Machine learning and artificial intelligence are helping automate an ever-increasing array of tasks, with ever-increasing accuracy. They are supported by the growing volume of data used to feed them, and the growing sophistication in algorithms.

The flip side of more complex algorithms, however, is less interpretability. In many cases, the ability to retrace and explain outcomes reached by machine learning models (ML) is crucial, as:

"Trust models based on responsible authorities are being replaced by algorithmic trust models to ensure privacy and security of data, source of assets and identity of individuals and things. Algorithmic trust helps to ensure that organizations will not be exposed to the risk and costs of losing the trust of their customers, employees and partners. Emerging technologies tied to algorithmic trust include secure access service edge, differential privacy, authenticated provenance, bring your own identity, responsible AI and explainable AI."

The above quote is taken from Gartner's newly released 2020 Hype Cycle for Emerging Technologies. In it, explainable AI is placed at the peak of inflated expectations. In other words, we have reached peak hype for explainable AI. To put that into perspective, a recap may be useful.

As experts such as Gary Marcus point out, AI is probably not what you think it is. Many people today conflate AI with machine learning. While machine learning has made strides in recent years, it's not the only type of AI we have. Rule-based, symbolic AI has been around for years, and it has always been explainable.

Incidentally, that kind of AI, in the form of "Ontologies and Graphs" is also included in the same Gartner Hype Cycle, albeit in a different phase -- the trough of disillusionment. Incidentally, again, that's conflating.Ontologies are part of AI, while graphs, not necessarily.

That said: If you are interested in getting a better understanding of the state of the art in explainable AI machine learning, reading Christoph Molnar's book is a good place to start. Molnar is a data scientist and Ph.D. candidate in interpretable machine learning. Molnar has written the bookInterpretable Machine Learning: A Guide for Making Black Box Models Explainable, in which he elaborates on the issue and examines methods for achieving explainability.

Gartner's Hype Cycle for Emerging Technologies, 2020. Explainable AI, meaning interpretable machine learning, is at the peak of inflated expectations. Ontologies, a part of symbolic AI which is explainable, is in the trough of disillusionment

Recently, Molnar and a group of researchers attempted to addresses ML practitioners by raising awareness of pitfalls and pointing out solutions for correct model interpretation, as well as ML researchers by discussing open issues for further research. Their work was published as a research paper, titledPitfalls to Avoid when Interpreting Machine Learning Models, by the ICML 2020 Workshop XXAI: Extending Explainable AI Beyond Deep Models and Classifiers.

Similar to Molnar's book, the paper is thorough. Admittedly, however, it's also more involved. Yet, Molnar has striven to make it more approachable by means of visualization, using what he dubs "poorly drawn comics" to highlight each pitfall. As with Molnar's book on interpretable machine learning, we summarize findings here, while encouraging readers to dive in for themselves.

The paper mainly focuses on the pitfalls of global interpretation techniques when the full functional relationship underlying the data is to be analyzed. Discussion of "local" interpretation methods, where individual predictions are to be explained, is out of scope. For a reference on global vs. local interpretations, you can refer to Molnar's book as previously covered on ZDNet.

Authors note that ML models usually contain non-linear effects and higher-order interactions. As interpretations are based on simplifying assumptions, the associated conclusions are only valid if we have checked that the assumptions underlying our simplifications are not substantially violated.

In classical statistics this process is called "model diagnostics," and the research claims that a similar process is necessary for interpretable ML (IML) based techniques. The research identifies and describes pitfalls to avoid when interpreting ML models, reviews (partial) solutions for practitioners, and discusses open issues that require further research.

Under- or overfitting models will result in misleading interpretations regarding true feature effects and importance scores, as the model does not match the underlying data generating process well. Evaluation of training data should not be used for ML models due to the danger of overfitting. We have to resort to out-of-sample validation such as cross-validation procedures.

Formally, IML methods are designed to interpret the model instead of drawing inferences about the data generating process. In practice, however, the latter is the goal of the analysis, not the former. If a model approximates the data generating process well enough, its interpretation should reveal insights into the underlying process. Interpretations can only be as good as their underlying models. It is crucial to properly evaluate models using training and test splits -- ideally using a resampling scheme.

