Category Archives: Machine Learning

The following was written byHarald Collet, CEO at Alkymi andHugues Chabanis, Product Portfolio Manager,Alternative Investments at SimCorp

Institutional investors are buckling under the operational constraint of processing hundreds of data streams from unstructured data sources such as email, PDF documents, and spreadsheets. These data formats bury employees in low-value copy-paste workflows andblockfirms from capturing valuable data. Here, we explore how Machine Learning(ML)paired with a better operational workflow, can enable firms to more quickly extract insights for informed decision-making, and help governthe value of data.

According to McKinsey, the average professional spends 28% of the workday reading and answering an average of 120 emails on top ofthe19% spent on searching and processing data.The issue is even more pronouncedininformation-intensive industries such as financial services,asvaluable employees are also required to spendneedlesshoursevery dayprocessing and synthesizing unstructured data. Transformational change, however,is finally on the horizon. Gartner research estimates thatby 2022, one in five workers engaged in mostly non-routine tasks will rely on artificial intelligence (AI) to do their jobs. And embracing ML will be a necessity for digital transformation demanded both by the market and the changing expectations of the workforce.

For institutional investors that are operating in an environment of ongoing volatility, tighter competition, and economic uncertainty, using ML to transform operations and back-office processes offers a unique opportunity. In fact, institutional investors can capture up to 15-30% efficiency gains by applying ML and intelligent process automation (Boston Consulting Group, 2019)inoperations,which in turn creates operational alpha withimproved customer service and redesigning agile processes front-to-back.

Operationalizingmachine learningworkflows

ML has finally reached the point of maturity where it can deliver on these promises. In fact, AI has flourished for decades, but the deep learning breakthroughs of the last decade has played a major role in the current AI boom. When it comes to understanding and processing unstructured data, deep learning solutions provide much higher levels of potential automation than traditional machine learning or rule-based solutions. Rapid advances in open source ML frameworks and tools including natural language processing (NLP) and computer vision have made ML solutions more widely available for data extraction.

Asset class deep-dive: Machine learning applied toAlternative investments

In a 2019 industry survey conducted byInvestOps, data collection (46%) and efficient processing of unstructured data (41%) were cited as the top two challenges European investment firms faced when supportingAlternatives.

This is no surprise as Alternatives assets present an acute data management challenge and are costly, difficult, and complex to manage, largely due to the unstructured nature ofAlternatives data. This data is typically received by investment managers in the form of email with a variety of PDF documents or Excel templates that require significant operational effort and human understanding to interpret, capture,and utilize. For example, transaction data istypicallyreceived by investment managers as a PDF document via email oran online portal. In order to make use of this mission critical data, the investment firm has to manually retrieve, interpret, and process documents in a multi-level workflow involving 3-5 employees on average.

The exceptionally low straight-through-processing (STP) rates already suffered by investment managers working with alternative investments is a problem that will further deteriorate asAlternatives investments become an increasingly important asset class,predictedbyPrequinto rise to $14 trillion AUM by 2023 from $10 trillion today.

Specific challenges faced by investment managers dealing with manual Alternatives workflows are:

WithintheAlternatives industry, variousattempts have been madeto use templatesorstandardize the exchange ofdata. However,these attempts have so far failed,or are progressing very slowly.

Applying ML to process the unstructured data will enable workflow automation and real-time insights for institutional investment managers today, without needing to wait for a wholesale industry adoption of a standardized document type like the ILPA template.

To date, the lack of straight-through-processing (STP) in Alternatives has either resulted in investment firms putting in significant operational effort to build out an internal data processing function,or reluctantly going down the path of adopting an outsourcing workaround.

However, applyinga digital approach,more specificallyML, to workflows in the front, middle and back office can drive a number of improved outcomes for investment managers, including:

Trust and control are critical when automating critical data processingworkflows.This is achieved witha human-in-the-loopdesign that puts the employee squarely in the drivers seat with features such as confidence scoring thresholds, randomized sampling of the output, and second-line verification of all STP data extractions. Validation rules on every data element can ensure that high quality output data is generated and normalized to a specific data taxonomy, making data immediately available for action. In addition, processing documents with computer vision can allow all extracted data to be traced to the exact source location in the document (such as a footnote in a long quarterly report).

