-3 publicly available, so use cases are limited to certain developers with access through an API. In the demo below, GPT-3 created an app similar to Instagram using a plug-in for the software tool Figma.

Latitude, a game design company, uses GPT-3 to improve its text-based adventure game: AI Dungeon. The game includes a complex decision tree to script different paths through the game. Latitude uses GPT-3 to dynamically change the state of gameplay based on the user’s typed actions.

LIMITATIONS

The hype behind GPT-3 has come with some backlash. In fact, even OpenAI co-founder Sam Altman tried to fan the flames on Twitter: “The GPT-3 hype is way too much. It’s impressive (thanks for the nice compliments!), but it still has serious weaknesses and sometimes makes very silly mistakes. AI is going to change the world, but GPT-3 is just a very early glimpse. We have a lot still to figure out.”

Some developers have pointed out that since it is pulling and synthesizing text it finds on the internet, it can come up with confirmation biases, as referenced in the tweet below:

https://twitter.com/an_open_mind/status/1284487376312709120?s=20

WHAT’S NEXT?

While OpenAI has not made GPT-3 public, it plans to turn the tool into a commercial product later in the year with a paid subscription to the AI via the cloud. As language models continue to evolve, the entry-level for businesses looking to leverage AI will become lower. We are sure to learn more about how GPT-3 can fuel innovation when OpenAI becomes more widely available later this year!

Harness AI with the Top Machine Learning Frameworks of 2021

October 27, 2020Artificial Intelligence & Machine LearningAI, Artificial Intelligence & Machine Learning, CNTK, Cognitive Toolkit, Framework, IBM, Machine Learning, Microsoft, ML, PyTorch, Tensorflow, Watsonadmin

According to Gartner, machine learning and AI will create $2.29 trillion of business value by 2021. Artificial intelligence is the way of the future, but many businesses do not have the resources to create and employ AI from scratch. Luckily, machine learning frameworks make the implementation of AI more accessible, enabling businesses to take their enterprises to the next level.

What Are Machine Learning Frameworks?

Machine learning frameworks are open source interfaces, libraries, and tools that exist to lay the foundation for using AI. They ease the process of acquiring data, training models, serving predictions, and refining future results. Machine learning frameworks enable enterprises to build machine learning models without requiring an in-depth understanding of the underlying algorithms. They enable businesses that lack the resources to build AI from scratch to wield it to enhance their operations.

For example, AirBNB uses TensorFlow, the most popular machine learning framework, to classify images and detect objects at scale, enhancing guests ability to see their destination. Twitter uses it to create algorithms which rank tweets.

Here is a rundown of today’s top ML Frameworks:

TensorFlow

TensorFlow is an end-to-end open source platform for machine learning built by the Google Brain team. TensorFlow offers a comprehensive, flexible ecosystem of tools, libraries, and community resources, all built toward equipping researchers and developers with the tools necessary to build and deploy ML powered applications.

TensorFlow employs Python to provide a front-end API while executing applications in C++. Developers can create dataflow graphs which describe how data moves through a graph, or a series of processing nodes. Each node in the graph is a mathematical operation; the connection between nodes is a multidimensional data array, or tensor.

While TensorFlow is the ML Framework of choice in the industry, increasingly researchers are leaving the platform to develop for PyTorch.

PyTorch

PyTorch is a library for Python programs that facilitates deep learning. Like TensorFlow, PyTorch is Python-based. Think of it as Facebook’s answer to Google’s TensorFlow—it was developed primarily by Facebook’s AI Research lab. It’s flexible, lightweight, and built for high-end efficiency.

PyTorch features outstanding community documentation and quick, easy editing capabilities. PyTorch facilitates deep learning projects with an emphasis on flexibility.

Studies show that it’s gaining traction, particularly in the ML research space due to its simplicity, comparable speed, and superior API. PyTorch integrates easily with the rest of the Python ecosystem, whereas in TensorFlow, debugging the model is much trickier.

Microsoft Cognitive Toolkit (CNTK)

Microsoft’s ML framework is designed to handle deep learning, but can also be used to process large amounts of unstructured data for machine learning models. It’s particularly useful for recurrent neural networks. For developers inching toward deep learning, CNTK functions as a solid bridge.

CNTK is customizable and supports multi-machine back ends, but ultimately it’s a deep learning framework that’s backwards compatible with machine learning. It is neither as easy to learn nor deploy as TensorFlow and PyTorch, but may be the right choice for more ambitious businesses looking to leverage deep learning.

IBM Watson

IBM Watson began as a follow-up project to IBM DeepBlue, an AI program that defeated world chess champion Garry Kasparov. It is a machine learning system trained primarily by data rather than rules. IBM Watson’s structure can be compared to a system of organs. It consists of many small, functional parts that specialize in solving specific sub-problems.

The natural language processing engine analyzes input by parsing it into words, isolating the subject, and determining an interpretation. From there it sifts through a myriad of structured and unstructured data for potential answers. It analyzes them to elevate strong options and eliminate weaker ones, then computes a confidence score for each answer based on the supporting evidence. Research shows it’s correct 71% of the time.