Flexible models should be part of the model selection process so that the true data-generating function is more likely to be discovered. This is important, as the Bayes error for most practical situations is unknown, and we cannot make absolute statements about whether a model already fits the data optimally.

Using opaque, complex ML models when an interpretable model would have been sufficient (i.e., having similar performance) is considered a common mistake. Starting with simple, interpretable models and gradually increasing complexity in a controlled, step-wise manner, where predictive performance is carefully measured and compared is recommended.

Measures of model complexity allow us to quantify the trade-off between complexity and performance and to automatically optimize for multiple objectives beyond performance. Some steps toward quantifying model complexity have been made. However, further research is required as there is no single perfect definition of interpretability but rather multiple, depending on the context.

This pitfall is further analyzed in three sub-categories: Interpretation with extrapolation, confusing correlation with dependence, and misunderstanding conditional interpretation.

Interpretation with Extrapolation refers to producing artificial data points that are used for model predictions with perturbations. These are aggregated to produce global interpretations. But if features are dependent, perturbation approaches produce unrealistic data points. In addition, even if features are independent, using an equidistant grid can produce unrealistic values for the feature of interest. Both issues can result in misleading interpretations.

Before applying interpretation methods, practitioners should check for dependencies between features in the data (e.g., via descriptive statistics or measures of dependence). When it is unavoidable to include dependent features in the model, which is usually the case in ML scenarios, additional information regarding the strength and shape of the dependence structure should be provided.

Confusing correlation with dependence is a typical error. The Pearson correlation coefficient (PCC) is a measure used to track dependency among ML features. But features with PCC close to zero can still be dependent and cause misleading model interpretations. While independence between two features implies that the PCC is zero, the converse is generally false.

Any type of dependence between features can have a strong impact on the interpretation of the results of IML methods. Thus, knowledge about (possibly non-linear) dependencies between features is crucial. Low-dimensional data can be visualized to detect dependence. For high-dimensional data, several other measures of dependence in addition to PCC can be used.

Misunderstanding conditional interpretation. Conditional variants to estimate feature effects and importance scores require a different interpretation. While conditional variants for feature effects avoid model extrapolations, these methods answer a different question. Interpretation methods that perturb features independently of others also yield an unconditional interpretation.

Conditional variants do not replace values independently of other features, but in such a way that they conform to the conditional distribution. This changes the interpretation as the effects of all dependent features become entangled. The safest option would be to remove dependent features, but this is usually infeasible in practice.

When features are highly dependent and conditional effects and importance scores are used, the practitioner has to be aware of the distinct interpretation. Currently, no approach allows us to simultaneously avoid model extrapolations and to allow a conditional interpretation of effects and importance scores for dependent features.

Global interpretation methods can produce misleading interpretations when features interact. Many interpretation methods cannot separate interactions from main effects. Most methods that identify and visualize interactions are not able to identify higher-order interactions and interactions of dependent features.

There are some methods to deal with this, but further research is still warranted. Furthermore, solutions lack in automatic detection and ranking of all interactions of a model as well as specifying the type of modeled interaction.

Due to the variance in the estimation process, interpretations of ML models can become misleading. When sampling techniques are used to approximate expected values, estimates vary, depending on the data used for the estimation. Furthermore, the obtained ML model is also a random variable, as it is generated on randomly sampled data and the inducing algorithm might contain stochastic components as well.

Hence, themodel variance has to be taken into account. The true effect of a feature may be flat, but purely by chance, especially on smaller data, an effect might algorithmically be detected. This effect could cancel out once averaged over multiple model fits. The researchers note the uncertainty in feature effect methods has not been studied in detail.

It's a steep fall to the peak of inflated expectations to the trough of disillusionment. Getting things done for interpretable machine learning takes expertise and concerted effort.

Simultaneously testing the importance of multiple features will result in false-positive interpretations if the multiple comparisons problem (MCP) is ignored. MCP is well known in significance tests for linear models and similarly exists in testing for feature importance in ML.