Reverse outsourcing to govern the value of your data

Big data is often considered the new oil or super power, and there are, of course, many third-party service providers standing at the ready, offering to help institutional investors extract and organize the ever-increasing amount of unstructured, big data which is not easily accessible, either because of the format (emails, PDFs, etc.) or location (web traffic, satellite images, etc.). To overcome this, some turn to outsourcing, but while this removes the heavy manual burden of data processing for investment firms, it generates other challenges, including governance and lack of control.

Embracing ML and unleashing its potential

Investment managers should think of ML as an in-house co-pilot that can help its employees in various ways: First, it is fast, documents are processed instantly and when confidence levels are high, processed data only requires minimum review. Second, ML is used as an initial set of eyes, to initiate proper workflows based on documents that have been received. Third, instead of just collecting the minimum data required, ML can collect everything, providing users with options to further gather and reconcile data, that may have been ignored and lost due to a lack of resources. Finally, ML will not forget the format of any historical document from yesterday or 10 years ago safeguarding institutional knowledge that is commonly lost during cyclical employee turnover.

ML has reached the maturity where it can be applied to automate narrow and well-defined cognitive tasks and can help transform how employees workin financial services. However many early adopters have paid a price for focusing too much on the ML technology and not enough on the end-to-end business process and workflow.

The critical gap has been in planning for how to operationalize ML for specific workflows. ML solutions should be designed collaboratively with business owners and target narrow and well-defined use cases that can successfully be put into production.

Alternatives assets are costly, difficult, and complex to manage, largely due to the unstructured nature of Alternatives data. Processing unstructured data with ML is a use case that generates high levels of STP through the automation of manual data extraction and data processing tasks in operations.

Using ML to automatically process unstructured data for institutional investors will generate operational alpha; a level of automation necessary to make data-driven decisions, reduce costs, and become more agile.

The views represented in this commentary are those of its author and do not reflect the opinion of Traders Magazine, Markets Media Group or its staff. Traders Magazine welcomes reader feedback on this column and on all issues relevant to the institutional trading community.

More:
Machine Learning: Making Sense of Unstructured Data and Automation in Alt Investments - Traders Magazine

Machine learning is by no means a new phenomenon. It has been used in various forms for decades, but it is very much a technology of the present due to the massive increase in the data upon which it thrives. It has been widely adopted by businesses, reducing the time and improving the value of the insight they can distil from large volumes of customer data.

However, in the public sector there is a different story. Despite being championed by some in government, machine learning has often faced a reaction of concern and confusion. This is not intended as general criticism and in many cases it reflects the greater value that civil servants place on being ethical and fair, than do some commercial sectors.

One fear is that, if the technology is used in place of humans, unfair judgements might not be noticed or costly mistakes in the process might occur. Furthermore, as many decisions being made by government can dramatically affect peoples lives and livelihood then often decisions become highly subjective and discretionary judgment is required. There are also those still scarred by films such as iRobot, but thats a discussion for another time.

Fear of the unknown is human nature, so fear of unfamiliar technology is thus common. But fears are often unfounded and providing an understanding of what the technology does is an essential first step in overcoming this wariness. So for successful digital transformation not only do the civil servants who are considering such technologies need to become comfortable with its use but the general public need to be reassured that the technology is there to assist, not replace, human decisions affecting their future health and well-being.

Theres a strong case to be made for greater adoption of machine learning across a diverse range of activities. The basic premise of machine learning is that a computer can derive a formula from looking at lots of historical data that enables the prediction of certain things the data describes. This formula is often termed an algorithm or a model. We use this algorithm with new data to make decisions for a specific task, or we use the additional insight that the algorithm provides to enrich our understanding and drive better decisions.

