Recently, Google AI has announced that the use of a machine learning system for rapid training to game-playing agents. This framework can be used by game developers to release human game-playing agents, who can efficiently focus on other priority duties to boost productivity. Google AI will provide an open-source library to show the techniques used in this practice. Google AI wants to harness machine learning algorithms for designers to balance their games, artists to generate assets of top-notch quality within a short period of time. The models can also be used to train challenging opponents to compete at the highest level of any game efficiently.

Traditionally, game developers leverage machine learning algorithms for direct access to the source, the uniquely interactive nature of video games, and many more functionalities. But, in the current scenario, Google AI has launched a machine learning system that can be used for rapid training to game-playing agents and seeking serious bugs instantly. The modern solution can work on some popular game genres and generate game actions from a game state within one hour on a single game. Google AI is determined to produce a machine learning system that can only play the game for the game developers while detecting and fixing bugs automatically.

This updated machine learning system can allow game developers to train multiple game-playing agents instead of one super-effective agent with a single end-to-end machine learning model. Game developers experienced a most fundamental barrier in implementing machine learning to computer games- bridging the gap between the simulation-centric world of games and the data-centric world of machine learning. The current machine learning system provides efficient as well as game-developer-friendly APIs with a Dagger-inspired interactive training flow to develop user-friendly video games by describing what a player perceives and the semantic actions related to it like joysticks, 3D objects, 3D locations, buttons, and many more.

Google AI is ready to provide an open-source library for game developers with no prior knowledge of machine learning while the training of game-playing agents can be finished within an hour on a single developer machine. The reason is the effective performance of imitation learning that teaches machine learning models by observing the behaviors of professional players in the games. The imitation model is inspired by the DAgger algorithm that allows taking an advantage of the interactivity of video games.

The post GOOGLE AI: MACHINE LEARNING FOR RAPID TRAINING TO GAME-PLAYING AGENTS appeared first on Artificial Intelligence.

Google has introduced an artificial intelligence-powered system aimed at helping lending institutions process coronavirus-related loans to small businesses.

The PPP Lending AI Solution, introduced at the end of last week, is specifically aimed at lenders participating in a US government scheme that offers loans to help small businesses stay in business through the Covid-19 pandemic.

The US Small Business Association’s Paycheck Protection Programme (PPP) offers loans that smaller organisations can use to keep staff on their payroll during the crisis.

But lenders participating in the programme have been overwhelmed by the number of applications coming in.

Analytics

Google said its scheme is an end-to-end system that can speed up the time it takes to make a decision on loans, while helping inform lenders’ liquidity analysis.

The programme’s tools extend “from the initial application submission to the underwriting process and SBA validation”, Google said.

The product is composed of three components that can be used together or individually, covering loan processing, document parsing and data analytics.

The Loan Processing Portal is a web-based application that allows lending agents or loan applicants create, submit and view the status of a PPP loan application, while the Document AI PPP Parser API allows lenders to use AI to extract structured information from PPP loan documents.

The third tool, Loan Analytics, allows lenders to access and analyse historical loan data, assist in the anonymisation of sensitive information and securely store data.

Cloud boom

Google added that the Document AI PPP Parser is available free of charge through 30 June.

Google’s cloud services arm has stepped up its enterprise focus over the past year, focusing on several key vertical industries, including financial services.

The company highlighted cloud features that allow financial organisations to quickly increase capacity when required and to shift workloads from mainframes to the cloud for added flexibility – features that have proven desireable during the pandemic.

“We are committed to maintaining the health of the systems that power the financial services industry, and will do everything we can to empower our customers’ business continuity planning and resilience,” Google said in a blog post.

The post Google AI Scheme Targets Banks Overwhelmed By Pandemic appeared first on Artificial Intelligence.

In an effort to help lenders expedite the processing of applications for the U.S. Small Business Administration’s (SBA) Paycheck Protection Program, which aims to keep workers employed during the coronavirus pandemic, Google developed an AI solution called PPP Lending AI that integrates with existing document ingestion tools. It’s available to eligible lending institutions through June 30.

As Google explains in a whitepaper, AI can automate the handling of volumes of loan applications by identifying patterns that would take a human worker longer to spot. Specifically, PPP Lending AI can classify and extract data in critical paperwork before readying documents for submission to the SBA.

PPP Lending AI, which Google says takes only days to implement, is a solution in three parts.

The first is the Loan Processing Portal, a web-based app that serves as a user interface and self-servicing center. In addition to providing administration views for loan officers and loan processors, it allows end users and loan applicants to create, submit, and view the status of their PPP loan.