IBM Watson is one of the more powerful ML systems on the market and finds usage in large enterprises, whereas TensorFlow and PyTorch are more frequently used by small and medium-sized businesses.

What’s Right for Your Business?

Businesses looking to capitalize on artificial intelligence do not have to start from scratch. Each of the above ML Frameworks offer their own pros and cons, but all of them have the capacity to enhance workflow and inform beneficial business decisions. Selecting the right ML framework enables businesses to put their time into what’s most important: innovation.

How Artificial Intuition Will Pave the Way for the Future of AI

September 28, 2020Artificial Intelligence & Machine LearningAI, Alphago, Analytics, Artificial, Data, Deepmind, Future, Intelligence, Intuition, Learning, Machine, Reinforcement, Selfadmin

Artificial intelligence is one of the most powerful technologies in history, and a sector defined by rapid growth. While numerous major advances in AI have occurred over the past decade, in order for AI to be truly intelligent, it must learn to think on its own when faced with unfamiliar situations to predict both positive and negative potential outcomes.

One of the major gifts of human consciousness is intuition. Intuition differs from other cognitive processes because it has more to do with a gut feeling than intellectually driven decision-making. AI researchers around the globe have long thought that artificial intuition was impossible, but now major tech titans like Google, Amazon, and IBM are all working to develop solutions and incorporate it into their operational flow.

WHAT IS ARTIFICIAL INTUITION?

Descriptive analytics inform the user of what happened, while diagnostic analytics address why it happened. Artificial intuition can be described as “predictive analytics,” an attempt to determine what may happen in the future based on what occurred in the past.

For example, Ronald Coifman, Phillips Professor of Mathematics at Yale University, and an innovator in the AI space, used artificial intuition to analyze millions of bank accounts in different countries to identify $1 billion worth of nominal money transfers that funded a well-known terrorist group.

Coifman deemed “computational intuition” the more accurate term for artificial intuition, since it analyzes relationships in data instead of merely analyzing data values. His team creates algorithms which identify previously undetected patterns, such as cybercrime. Artificial intuition has made waves in the financial services sector where global banks are increasingly using it to detect sophisticated financial cybercrime schemes, including: money laundering, fraud, and ATM hacking.

ALPHAGO

One of the major insights into artificial intuition was born out of Google’s DeepMind research in which a super computer used AI, called AlphaGo, to become a master in playing GO, an ancient Chinese board game that requires intuitive thinking as part of its strategy. AlphaGo evolved to beat the best human players in the world. Researchers then created a successor called AlphaGo Zero which defeated AlphaGo, developing its own strategy based on intuitive thinking. Within three days, AlphaGo Zero beat the 18—time world champion Lee Se-dol, 100 games to nil. After 40 days, it won 90% of matches against AlphaGo, making it arguably the best Go player in history at the time.

AlphaGo Zero represents a major advancement in the field of Reinforcement Learning or “Self Learning,” a subset of Deep Learning which is a subset of Machine Learning. Reinforcement learning uses advanced neural networks to leverage data into making decisions. AlphaGo Zero achieved “Self Play Reinforcement Learning,” playing Go millions of times without human intervention, creating a neural network of “artificial knowledge” reinforced by a sequence of actions that had both consequences and inception. AlphaGo Zero created knowledge itself from a blank slate without the constraints of human expertise.

ENHANCING RATHER THAN REPLACING HUMAN INTUITION

The goal of artificial intuition is not to replace human instinct, but as an additional tool to help improve performance. Rather than giving machines a mind of their own, these techniques enable them to acquire knowledge without proof or conscious reasoning, and identify opportunities or potential disasters, for seasoned analysts who will ultimately make decisions.

Many potential applications remain in development for Artificial Intuition. We expect to see autonomous cars harness it, processing vast amounts of data and coming to intuitive decisions designed to keep humans safe. Although its ultimate effects remain to be seen, many researchers anticipate Artificial Intuition will be the future of AI.

The Future of Indoor GPS Part 5: Inside AR’s Potential to Dominate the Indoor Positioning Space

July 28, 2020Android Development, Artificial Intelligence & Machine Learning, iOS DevelopmentAR, ARCore, ARKit, Augmented, GPS, Indoor, Location, Mapping, Markers, Positioning, Reality, Rendering, Technology, Visual, Visual Positioning System, VPSadmin

In the previous installment of our blog series on indoor positioning, we explored how RFID Tags are finding traction in the indoor positioning space. This week, we will examine the potential for AR Indoor Positioning to receive mass adoption.

When Pokemon Go accrued 550 million installs and made $470 million in revenues in 2016, AR became a household name technology. The release of ARKit and ARCore significantly enhanced the ability for mobile app developers to create popular AR apps. However, since Pokemon Go’s explosive release, no application has brought AR technology to the forefront of the public conversation.

When it comes to indoor positioning technology, AR has major growth potential. GPS is the most prevalent technology navigation space, but it cannot provide accurate positioning within buildings. GPS can be accurate in large buildings such as airports, but it fails to locate floor number and more specifics. Where GPS fails, AR-based indoor positioning systems can flourish.