For example, machine learning can analyse patients interactions in the healthcare system and highlight which combinations of therapies in what sequence offer the highest success rates for patients; and maybe how this regime is different for different age ranges. When combined with some decisioning logic that incorporates resources (availability, effectiveness, budget, etc.) its possible to use the computers to model how scarce resources could be deployed with maximum efficiency to get the best tailored regime for patients.

When we then automate some of this, machine learning can even identify areas for improvement in real time and far faster than humans and it can do so without bias, ulterior motives or fatigue-driven error. So, rather than being a threat, it should perhaps be viewed as a reinforcement for human effort in creating fairer and more consistent service delivery.

Machine learning is an iterative process; as the machine is exposed to new data and information, it adapts through a continuous feedback loop, which in turn provides continuous improvement. As a result, it produces more reliable results over time and evermore finely tuned and improved decision-making. Ultimately, its a tool for driving better outcomes.

The opportunities for AI to enhance service delivery are many. Another example in healthcare is Computer Vision (another branch of AI), which is being used in cancer screening and diagnosis. Were already at the stage where AI, trained from huge libraries of images of cancerous growths, is better at detecting cancer than human radiologists. This application of AI has numerous examples, such as work being done at Amsterdam UMC to increase the speed and accuracy of tumour evaluations.

But lets not get this picture wrong. Here, the true value is in giving the clinician more accurate insight or a second opinion that informs their diagnosis and, ultimately, the patients final decision regarding treatment. A machine is there to do the legwork, but the human decision to start a programme for cancer treatment, remains with the humans.

Acting with this enhanced insight enables doctors to become more efficient as well as effective. Combining the results of CT scans with advanced genomics using analytics, the technology can assess how patients will respond to certain treatments. This means clinicians avoid the stress, side effects and cost of putting patients through procedures with limited efficacy, while reducing waiting times for those patients whose condition would respond well. Yet, full-scale automation could run the risk of creating a lot more VOMIT.

Victims Of Modern Imaging Technology (VOMIT) is a new phenomenon where a condition such as a malignant tumour is detected by imaging and thus at first glance it would seem wise to remove it. However, medical procedures to remove it carry a morbidity risk which may be greater than the risk the tumour presents during the patients likely lifespan. Here, ignorance could be bliss for the patient and doctors would examine the patient holistically, including mental health, emotional state, family support and many other factors that remain well beyond the grasp of AI to assimilate into an ethical decision.

All decisions like these have a direct impact on peoples health and wellbeing. With cancer, the faster and more accurate these decisions are, the better. However, whenever cost and effectiveness are combined there is an imperative for ethical judgement rather than financial arithmetic.

Healthcare is a rich seam for AI but its application is far wider. For instance, machine learning could also support policymakers in planning housebuilding and social housing allocation initiatives, where they could both reduce the time for the decision but also make it more robust. Using AI in infrastructural departments could allow road surface inspections to be continuously updated via cheap sensors or cameras in all council vehicles (or cloud-sourced in some way). The AI could not only optimise repair work (human or robot) but also potentially identify causes and then determine where strengthened roadways would cost less in whole-life costs versus regular repairs or perhaps a different road layout would reduce wear.

In the US, government researchers are already using machine learning to help officials make quick and informed policy decisions on housing. Using analytics, they analyse the impact of housing programmes on millions of lower-income citizens, drilling down into factors such as quality of life, education, health and employment. This instantly generates insightful, accessible reports for the government officials making the decisions. Now they can enact policy decisions as soon as possible for the benefit of residents.

While some of the fears about AI are fanciful, there is a genuine cause for concern about the ethical deployment of such technology. In our healthcare example, allocation of resources based on gender, sexuality, race or income wouldnt be appropriate unless these specifically had an impact on the prescribed treatment or its potential side-effects. This is self-evident to a human, but a machine would need this to be explicitly defined. Logically, a machine would likely display bias to those groups whose historical data gave better resultant outcomes, thus perpetuating any human equality gap present in the training data.

The recent review by the Committee on Standards in Public Life into AI and its ethical use by government and other public bodies concluded that there are serious deficiencies in regulation relating to the issue, although it stopped short of recommending the establishment of a new regulator.