The second piece of PPP Lending AI is the Document AI PPP Parser, which allows lenders to use AI to extract structured information from loan documents submitted by the loan applicants. It’s built atop Google Cloud‘s Document AI, a service that leverages optical character recognition, form parsing, and natural language processing to capture and enrich unstructured data.

The third is Loan Analytics, which lets servicers or lenders onboard structured historical loan data, perform de-identification anonymization on sensitive information, store it securely with fine-grained data access control, and perform queries on the data.

“Leveraging artificial intelligence, we’ve created an end-to-end solution that speeds up the time-to-decision on loans and helps inform lenders’ liquidity analysis — from the initial application submission to the underwriting process and SBA validation,” wrote Google Cloud global financial services and solutions lead Christin Brown in a blog post. “The solution is also equipped with Google’s security capabilities, enabling lenders to meet policy requirements and protect critical assets.”

Google says lenders can speak with a Google Cloud account manager for more information.

PPP Lending AI appears to skirt around a newly imposed U.S. Treasury and SBA rule prohibiting the submission of PPA loans prepared by robotic process automation (RPA), or AI systems that perform repetitive, monotonous tasks at scale with greater speed and accuracy than humans. The agencies blamed RPA for overburdening E-Tran, the SBA’s electronic loan servicing portal, and reducing its capabilities.

On Monday, E-Tran crashed minutes after the opening of $310 billion in additional PPP funding. The funds were approved last week following the first $349 billion round, which ran out in early April. That’s in spite of the fact that the SBA limited application submissions to 350 per hour and allowed banks with a minimum of 5,000 loans to bulk-file their applications.

PPP loans are available to small businesses that were in operation as of February 15 with 500 or fewer employees, including not-for-profits, veterans’ organizations, tribal concerns, self-employed individuals, sole proprietorships, and independent contractors. Businesses with more than 500 employees in certain industries can also apply for loans, according to the SBA and Treasury.

The post Google releases AI tool for processing Paycheck Protection Program loans appeared first on Artificial Intelligence.

In a paper originally published last October and accepted to the International Conference on Learning Representations (ICLR) 2020, researchers affiliated with Google and the University College London propose an AI model that enables control of speech characteristics like pitch, emotion, and speaking rate with as little as 30 minutes of data.

The work has obvious commercial implications. Brand voices such as Progressive’s Flo (played by comedian Stephanie Courtney) are often pulled in for pick-ups — sessions to address mistakes, changes, or additions in voiceover scripts — long after a recording finishes. AI-assisted voice correction could eliminate the need for these, saving time and money on the part of the actors’ employers.

A previous study investigated the use of so-called style tokens (which represented different categories of emotion) to control speech affect. The method achieved good results with only 5% of labeled data, but it couldn’t handle speech samples with varying prosody (i.e., intonation, tone, stress, and rhythm) and fixed emotion. The work from Google and the University of College London addresses this limitation.

The researchers trained the system for 300,000 steps across 32 of Google’s custom-designed tensor processing units (TPUs), a scale of compute exceeding that used in previous work. They report that using 30 minutes of labeled data allowed for a “significant degree” of control over speech rate, valence, and arousal, and that affect accuracy didn’t degrade noticeably with at least 10% of labeled data. The researchers said that just 3 minutes of data allowed for control of speech rate and extrapolation outside data seen during training — a result the researchers claim beat out state-of-the-art baselines.

The researchers’ system taps a trained generative model that can synthesize acoustic features from text. Similar to Google’s Tacotron 2, a text-to-speech (TTS) system that generates natural-sounding speech from raw transcripts, the new system can produce visual representations of frequencies called spectrograms by training a second model such as DeepMind’s WaveNet to act as a vocoder, a voice codec that analyzes and synthesizes voice data. (This system uses WaveRNN.)

An annotated data set comprising 72,405 roughly 5-second recordings from 40 English speakers, amounting to 45 hours of audio, was used to train the system. The speakers, all of whom were trained voice actors, were prompted to read text snippets with varying levels of valence (emotions like sadness or happiness) and arousal (excitement or energy). From these sessions, the researchers obtained six possible affective states, which they modeled and use as labels along with labels for speaking rate (here defined as the number of syllables per second in each utterance).

Here’s one of the voices the system modified (which sounds not unlike the default Google Assistant voice, interestingly) to have high arousal and an “angry” valence:

The study’s coauthors acknowledge that the work might raise ethical concerns because it could be misused for misinformation or to commit fraud. Indeed, deepfakes — media that takes a person in an existing image, audio recording, or video and replaces them with someone else’s likeness using AI — are multiplying quickly, and have already been used to defraud a major energy producer. In tandem with tools like Resemble, Baidu’s Deep Voice, and Lyrebird, which need only seconds to minutes of audio samples to clone someone’s voice, it’s not difficult to imagine how this new system might add fuel to the fire.