For example, when simultaneously testing the importance of 50 features, even if all features are unimportant, the probability of observing that at least one feature is significantly important is 0.923. Multiple comparisons will even be more problematic, the higher dimensional a dataset is. Since MCP is well known in statistics, the authors refer practitioners to existing overviews and discussions of alternative adjustment methods.

Practitioners are often interested in causal insights into the underlying data-generating mechanisms, which IML methods, in general, do not provide. Common causal questions include the identification of causes and effects, predicting the effects of interventions, and answering counterfactual questions. In the search for answers, researchers can be tempted to interpret the result of IML methods from a causal perspective.

However, a causal interpretation of predictive models is often not possible. Standard supervised ML models are not designed to model causal relationships but to merely exploit associations. A model may, therefore, rely on the causes and effects of the target variable as well as on variables that help to reconstruct unobserved influences.

Consequently, the question of whether a variable is relevant to a predictive model does not directly indicate whether a variable is a cause, an effect, or does not stand in any causal relation to the target variable.

As the researchers note, the challenge of causal discovery and inference remains an open key issue in the field of machine learning. Careful research is required to make explicit under which assumptions what insight about the underlying data generating mechanism can be gained by interpreting a machine learning model

Molnar et. al. offer an involved review of the pitfalls of global model-agnostic interpretation techniques for ML. Although as they note their list is far from complete, they cover common ones that pose a particularly high risk.

They aim to encourage a more cautious approach when interpreting ML models in practice, to point practitioners to already (partially) available solutions, and to stimulate further research.

Contrasting this highly involved and detailed groundwork to high-level hype and trends on explainable AI may be instructive.

The rest is here:

Explainable AI: From the peak of inflated expectations to the pitfalls of interpreting machine learning models - ZDNet

Automation offers substantive benefits as companies look for ways to manage evolving workforces and workplace expectations. More than half of U.S. businesses now plan to increase their automation investment to help increase their agility and improve their ability to handle changing conditions quickly, according to Robotics and Automation News.

Businesses also need to be able to solve problems at scale, something that organizations are increasingly turning to machine learning to do. By creating algorithms that learn over time, its possible for companies to streamline decision-making with data-driven predictions. But creating the models can be complex and time-consuming, putting an added strain on businesses that may be low on resources.

Automated machine learning combines these two technologies to tap the best of both worlds, allowing companies to gain actionable insights while reducing total complexity. Once implemented, AutoML can help businesses gather and analyze data, respond to it quickly and better manage resources.

WATCH: Find out how organizations can empower digital transformation and secure remote work.

AutoML goes a step further than classic machine learning, says Earnest Collins, managing member of Regulatory Compliance and Examination Consultants and a member of the ISACA Emerging Technologies Advisory Group.

AutoML goes beyond creating machine learning architecture models, says Collins. It can automate many aspects of machine learning workflow, which include data preprocessing, feature engineering, model selection, architecture search and model deployment.

AutoML deployments can also be categorized by the format of training data used. Collins points to examples such as independent, identically distributed (IID) tabular data, raw text or image data, and notes that some AutoML solutions can handle multiple data types and algorithms.

There is no single algorithm that performs best on all data sets, he says.

Leveraging AutoML solutions offers multiple benefits that go beyond traditional machine learning or automation. The first is speed, according to Collins.

AutoML allows data scientists to build a machine learning model with a high degree of automation more quickly and conduct hyperparameter search over different types of algorithms, which can otherwise be time-consuming and repetitive, he says. By automating key processes from raw data set capture to eventual analysis and learningteams can reduce the amount of time required to create functional models.

Another benefit is scalability. While machine learning models cant compete with the in-depth nature of human cognition, evolving technology makes it possible to create effective analogs of specific human learning processes. Introducing automation, meanwhile, helps apply this process at scale in turn, enabling data scientists, engineers and DevOps teams to focus on business problems instead of iterative tasks, Collins says.

A third major benefit is simplicity, according to Collins. AutoML is a tool that assists in automating the process of applying machine learning to real-world problems, he says.

By reducing the complexity that comes with building, testing and deploying entirely new ML frameworks, AutoML streamlines the processes required to solve line-of-business challenges.