The review was chaired by crossbench peer Lord Jonathan Evans, who commented:

Explaining AI decisions will be the key to accountability but many have warned of the prevalence of Black Box AI. However our review found that explainable AI is a realistic and attainable goal for the public sector, so long as government and private companies prioritise public standards when designing and building AI systems.

Fears of machine learning replacing all human decision-making need to be debunked as myth: this is not the purpose of the technology. Instead, it must be used to augment human decision-making, unburdening them from the time-consuming job of managing and analysing huge volumes of data. Once its role can be made clear to all those with responsibility for implementing it, machine learning can be applied across the public sector, contributing to life-changing decisions in the process.

Find out more on the use of AI and machine learning in government.

Simon Dennis, Director of AI & Analytics Innovation, SAS UK

Here is the original post:
Machine learning: the not-so-secret way of boosting the public sector - ITProPortal

The legal profession, the technology industry and the relationship between the two are in a state of transition. Computer processing power has doubled every year for decades, leading to an explosion in corporate data and increasing pressure on lawyers entrusted with reviewing all of this information.

Now, the legal industry is undergoing significant change, with the advent of machine learning technology fundamentally reshaping the way lawyers conduct their day-to-day practice. Indeed, whilst technological gains might once have had lawyers sighing at the ever-increasing stack of documents in the review pile, technology is now helping where it once hindered. For the first time ever, advanced algorithms allow lawyers to review entire document sets at a glance, releasing them from wading through documents and other repetitive tasks. This means legal professionals can conduct their legal review with more insight and speed than ever before, allowing them to return to the higher-value, more enjoyable aspect of their job: providing counsel to their clients.

In this article, we take a look at how this has been made possible.

Practicing law has always been a document and paper-heavy task, but manually reading huge volumes of documentation is no longer feasible, or even sustainable, for advisors. Even conservatively, it is estimated that we create 2.5 quintillion bytes of data every day, propelled by the usage of computers, the growth of the Internet of Things (IoT) and the digitalisation of documents. Many lawyers have had no choice but resort to sampling only 10 per cent of documents, or, alternatively, rely on third-party outsourcing to meet tight deadlines and resource constraints. Whilst this was the most practical response to tackle these pressures, these methods risked jeopardising the quality of legal advice lawyers could give to their clients.

Legal technology was first developed in the early 1970s to take some of the pressure off lawyers. Most commonly, these platforms were grounded on Boolean search technology, requiring months and even years building the complex sets of rules. As well as being expensive and time-intensive, these systems were also unable to cope with the unpredictable, complex and ever-changing nature of the profession, requiring significant time investment and bespoke configuration for every new challenge that arose. Not only did this mean lawyers were investing a lot of valuable time and resources training a machine, but the rigidity of these systems limited the advice they could give to their clients. For instance, trying to configure these systems to recognise bespoke clauses or subtle discrepancies in language was a near impossibility.

Today, machine learning has become advanced enough that it has many practical applications, a key one being legal document review.

Machine learning can be broadly categorised into two types: supervised and unsupervised machine learning. Supervised machine learning occurs when a human interacts with the system in the case of the legal profession, this might be tagging a document, or categorising certain types of documents, for example. The machine then builds its understanding to generate insights to the user based on this human interaction.

Unsupervised machine learning is where the technology forms an understanding of a certain subject without any input from a human. For legal document review, the unsupervised machine learning will cluster similar documents and clauses, along with clear outliers from those standards. Because the machine requires no a priori knowledge of what the user is looking for, the system may indicate anomalies or unknown unknowns- data which no one had set out to identify because they didnt know what to look for. This allows lawyers to uncover critical hidden risks in real time.

It is the interplay between supervised and unsupervised machine learning that makes technology like Luminance so powerful. Whilst the unsupervised part can provide lawyers with an immediate insight into huge document sets, these insights only increase with every further interaction, with the technology becoming increasingly bespoke to the nuances and specialities of a firm.