But the coauthors also assert that in this case, since the focus of this work is on improved prosody with potential benefits to human-computer interfaces, the benefits likely outweigh the risks. “We … urge the research community to take seriously the potential for misuse both of this work and broader advances in TTS,” they wrote.

The post Google’s AI can adjust voice emotion, pitch, and speed with 30 minutes of data appeared first on Artificial Intelligence.

In a recent study conducted in collaboration with Calico Life Sciences, Google researchers built a “genome-wide” machine learning model for the regulation of gene expression — the process by which information from a gene is used to create functional protein or RNA — in a species of yeast. While the work focused on yeast, it could be applicable to humans because it reveals how genes work together as a system, a core and only partially understood piece of the microbiological puzzle.

As the team explains in a technical paper and a blog post, yeast — which are single-celled organisms — grow old and die after budding (i.e., producing almost genetically identical offspring) 30 times. Budding produces “scars” on yeast cells that are visible under a powerful microscope, making it possible to determine the age of a cell from its appearance.

Leveraging this, Google Research’s Ted Baltz and team trained a model on a yeast growth data set produced by Calico, which contained the results of over 200 experiments on different yeast strains. In the course of each experiment, a single gene within the strains was activated and the expression levels of 6,000 genes were measured 8 times over 90 minutes, yielding a total of almost 20 million individual measurements.

The Google researchers’ approach was to model the whole data set as a system of differential equations, such that the rate of change of the expression of a gene was proportional to a weighted sum of the expression levels of all genes. Baltz reports that in the end, the work amounted to more than 50 million regularization paths, which informed predictions about which genes would code for regulators (i.e., genes involved in controlling the expression of one or more other genes).

To verify the model’ss predictions, the researchers tested it against a validation data set comprising ten new yeast strains. They report that three out of the ten predictions held up in experiments, including one gene that hadn’t previously been identified by scientists.

“Based on exhaustive experiments, we built a genome-wide model for the regulation of gene expression in [yeast] and verified some of the results experimentally, enabling future investigations into less well understood biological systems,” wrote Baltz. “Our model was able to identify these without prior biological knowledge, demonstrating that these [machine learning] techniques might scale to other domains or organisms that are much less well studied.”

Google’s work in AI and gene expression follows the publication of a study describing a “massively parallel reporter assay” (MPRA),” a framework designed to investigate DNA. The researchers claimed it could be used to create AI models that predict gene regulation for industrial and life science applications. An older work proposes a unified AI architecture to model and interpret how chromatin, a complex of DNA and protein found in eukaryotic cells, controls gene regulation.

The post Google’s AI attempts to predict how yeast cell genes will express themselves appeared first on Artificial Intelligence.

Google’s AI teams used the comparatively quiet post-easter days to get ML practitioners up to speed with their latest research in reinforcement learning, natural language processing, and computer vision.

In “An optimistic perspective on offline reinforcement learning”, a team of researchers has looked into ways to use a fixed offline dataset of logged interactions to teach agents how to handle themselves in real world situations. While agents normally learn by getting live feedback from their environment, this approach is meant to be useful in certain robotics use cases or autonomous driving, where enough recorded interaction data is available and other ways of collecting information either seem insufficient or are too expensive to realise.

These are usually seen as tricky to implement, since there’s no real way of knowing how an agent should be rewarded when it takes an action that differs from the dataset provided. To tackle that, Google’s AI team added some supervised learning methods into the mix which helps to improve generalisation and make the whole system more robust. The results are called Ensemble-DQN and Random Ensemble Mixture and can be investigated here.

Meanwhile another team has been busy improving the way objects are detected. The outcome has been dubbed EfficientDet and will be presented at the renowned computer vision conference CVPR in Seattle in June – if COVID refrains from putting a spoke in their wheel. It aims at introducing a “new family of scalable and efficient object detectors” to the computer vision community, building upon earlier work concerning the scaling of neural networks (EfficientNet).

In EfficientDet, EfficientNet is used as a backbone to more effectively extract features from images, while a new bi-directional feature network in combination with a fresh normalised fusion technique is meant to get to image characteristics faster at a lower computation cost.

If you’re more interested in NLP, Google has also been busy setting up a benchmark to make comparing multilingual representations easier. XTREME covers 40 languages from 12 language families and includes nine tasks ranging from sentence classification to question answering to evaluate methods making the most of the shared structures of languages. The project can be found at GitHub.