For machine learning solutions to deliver business value, ML models must be optimized based on current conditions and desired outputs. Doing so requires the use of hyperparameters, which Collins defines as adjustable parameters that govern the training of ML models.

Optimal ML model performance depends on the hyperparameter configuration value selection; this can be a time-consuming, manual process, which is where AutoML can come into play, Collins adds.

By using AutoML platforms to automate key hyperparameter selection and balancing including learning rate, batch size and drop rate its possible to reduce the amount of time and effort required to get ML algorithms up and running.

While AutoML isnt new, evolution across machine learning and artificial intelligence markets is now driving a second generation of automated machine learning platforms, according to RTInsights. The first wave of AutoML focused on building and validating models, but the second iterations include key features such as data preparation and feature engineering to accelerate data science efforts.

But this market remains both fragmented and complex, according to Forbes, because of a lack of established standards and expectations in the data science and machine learning (DSML) industry. Businesses can go with an established provider, such as Microsoft Azure Databricks, or they can opt for more up-and-coming solutions such as Google Cloud AutoML.

There are more tools around the corner. According to Synced, Google researchers are now developing AutoML-Zero, which is capable of searching for applicable ML algorithms within a defined space to reduce the need to create them from scratch. The search giant is also applying its AutoML to unique use cases; for example, the companys new Fabricius tool which leverages Googles AutoML vision toolset is designed to decode ancient Egyptian hieroglyphics.

Technological advancements combined with shifting staff priorities are somewhat driving robotic replacements. According to Time, companies are replacing humans wherever possible to reduce risk and improve operational output. But that wont necessarily apply to data scientists as AutoML rises, according to Collins.

The skills of professional, well-trained data scientists will be essential to interpreting data and making recommendations for how information should be used, he says. AutoML will be a key tool for improving their productivity, and the citizen data scientist, with no training in the field, would not be able to do machine learning without AutoML.

In other words, while AutoML platforms provide business benefits, recognizing the full extent of automated advantages will always require human expertise.

Go here to read the rest:

What is AutoML and Why Should Your Business Consider It - BizTech Magazine

Machine Learning Courses and Market research is an intelligence report with meticulous efforts undertaken to study the right and valuable information. It offers an overview of the market including its definition, applications, key drivers, key market players, key segments, and manufacturing technology. Moreover, the report is a detailed study exhibiting current market trends with an overview of future market study.

Get Sample Copy athttps://www.orianresearch.com/request-sample/760330

Regions and Countries Level Analysis

Regional analysis is another highly comprehensive part of the research and analysis study of the global Machine Learning Courses and market presented in the report. This section sheds light on the sales growth of different regional and country-level Machine Learning Courses and markets. For the historical and forecast period 2015 to 2026, it provides detailed and accurate country-wise volume analysis and region-wise market size analysis of the global Machine Learning Courses and market.

The key players covered in this study

EdX

Ivy Professional School

NobleProg

Udacity

Edvancer

Udemy

Simplilearn

Jigsaw Academy

BitBootCamp

Metis

DataCamp

Inquire more or share questions if any before the purchase on this report @https://www.orianresearch.com/enquiry-before-buying/760330

No of Pages: 121

Market segmentation

Machine Learning Courses and market is split by Type and by Application. For the period 2015-2026, the growth among segments provide accurate calculations and forecasts for sales by Type and by Application in terms of volume and value. This analysis can help you expand your business by targeting qualified niche markets.

Market segment by Type, the product can be split into Rote Learning Learning From Instruction Learning By Deduction Learning By Analogy Explanation-Based Learning Learning From Induction

Market segment by Application, split into Data Mining Computer Vision Natural Language Processing Biometrics Recognition Search Engines Medical Diagnostics Detection Of Credit Card Fraud Securities Market Analysis DNA Sequencing

What our report offers:

Market share assessments for the regional and country level segments

Market share analysis of the top industry players

Strategic recommendations for the new entrants

Market forecasts for a minimum of 9 years of all the mentioned segments, sub segments and the regional markets

Market Trends (Drivers, Constraints, Opportunities, Threats, Challenges, Investment Opportunities, and recommendations)

Strategic recommendations in key business segments based on the market estimations

Competitive landscaping mapping the key common trends

Company profiling with detailed strategies, financials, and recent developments

Supply chain trends mapping the latest technological advancements

Global Machine Learning Courses and Market report has been compiled through extensive primary research (through analytical research, market survey and observations) and secondary research. The Machine Learning Courses and Market report also features a complete qualitative and quantitative assessment by analyzing data gathered from industry analysts, key vendors, business news, row material supplier, regional clients, company journals, and market participants across key points in the industrys value chain.