This goes far beyond more simplistic contract review platforms. Machine learning algorithms, such as those developed by Luminance, are able to identify patterns and anomalies in a matter of minutes and can form an understanding of documents both on a singular level and in their relationship to each another. Gone are the days of implicit bias being built into search criteria, since the machine surfaces all relevant information, it remains the responsibility of the lawyer to draw the all-important conclusions. But crucially, by using machine learning technology, lawyers are able to make decisions fully appraised of what is contained within their document sets; they no longer need to rely on methods such as sampling, where critical risk can lay undetected. Indeed, this technology is designed to complement the lawyers natural patterns of working, for example, providing results to a clause search within the document set rather than simply extracting lists of clauses out of context. This allows lawyers to deliver faster and more informed results to their clients, but crucially, the lawyer is still the one driving the review.

With the right technology, lawyers can cut out the lower-value, repetitive work and focus on complex, higher-value analysis to solve their clients legal and business problems, resulting in time-savings of at least 50 per cent from day one of the technology being deployed. This redefines the scope of what lawyers and firms can achieve, allowing them to take on cases which would have been too time-consuming or too expensive for the client if they were conducted manually.

Machine learning is offering lawyers more insight, control and speed in their day-to-day legal work than ever before, surfacing key patterns and outliers in huge volumes of data which would normally be impossible for a single lawyer to review. Whether it be for a due diligence review, a regulatory compliance review, a contract negotiation or an eDiscovery exercise, machine learning can relieve lawyers from the burdens of time-consuming, lower value tasks and instead frees them to spend more time solving the problems they have been extensively trained to do.

In the years to come, we predict a real shift in these processes, with the latest machine learning technology advancing and growing exponentially, and lawyers spending more time providing valuable advice and building client relationships. Machine learning is bringing lawyers back to the purpose of their jobs, the reason they came into the profession and the reason their clients value their advice.

James Loxam, CTO, Luminance

Follow this link:
The impact of machine learning on the legal industry - ITProPortal

Both weather and climate models have improved drastically in recent years, as advances in one field have tended to benefit the other. But there is still significant uncertainty in model outputs that are not quantified accurately. Thats because the processes that drive climate and weather are chaotic, complex, and interconnected in ways that researchers have yet to describe in the complex equations that power numerical models.

Historically, researchers have used approximations called parameterizations to model the relationships underlying small-scale atmospheric processes and their interactions with large-scale atmospheric processes. Stochastic parameterizations have become increasingly common for representing the uncertainty in subgrid-scale processes, and they are capable of producing fairly accurate weather forecasts and climate projections. But its still a mathematically challenging method. Now researchers are turning to machine learning to provide more efficiency to mathematical models.

Here Gagne et al. evaluate the use of a class of machine learning networks known as generative adversarial networks (GANs) with a toy model of the extratropical atmospherea model first presented by Edward Lorenz in 1996 and thus known as the L96 system that has been frequently used as a test bed for stochastic parameterization schemes. The researchers trained 20 GANs, with varied noise magnitudes, and identified a set that outperformed a hand-tuned parameterization in L96. The authors found that the success of the GANs in providing accurate weather forecasts was predictive of their performance in climate simulations: The GANs that provided the most accurate weather forecasts also performed best for climate simulations, but they did not perform as well in offline evaluations.

The study provides one of the first practically relevant evaluations for machine learning for uncertain parameterizations. The authors conclude that GANs are a promising approach for the parameterization of small-scale but uncertain processes in weather and climate models. (Journal of Advances in Modeling Earth Systems (JAMES), https://doi.org/10.1029/2019MS001896, 2020)

Kate Wheeling, Science Writer

See more here:
Machine Learning Improves Weather and Climate Models - Eos

Despite the huge contributions of deep learning to the field of artificial intelligence, theres something very wrong with it: It requires huge amounts of data. This is one thing that boththe pioneersandcritics of deep learningagree on. In fact, deep learning didnt emerge as the leading AI technique until a few years ago because of the limited availability of useful data and the shortage of computing power to process that data.