Git pushes out security updates to stop tricksters

This week, Git maintainer Junio C Hamano has unleashed versions v2.26.1, v2.25.3, v2.24.2, v2.23.2, v2.22.3, v2.21.2, v2.20.3, v2.19.4, v2.18.3, and v2.17.4 of the version control system onto the coding masses. 

Updating is strongly advised, since the security fixes mediate an issue which “allowed a crafted URL to trick a Git client to send credential information for a wrong host to the attacker’s site”. 

Gloo lures admins with new dev portal

Envoy-based API gateway Gloo hit version 1.3 earlier this week, focusing on performance, stability and extensibility improvements. However, the release also includes a developer portal, so that admins have an easier way of controlling who gets access to which APIs.

Once set up, they can select which interfaces should be shared at all and decide which users and groups get to see them once they’ve logged into the portal. The whole apparatus is designed for self-service, with the Gloo team promising easy integration into continuous delivery processes.

The post What’s the point: Google AI advances, Git security updates, and API gateway Gloo appeared first on Artificial Intelligence.

Machine learning (ML) is a method by which algorithms adapt their activity using inputted data, rather than being programmed to do so. But building and “training” these algorithms takes time, and can often ingrain human biases.

To overcome these limitations, and enable further innovation in machine learning, researchers have explored the field of AutoML, whereby the machine learning process can be progressively automated, relying on machine compute time, rather than human research time.

So far, although some steps have been automated, the benchmark of virtually zero human input has yet to be attained. However, a team of scientists from Google have seen some “preliminary success” in discovering machine learning algorithms from scratch, indicating a “promising new direction for the field.”

In a paper, published on the preprint server arXiv, Quoc Le, a computer scientist at Google, and colleagues, employed concepts from Darwinian evolution, such as natural selection, to enable ML algorithms to improve generation upon generation. Combining basic mathematical operations, their program, called AutoML-Zero, generated 100 unique algorithms that they then tested on simple tasks, such as image recognition.

After their performance was compared to hand-designed algorithms, the best were kept, and small random “mutations” in their code were introduced, whilst the weaker candidates were removed. As the cycle continued, a high-performing set of algorithms were found, some of which are comparable to a number of classic machine learning techniques – such as neural networks (a kind of computer program that loosely mimics how our brain cells work together to make decisions).

This proves the team’s concept, Le told Science Magazine, but he is hopeful that the processes can be scaled up to eventually create much more complex AIs, which human researchers could never find.

“Our goal is to show that AutoML can go further: it is possible today to automatically discover complete machine learning algorithms just using basic mathematical operations as building blocks,” the team wrote in the paper, which is awaiting peer-review.

“Starting from empty component functions and using only basic mathematical operations, we evolved linear regressors, neural networks, gradient descent, multiplicative interactions, weight averaging, normalized gradients, etc.” the authors continued. “These results are promising, but there is still much work to be done.”

The post Google Scientists Develop Software That Could Enable AI To Evolve With No Human Input appeared first on Artificial Intelligence.

Science-fiction author Vernor Vinge once said that mankind’s last great invention will be the first self-replicating machine. Now, AI scientists working in Google Brain division are testing how machine learning algorithms can be created from scratch, then evolve naturally, based on simple math, according to Google’s AutoML team who suggest the software could potentially be updated to “automatically discover” completely unknown algorithms while also reducing human bias during the data input process. The software, known as AutoML-Zero, resembles the process of evolution, with code improving every generation with little human interaction.

Machine learning tools are “trained” to find patterns in vast amounts of data while automating such processes and constantly being refined based on past experience.

But there’s a draw back, “Human-designed components bias the search results in favor of human-designed algorithms, possibly reducing the innovation potential of AutoML,” according to the team’s paper. “Innovation is also limited by having fewer options: you cannot discover what you cannot search for.” The analysis, which was published last month on arXiv, is titled “Evolving Machine Learning Algorithms From Scratch”.

In an interview with Newsweek, Haran Jackson, the chief technology officer (CTO) at Techspert, who has a PhD in Computing from the University of Cambridge, said that AutoML tools are typically used to “identify and extract” the most useful features from datasets—and this approach is a welcome development.

“There is a sense,” he added, “that among many members of the community that the most impressive feats of artificial intelligence will only be achieved with the invention of new algorithms that are fundamentally different to those that we as a species have so far devised. This is what makes the aforementioned paper so interesting. It presents a method by which we can automatically construct and test completely novel machine learning algorithms.”

Jackson concluded that the approach taken was similar to the theory of evolution proposed by Charles Darwin, noting how the Google team was able to induce “mutations” into the set of algorithms. “The mutated algorithms that did a better job of solving real-world problems were kept alive, with the poorly-performing ones being discarded.