Order a Copy of this Report @https://www.orianresearch.com/checkout/760330

Table of Contents

1 Industry Overview of Machine Learning Courses and

2 Industry Chain Analysis of Machine Learning Courses and

3 Manufacturing Technology of Machine Learning Courses and

4 Major Manufacturers Analysis of Machine Learning Courses and

5 Global Productions, Revenue and Price Analysis of Machine Learning Courses and by Regions, Manufacturers, Types and Applications

6 Global and Major Regions Capacity, Production, Revenue and Growth Rate of Machine Learning Courses and 2014-2019

7 Consumption Volumes, Consumption Value, Import, Export and Sale Price Analysis of Machine Learning Courses and by Regions

8 Gross and Gross Margin Analysis of Machine Learning Courses and

9 Marketing Traders or Distributor Analysis of Machine Learning Courses and

10 Global and Chinese Economic Impacts on Machine Learning Courses and Industry

11 Development Trend Analysis of Machine Learning Courses and

12 Contact information of Machine Learning Courses and

13 New Project Investment Feasibility Analysis of Machine Learning Courses and

14 Conclusion of the Global Machine Learning Courses and Industry 2019 Market Research Report

Customization Service of the Report:-

Orian Research provides customization of Reports as your need. This Report can be personalized to meet all your requirements. If you have any question get in touch with our sales team, who will guarantee you to get a Report that suits your necessities.

About Us

Orian Research is one of the most comprehensive collections of market intelligence reports on The World Wide Web. Our reports repository boasts of over 500000+ industry and country research reports from over 100 top publishers. We continuously update our repository so as to provide our clients easy access to the worlds most complete and current database of expert insights on global industries, companies, and products. We also specialize in custom research in situations where our syndicate research offerings do not meet the specific requirements of our esteemed clients.

Contact Us

Ruwin Mendez

Vice President Global Sales & Partner Relations

Orian Research Consultants

US: +1 (415) 830-3727 | UK: +44 020 8144-71-27

Email:[emailprotected]

Read more from the original source:

Machine Learning Courses and Market 2020: Growing Tends in Global Regions with COVID-19 Pandemic Analysis, Growth Size, Share, Types, Applications,...

The World Economic Forum and associate partner Appen are wrestling with the thorny issue of how to create artificial intelligence with a sense of ethics.

Their main area of focus is to design standards and best practices for responsible training data used in building machine learning and AI applications. It has already been a long process and continues.

A solid training data platform and management strategy is often the most critical component of launching a successful, responsible machine learning-powered product into production, said Mark Brayan, CEO of Appen in a statement. Appen has been providing training data to companies building AI for more than 20 years. In 2019, Appen created its own Crowd Code of Ethics.

The electronics industry remains in flux as constant innovation fuels market trends. FierceElectronics subscribers rely on our suite of newsletters as their must-read source for the latest news, developments and predictions impacting their world. Sign up today to get electronics news and updates delivered to your inbox and read on the go.

Ethical, diverse training data is essential to building a responsible AI system, Brayan added.

Kay Firth-Butterfield, head of AI and machine learning at WEF, said the industry needs guidelines for acquiring and using responsible training data. Companies should address questions around user permissions, privacy, security, bias, safety and how people are compensated for their work in the AI supply chain, she said.

Every business needs a plan to understand AI and deploy AI safely and ethically, she added in a video overview of Forums AI agenda. The purpose is to think about what are the big issues in AI that really require something be done in the governance area so that AI can flourish.

Were very much advocating asoft law approach, thinking about standards and guidelines rather than looking to regulation, she said.