Reducing the data-dependency of deep learning is currently among the top priorities of AI researchers.

In hiskeynote speech at the AAAI conference, computer scientist Yann LeCun discussed the limits of current deep learning techniques and presented the blueprint for self-supervised learning, his roadmap to solve deep learnings data problem. LeCun is one of thegodfathers of deep learningand the inventor ofconvolutional neural networks (CNN), one of the key elements that have spurred a revolution in artificial intelligence in the past decade.

Self-supervised learning is one of several plans to create data-efficient artificial intelligence systems. At this point, its really hard to predict which technique will succeed in creating the next AI revolution (or if well end up adopting a totally different strategy). But heres what we know about LeCuns masterplan.

First, LeCun clarified that what is often referred to as the limitations of deep learning is, in fact, a limit ofsupervised learning. Supervised learning is the category of machine learning algorithms that require annotated training data. For instance, if you want to create an image classification model, you must train it on a vast number of images that have been labeled with their proper class.

[Deep learning] is not supervised learning. Its not justneural networks. Its basically the idea of building a system by assembling parameterized modules into a computation graph, LeCun said in his AAAI speech. You dont directly program the system. You define the architecture and you adjust those parameters. There can be billions.

Deep learning can be applied to different learning paradigms, LeCun added, including supervised learning,reinforcement learning, as well as unsupervised or self-supervised learning.

But the confusion surrounding deep learning and supervised learning is not without reason. For the moment, the majority of deep learning algorithms that have found their way into practical applications are based on supervised learning models, which says a lot aboutthe current shortcomings of AI systems. Image classifiers, facial recognition systems, speech recognition systems, and many of the other AI applications we use every day have been trained on millions of labeled examples.

Reinforcement learning and unsupervised learning, the other categories of learning algorithms, have so far found very limited applications.

Supervised deep learning has given us plenty of very useful applications, especially in fields such ascomputer visionand some areas of natural language processing. Deep learning is playing an increasingly important role in sensitive applications, such as cancer detection. It is also proving to be extremely useful in areas where the scale of the problem is beyond being addressed with human efforts, such aswith some caveatsreviewing the huge amount of content being posted on social media every day.

If you take deep learning from Facebook, Instagram, YouTube, etc., those companies crumble, LeCun says. They are completely built around it.

But as mentioned, supervised learning is only applicable where theres enough quality data and the data can capture the entirety of possible scenarios. As soon as trained deep learning models face novel examples that differ from their training examples, they start to behave in unpredictable ways. In some cases,showing an object from a slightly different anglemight be enough to confound a neural network into mistaking it with something else.

ImageNet vs reality: In ImageNet (left column) objects are neatly positioned, in ideal background and lighting conditions. In the real world, things are messier (source: objectnet.dev)

Deep reinforcement learning has shownremarkable results in games and simulation. In the past few years, reinforcement learning has conquered many games that were previously thought to off-limits for artificial intelligence. AI programs have already decimated human world champions atStarCraft 2, Dota, and the ancient Chinese board game Go.

But the way these AI programs learn to solve problems is drastically different from that of humans. Basically, a reinforcement learning agent starts with a blank slate and is only provided with a basic set of actions it can perform in its environment. The AI is then left on its own to learn through trial-and-error how to generate the most rewards (e.g., win more games).

This model works when the problem space is simple and you have enough compute power to run as many trial-and-error sessions as possible. In most cases, reinforcement learning agents take an insane amount of sessions to master games. The huge costs have limited reinforcement learning research to research labsowned or funded by wealthy tech companies.

Reinforcement learning agents must be trained on hundreds of years worth of session to master games, much more than humans can play in a lifetime (source: Yann LeCun).

Reinforcement learning systems are very bad attransfer learning. A bot that plays StarCraft 2 at grandmaster level needs to be trained from scratch if it wants to play Warcraft 3. In fact, even small changes to the StarCraft game environment can immensely degrade the performance of the AI. In contrast, humans are very good at extracting abstract concepts from one game and transferring it to another game.