The post Google AI Scientists –“Developing Algorithms that Mirror Darwinian Evolution” appeared first on Artificial Intelligence.

Google today announced TensorFlow Lite Model Maker, a tool that adapts state-of-the-art machine learning models to custom data sets using a technique known as transfer learning. It wraps machine learning concepts with an API that enables developers to train models in Google’s TensorFlow AI framework with only a few lines of code, and to deploy those models for on-device AI applications.

Tools like Model Maker could help companies incorporate AI into their workflows faster than before. According to a study conducted by Algorithmia, 50% of organizations spend between 8 and 90 days deploying a single machine learning model, with most blaming the duration on a failure to scale.

Model Maker, which currently only supports image and text classification use cases, works with many of the models in TensorFlow Hub, Google’s library for reusable machine learning modules. (“Modules” in this context refers to self-contained algorithms along with assets that can be used across different AI tasks.) Essentially, Model Maker applies models trained on one task to another related task at varying levels of accuracy, according to several parameters specified at the outset.

Model accuracy can be improved with Model Maker by changing the model architecture, which requires editing one line of code. After the input data specific to an on-device AI is loaded in, Model Maker evaluates the model and exports it as a TensorFlow Lite model. (TensorFlow Lite is a version of TensorFlow that’s optimized for mobile, embedded, and internet of things devices.)

Models created by TensorFlow Lite Model Maker have metadata attached to them, including machine-readable parameters like mean, standard deviation, category label files, and human-readable parameters such as model descriptions and licenses. Google notes that fields like licenses can be critical in deciding whether a model can be used, while other systems can use the machine-readable parameters to generate wrapper code.

In the coming months, Google intends to enhance Model Maker to support more tasks, including object detection and several natural language processing tasks. Specifically, it says it’ll add BERT, a pretraining technique for natural language processing, for applications such as question-and-answer.

The launch of Model Maker follows on the heels of an API — Quantization Aware Training (QAT) — that trains smaller, faster TensorFlow models with the performance benefits of quantization (the process of mapping input values from a large set to output values in a smaller set) while retaining close to their original accuracy. Earlier in the year, Google unveiled TensorFlow Quantum, a machine learning framework for training quantum models, at the TensorFlow Dev Summit.

The post Google’s TensorFlow Lite Model Maker adapts state-of-the-art models for on-device AI appeared first on Artificial Intelligence.

Carnegie Mellon, Google, and Stanford researchers write in a paper that they’ve developed a framework for using weak supervision — a form of AI training where the model learns from large amounts of limited, imprecise, or noisy data — that enables robots to efficiently explore a challenging environment. By learning to reach only areas of its surroundings relevant to tasks as opposed to every corner, the researchers say their approach speeds up training on various robot manipulation tasks.

The teams’ framework — Weakly-Supervised Control (WSC) — learns a corpus with which a software agent can generate its own goals and perform exploration. It incorporates reinforcement learning, a form of training that spurs agents to accomplish goals via rewards. But unlike traditional reinforcement learning, which requires hand-designed rewards that are computationally expensive to obtain, WSC frames the weakly-supervised learning problem in a way that provides a form of supervision scalable with the collection of data, and that doesn’t require labels in the reinforcement learning loop.

In experiments, the researchers sought to determine whether weak supervision was necessary for learning a disentangled state representation — i.e., a set of features influenced by the actions of the agent. They tasked several models with simulated vision-based, goal-conditioned manipulation tasks of varying complexity: in one environment, agents were tasked with moving a specific object to a goal location, and in another, the agents had to open a door to match a goal angle.

The coauthors report that WSC learned more quickly than prior state-of-the-art goal-conditioned reinforcement learning methods, particularly as the complexity of the agents’ various environments grew. Moreover, they say that WSC attained higher correlation between latent goals and final states, indicating that it learned a more interpretable goal-conditioned policy.

However, the researchers concede that WSC isn’t without its limitations. It requires a user to indicate the factors relevant for downstream tasks, which might require expertise, and it only uses weak supervision during pre-training, which might produce representations that don’t generalize to new interactions encountered by the agent. This said, they hope in future work to investigate other forms of weak supervision that can provide useful signals to agents, as well as other ways to leverage these labels.

“Given the promising results in increasingly complex environments, evaluating this approach with robots in real-world environments is an exciting future direction,” wrote the coauthors. “Overall, we believe that our framework provides a new perspective on supervising the development of general-purpose agents acting in complex environments.”

The post Google’s AI learns how to navigate environments from limited data appeared first on Artificial Intelligence.