The Forum has issued a number of white papers dating to 2018 on ethics and related topics, with a white paper on responsible limits on facial recognition issued in March.

RELATED: Researchers deploy AI to detect bias in AI and humans

In January, the Forum published its AI toolkit for boards of directors with 12 modules for the impacts and potential of AI in company strategy and is currently building a toolkit for transferring those insights to CEOs and other C-suite executives.

Another focus area is on human-centered AI for human resources to create a toolkit for HR professionals that will help promote ethical human-centered use of AI. Various HR tools have been developed in recent years that rely on AI to hire and retain talent and the Forum notes that concerns have been raised about AI algorithms encoding bias and discrimination. Errors in the adoption of AI-based products can also undermine employee trust, leading to lower productivity and job satisfaction, the Forum added.

Firth-Butterfield will be a keynote speaker at Appen annual Train AI conference on October 14.

RELATED: Tech firms grapple with diversity after George Floyd protests

Continue reading here:

Focusing on ethical AI in business and government - FierceElectronics

HONG KONG, CHINA -Media OutReach- 27 August 2020 -It's been called the holy grail of finance. Is it possible to harness the promise of artificial intelligence to make money trading stocks? Many have tried with varying degrees of success. For example, BlackRock, the world's largest money manager, has said its Artificial Intelligence (AI) algorithms have consistently beaten portfolios managed by human stock pickers. However, a recent research study by The Chinese University of Hong Kong (CUHK) reveals that the effectiveness of machine learning methods may require a second look.

The study, titled "Machine Learning versus Economic Restrictions: Evidence from Stock Return Predictability", analysed a large sample of U.S. stocks between 1987 and 2017. Using three well-established deep-learning methods, researchers were able to generate a monthly value-weighted risk-adjusted return of as much as 0.75 percent to 1.87 percent, reflecting the success of machine learning in generating a superior payoff. However, the researchers found that this performance would attenuate if the machine learning algorithms were limited to working with stocks that were relatively easy and cheap to trade.

"We find that the return predictability of deep learning methods weakens considerably in the presence of standard economic restrictions in empirical finance, such as excluding microcaps or distressed firms," says Si Cheng, Assistant Professor at CUHK Business School's Department of Finance and one of the study's authors.

Disappearing Returns

Prof. Cheng, along with her collaborators Prof. Doron Avramov at IDC Herzliya and Lior Metzker, a research student at Hebrew University of Jerusalem, found the portfolio payoff declined by 62 percent when excluding microcaps -- stocks which can be difficult to trade because of their small market capitalisations, 68 percent lower when excluding non-rated firms -- stocks which do not receive Standard & Poor's long-term issuer credit rating, and 80 percent lower excluding distressed firms around credit rating downgrades.

According to the study, machine learning-based trading strategies are more profitable during periods when arbitrage becomes more difficult, such as when there is high investor sentiment, high market volatility, and low market liquidity.

One caveat of the machine-learning based strategies highlighted by the study is high transaction costs. "Machine learning methods require high turnover and taking extreme stock positions. An average investor would struggle to achieve meaningful alpha after taking transaction costs into account," she says, adding, however, that this finding did not imply that machine learning-based strategies are unprofitable for all traders.

"Instead, we show that machine learning methods studied here would struggle to achieve statistically and economically meaningful risk-adjusted performance in the presence of reasonable transaction costs. Investors thus should adjust their expectations of the potential net-of-fee performance," says Prof. Cheng.

The Future of Machine Learning

"However, our findings should not be taken as evidence against applying machine learning techniques in quantitative investing," Prof. Cheng explains. "On the contrary, machine learning-based trading strategies hold considerable promise for asset management." For instance, they have the capability to process and combine multiple weak stock trading signals into meaningful information that could form the basis for a coherent trading strategy.

Machine learning-based strategies display less downside risk and continue to generate positive payoff during crisis periods. The study found that during several major market downturns, such as the 1987 market crash, the Russian default, the burst of the tech bubble, and the recent financial crisis, the best machine-learning investment method generated a monthly value-weighted return of 3.56 percent, excluding microcaps, while the market return came in at a negative 6.91 percent during the same period.