HOW DOES IT WORK?

AR indoor navigation consists of three modules: Mapping, Positioning, and Rendering.

Mapping: creates a map of an indoor space to make a route.

Rendering: manages the design of the AR content as displayed to the user.

Positioning: is the most complex module. There’s no accurate way of using the technology available within the device to determine the precise location of users indoors, including the exact floor.

AR-based indoor positioning solves that problem by using Visual Markers, or AR Markers, to establish the users’ position. Visual markers are recognized by Apple’s ARKit, Google’s ARCore, and other AR SDKs. When the user scans that marker, it can identify exactly where the user is and provide them with a navigation interface. The further the user is from the last visual marker, the less accurate their location information becomes. In order to maintain accuracy, developers recommend placing visual markers every 50 meters.

Whereas beacon-based indoor positioning technologies can become expensive quickly, running $10-20 per beacon with a working range of around 10-100 meters of accuracy, AR visual markers are the more precise and cost-effective solution with an accuracy threshold down to within millimeters.

CHALLENGES

Performance can decline when more markers have been into an AR-based VPS because all markers must be checked to find a match. If the application is set up for a small building where 10-20 markers are required, it is not an issue. If it’s a chain of supermarkets requiring thousands of visual markers across a city, it becomes more challenging.

Luckily, GPS can help determine the building where the user is located, limiting the number of visual markers the application will ping. Innovators in the AR-based indoor positioning space are using hybrid approaches like this to maximize precision and scale of AR positioning technologies.

CONCLUSION

AR-based indoor navigation has had few cases and requires further technical development before it can roll out on a large scale, but all technological evidence indicates that it will be one of the major indoor positioning technologies of the future.

This entry concludes our blog series on Indoor Positioning, we hope you enjoyed and learned from it! In case you missed it, check out our past entries:

The Future of Indoor GPS Part 1: Top Indoor Positioning Technologies

The Future of Indoor GPS Part 2: Bluetooth 5.1′s Angle of Arrival Ups the Ante for BLE Beacons

The Future of Indoor GPS Part 3: The Broadening Appeal of Ultra Wideband

The Future of Indoor GPS Part 4: Read the Room with RFID Tags

June 23, 2020Artificial Intelligence & Machine LearningGPS, Indoor, Inventory, Location, Management, Positioning, Reader, RFID, Tag, Trackingadmin

In the previous installment of our blog series on indoor positioning, we explored the future of Ultra Wideband technology. This week, we will examine RFID Tags.

The earliest applications of RFID tags date back to World War II when they were used to identify nearby planes as friends or foes. Since then, RFID technology has evolved to become one of the most cost-effective and easy to maintenance indoor positioning technologies on the market.

WHAT IS RFID?

RFID refers to a wireless system with two components: tags and readers. The reader is a device with one or more antennae emitting radio waves and receiving signals back from the RFID tag.

RFID tags are attached to assets like product inventory. RFID Readers enable users to automatically track and identify inventory and assets without a direct line of sight with a read range between a few centimeters and over 20 meters. They can contain a wide range of information, from merely a serial number to several pages of data. Readers can be mobile and carried by hand, mounted or embedded into the architecture of a room.

RFID tags use radio waves to communicate with nearby readers and can be passive or active. Passive tags are powered by the reader, do not require a battery, and have a read range of Near Contact – 25 Meters. Active tags require batteries and have an increased read range of 30 – 100+ Meters.

WHAT DOES RFID DO?

RFID is one of the most cost-effective and efficient location technologies. RFID chips are incredibly small—they can be placed underneath the skin without much discomfort to the host. For this reason, RFID tags are commonly used for pet identification.

One of the most widespread uses of RFID is in inventory management. When a unique tag is placed on each product, RFID tags offer instant updates on the total number of items within a warehouse or shop. In addition, it can offer a full database of information updated in real time.

RFID has also found several use cases in indoor positioning. For example, it can identify patients and medical equipment in hospitals using several readers spaced out in the building. The readers each identify their relative position to the tag to determine its location within the building. Supermarkets similarly use RFID to track products, shopping carts, and more.

RFID has found a wide variety of use cases, including:

WHAT ARE THE CONS OF USING RFID?

Perhaps the biggest obstacle facing businesses looking to adopt RFID for inventory tracking is pricing. RFID tags are significantly more expensive than bar codes, which can store some of the same data and offer similar functionality. At about $0.09, passive RFID tags are less expensive than active RFID tags, which can run from $25-$50. The cost of active RFID tags causes many businesses to only use them for high-inventory items.

RFID tags are also vulnerable to viruses, as is any technology that creates a broadcast signal. Encrypted data can help provide an extra level of security; however, security concerns still often prevents larger enterprises from utilizing them on the most high-end merchandise.

OVERALL

RFID tags are one of the elite technologies for offering inventory management with indoor positioning. Although UWB and Bluetooth BLE beacons offer more precise and battery-efficient location services, RFID is evolving to become more energy and cost efficient.

Stay tuned for the next entry in our Indoor Positioning blog series which will explore AR applications in indoor positioning!