Reinforcement learning really shows its limits when it wants to learn to solve real-world problems that cant be simulated accurately. What if you want to train a car to drive itself? And its very hard to simulate this accurately, LeCun said, adding that if we wanted to do it in real life, we would have to destroy many cars. And unlike simulated environments, real life doesnt allow you to run experiments in fast forward, and parallel experiments, when possible, would result in even greater costs.

LeCun breaks down the challenges of deep learning into three areas.

First, we need to develop AI systems that learn with fewer samples or fewer trials. My suggestion is to use unsupervised learning, or I prefer to call it self-supervised learning because the algorithms we use are really akin to supervised learning, which is basically learning to fill in the blanks, LeCun says. Basically, its the idea of learning to represent the world before learning a task. This is what babies and animals do. We run about the world, we learn how it works before we learn any task. Once we have good representations of the world, learning a task requires few trials and few samples.

Babies develop concepts of gravity, dimensions, and object persistence in the first few months after their birth. While theres debate on how much of these capabilities are hardwired into the brain and how much of it is learned, what is for sure is that we develop many of our abilities simply by observing the world around us.

The second challenge is creating deep learning systems that can reason. Current deep learning systems are notoriously bad at reasoning and abstraction, which is why they need huge amounts of data to learn simple tasks.

The question is, how do we go beyond feed-forward computation and system 1? How do we make reasoning compatible with gradient-based learning? How do we make reasoning differentiable? Thats the bottom line, LeCun said.

System 1 is the kind of learning tasks that dont require active thinking, such as navigating a known area or making small calculations. System 2 is the more active kind of thinking, which requires reasoning.Symbolic artificial intelligence, the classic approach to AI, has proven to be much better at reasoning and abstraction.

But LeCun doesnt suggest returning to symbolic AI or tohybrid artificial intelligence systems, as other scientists have suggested. His vision for the future of AI is much more in line with that of Yoshua Bengio, another deep learning pioneer, who introduced the concept ofsystem 2 deep learningat NeurIPS 2019 and further discussed it at AAAI 2020. LeCun, however, did admit that nobody has a completely good answer to which approach will enable deep learning systems to reason.

The third challenge is to create deep learning systems that can lean and plan complex action sequences, and decompose tasks into subtasks. Deep learning systems are good at providing end-to-end solutions to problems but very bad at breaking them down into specific interpretable and modifiable steps. There have been advances in creatinglearning-based AI systems that can decompose images, speech, and text. Capsule networks, invented by Geoffry Hinton, address some of these challenges.

But learning to reason about complex tasks is beyond todays AI. We have no idea how to do this, LeCun admits.

The idea behind self-supervised learning is to develop a deep learning system that can learn to fill in the blanks.

You show a system a piece of input, a text, a video, even an image, you suppress a piece of it, mask it, and you train a neural net or your favorite class or model to predict the piece thats missing. It could be the future of a video or the words missing in a text, LeCun says.

The closest we have to self-supervised learning systems are Transformers, an architecture that has proven very successful innatural language processing. Transformers dont require labeled data. They are trained on large corpora of unstructured text such as Wikipedia articles. And theyve proven to be much better than their predecessors at generating text, engaging in conversation, and answering questions. (But they are stillvery far from really understanding human language.)

Transformers have become very popular and are the underlying technology for nearly all state-of-the-art language models, including Googles BERT, Facebooks RoBERTa,OpenAIs GPT2, and GooglesMeena chatbot.

More recently, AI researchers have proven thattransformers can perform integration and solve differential equations, problems that require symbol manipulation. This might be a hint that the evolution of transformers might enable neural networks to move beyond pattern recognition and statistical approximation tasks.

So far, transformers have proven their worth in dealing with discreet data such as words and mathematical symbols. Its easy to train a system like this because there is some uncertainty about which word could be missing but we can represent this uncertainty with a giant vector of probabilities over the entire dictionary, and so its not a problem, LeCun says.