Prof. Cheng says that the profitability of trading strategies based on identifying individual stock market anomalies -- stocks whose behaviour run counter to conventional capital market pricing theory predictions -- is primarily driven by short positions and is disappearing in recent years. However, machine-learning based strategies are more profitable in long positions and remain viable in the post-2001 period.

"This could be particularly valuable for real-time trading, risk management, and long-only institutions. In addition, machine learning methods are more likely to specialise in stock picking than industry rotation," Prof. Cheng adds, referring to strategy which seeks to capitalise on the next stage of economic cycles by moving funds from one industry to the next.

The study is the first to provide large-scale evidence on the economic importance of machine learning methods, she adds.

"The collective evidence shows that most machine learning techniques face the usual challenge of cross-sectional return predictability, and the anomalous return patterns are concentrated in difficult-to-arbitrage stocks and during episodes of high limits to arbitrage," Prof. Cheng says. "Therefore, even though machine learning offers unprecedented opportunities to shape our understanding of asset pricing formulations, it is important to consider the common economic restrictions in assessing the success of newly developed methods, and confirm the external validity of machine learning models before applying them to different settings."

Reference:

Avramov, Doron and Cheng, Si and Metzker, Lior, Machine Learning versus Economic Restrictions: Evidence from Stock Return Predictability (April 5, 2020). Available at SSRN: https://ssrn.com/abstract=3450322 or http://dx.doi.org/10.2139/ssrn.3450322

This article was first published in the China Business Knowledge (CBK) website by CUHK Business School: https://bit.ly/3fX2ydr.

CUHK Business School comprises two schools -- Accountancy and Hotel and Tourism Management -- and four departments -- Decision Sciences and Managerial Economics, Finance, Management and Marketing. Established in Hong Kong in 1963, it is the first business school to offer BBA, MBA and Executive MBA programmes in the region. Today, the School offers 11 undergraduate programmes and 20 graduate programmes including MBA, EMBA, Master, MSc, MPhil and Ph.D.

In the Financial Times Global MBA Ranking 2020, CUHK MBA is ranked 50th. In FT's 2019 EMBA ranking, CUHK EMBA is ranked 24th in the world. CUHK Business School has the largest number of business alumni (37,000+) among universities/business schools in Hong Kong -- many of whom are key business leaders. The School currently has about 4,800 undergraduate and postgraduate students and Professor Lin Zhou is the Dean of CUHK Business School.

More information is available at http://www.bschool.cuhk.edu.hk or by connecting with CUHK Business School on:

Facebook: http://www.facebook.com/cuhkbschool

Instagram: http://www.instagram.com/cuhkbusinessschool

LinkedIn: http://www.linkedin.com/school/cuhkbusinessschool

WeChat: CUHKBusinessSchool

Go here to see the original:

CUHK Business School Research Looks at the Limitations of Using Artificial Intelligence to Pick Stocks - Taiwan News

According to the approach embraced by McRO and BASCOM, while machine learning algorithms bringing a slight improvement can pass the eligibility test, algorithms paving the way for a whole new technology can be excluded from the benefits of patent protection simply because there are no alternatives.

In the past decade or so, humanity has gone through drastic changes as Artificial intelligence (AI) technologies such as recommendation systems and voice assistants have seeped into every facet of our lives. Whereas the number of patent applications for AI inventions skyrocketed, almost a third of these applications are rejected by the U.S. Patent and Trademark Office (USPTO) and the majority of these rejections are due to the claimed invention being ineligible subject matter.

The inventive concept may be attributed to different components of machine learning technologies, such as using a new algorithm, feeding more data, or using a new hardware component. However, this article will exclusively focus on the inventions achieved by Machine Learning (M.L.) algorithms and the effect of the preemption test adopted by U.S. courts on the patent-eligibility of such algorithms.

Since the Alice decision, the U.S. courts have adopted different views related to the role of the preemption test in eligibility analysis. While some courts have ruled that lack of preemption of abstract ideas does not make an invention patent-eligible [Ariosa Diagnostics Inc. v. Sequenom Inc.], others have not referred to it at all in their patent eligibility analysis. [Enfish LLC v. Microsoft Corp., 822 F.3d 1327]

Contrary to those examples, recent cases from Federal Courts have used the preemption test as the primary guidance to decide patent eligibility.