But the success of Transformers has not transferred to the domain of visual data. It turns out to be much more difficult to represent uncertainty and prediction in images and video than it is in text because its not discrete. We can produce distributions over all the words in the dictionary. We dont know how to represent distributions over all possible video frames, LeCun says.

For each video segment, there are countless possible futures. This makes it very hard for an AI system to predict a single outcome, say the next few frames in a video. The neural network ends up calculating the average of possible outcomes, which results in blurry output.

This is the main technical problem we have to solve if we want to apply self-supervised learning to a wide variety of modalities like video, LeCun says.

LeCuns favored method to approach supervised learning is what he calls latent variable energy-based models. The key idea is to introduce a latent variable Z which computes the compatibility between a variable X (the current frame in a video) and a prediction Y (the future of the video) and selects the outcome with the best compatibility score. In his speech, LeCun further elaborates on energy-based models and other approaches to self-supervised learning.

Energy-based models use a latent variable Z to compute the compatibility between a variable X and a prediction Y and select the outcome with the best compatibility score (image credit: Yann LeCun).

I think self-supervised learning is the future. This is whats going to allow to our AI systems, deep learning system to go to the next level, perhaps learn enough background knowledge about the world by observation, so that some sort of common sense may emerge, LeCun said in his speech at the AAAI Conference.

One of the key benefits of self-supervised learning is the immense gain in the amount of information outputted by the AI. In reinforcement learning, training the AI system is performed at scalar level; the model receives a single numerical value as reward or punishment for its actions. In supervised learning, the AI system predicts a category or a numerical value for each input.

In self-supervised learning, the output improves to a whole image or set of images. Its a lot more information. To learn the same amount of knowledge about the world, you will require fewer samples, LeCun says.

We must still figure out how the uncertainty problem works, but when the solution emerges, we will have unlocked a key component of the future of AI.

If artificial intelligence is a cake, self-supervised learning is the bulk of the cake, LeCun says. The next revolution in AI will not be supervised, nor purely reinforced.

This story is republished fromTechTalks, the blog that explores how technology is solving problems and creating new ones. Like them onFacebookhere and follow them down here:

Published April 5, 2020 05:00 UTC

See the original post here:
Self-supervised learning is the future of AI - The Next Web

Google has rolled out a new technology to improve audio quality in Duo calls when the service cant maintain a steady connection called WaveNetEQ. Its based on technology from Googles DeepMind division that aims to replace audio jitter with artificial noise that sounds just like human speech, generated using machine learning.

If youve ever made a call over the internet, chances are youve experienced audio jitter. It happens when packets of audio data sent as part of the call get lost along the way or otherwise arrive late or in the wrong order. Google says that 99 percent of Duo calls experience packet loss: 20 percent of these lose over 3 percent of their audio, and 10 percent lose over 8 percent. Thats a lot of audio to replace.

Every calling app has to deal with this packet loss somehow, but Google says that these packet loss concealment (PLC) processes can struggle to fill gaps of 60ms or more without sounding robotic or repetitive. WaveNetEQs solution is based on DeepMinds neural network technology, and it has been trained on data from over 100 speakers in 48 different languages.

Here are a few audio samples from Google comparing WaveNetEQ against NetEQ, a commonly used PLC technology. Heres how it sounds when its trying to replace 60ms of packet loss:

Heres a comparison when a call is experiencing packet loss of 120ms:

Theres a limit to how much audio the system can replace, though. Googles tech is designed to replace short sounds, rather than whole words. So after 120ms, it fades out and produces silence. Google says it evaluated the system to make sure it wasnt introducing any significant new sounds. Plus, all of the processing also needs to happen on-device since Google Duo calls are end-to-end encrypted by default. Once the calls real audio resumes, WaveNetEQ will seamlessly fade back to reality.

Its a neat little bit of technology that should make calls that much bit easier to understand when the internet fails them. The technology is already available for Duo calls made on Pixel 4 phones, thanks to the handsets December feature drop, and Google says its in the process of rolling it out to other unnamed handsets.

Link:
Google is using machine learning to improve the quality of Duo calls - The Verge