In McRO, the Federal Circuit ruled that the algorithms in the patent application prevent pre-emption of all processes for achieving automated lip-synchronization of 3-D characters. The court based this conclusion on the evidence of availability of an alternative set of rules to achieve the automation process other than the patented method. It held that the patent was directed to a specific structure to automate the synchronization and did not preempt the use of all of the rules for this method given that different sets of rules to achieve the same automated synchronization could be implemented by others.

Similarly, The Court in BASCOM ruled that the claims were patent eligible because they recited a specific, discrete implementation of the abstract idea of filtering contentand they do not preempt all possible ways to implement the image-filtering technology.

The analysis of the McRO and BASCOM cases reveals two important principles for the preemption analysis:

Machine learning can be defined as a mechanism which searches for patterns and which feeds intelligence into a machine so that it can learn from its own experience without explicit programming. Although the common belief is that data is the most important component in machine learning technologies, machine learning algorithms are equally important to proper functioning of these technologies and their importance cannot be understated.

Therefore, inventive concepts enabled by new algorithms can be vital to the effective functioning of machine learning systemsenabling new capabilities, making systems faster or more energy efficient are examples of this. These inventions are likely to be the subject of patent applications. However, the preemption test adopted by courts in the above-mentioned cases may lead to certain types of machine learning algorithms being held ineligible subject matter. Below are some possible scenarios.

The first situation relates to new capabilities enabled by M.L. algorithms. When a new machine learning algorithm adds a new capability or enables the implementation of a process, such as image recognition, for the first time, preemption concerns will likely arise. If the patented algorithm is indispensable for the implementation of that technology, it may be held ineligible based on the McRO case. This is because there are no other alternative means to use this technology and others would be prevented from using this basic tool for further development.

For example, a M.L. algorithm which enabled the lane detection capability in driverless cars may be a standard/must-use algorithm in the implementation of driverless cars that the court may deem patent ineligible for having preemptive effects. This algorithm clearly equips the computer vision technology with a new capability, namely, the capability to detect boundaries of road lanes. Implementation of this new feature on driverless cars would not pass the Alice test because a car is a generic tool, like a computer, and even limiting it to a specific application may not be sufficient because it will preempt all uses in this field.

Should the guidance of McRO and BASCOM be followed, algorithms that add new capabilities and features may be excluded from patent protection simply because there are no other available alternatives to these algorithms to implement the new capabilities. These algorithms use may be so indispensable for the implementation of that technology that they are deemed to create preemptive effects.

Secondly, M.L. algorithms which are revolutionary may also face eligibility challenges.

The history of how deep neural networks have developed will be explained to demonstrate how highly-innovative algorithms may be stripped of patent protection because of the preemption test embraced by McRO and subsequent case law.

Deep Belief Networks (DBNs) is a type of Artificial Neural Networks (ANNs). The ANNs were trained with a back-propagation algorithm, which adjusts weights by propagating the outputerror backwardsthrough the network However, the problem with the ANNs was that as the depth was increased by adding more layers, the error vanished to zero and this severely affected the overall performance, resulting in less accuracy.

From the early 2000s, there has been a resurgence in the field of ANNs owing to two major developments: increased processing power and more efficient training algorithms which made trainingdeep architecturesfeasible. The ground-breaking algorithm which enabled the further development of ANNs in general and DBNs in particular was Hintons greedy training algorithm.

Thanks to this new algorithm, DBNs has been applicable to solve a variety of problems that were the roadblock before the use of new technologies, such as image processing,natural language processing, automatic speech recognition, andfeature extractionand reduction.

As can be seen, the Hiltons fast learning algorithm revolutionized the field of machine learning because it made the learning easier and, as a result, technologies such as image processing and speech recognition have gone mainstream.

If patented and challenged at court, Hiltons algorithm would likely be invalidated considering previous case law. In McRO, the court reasoned that the algorithm at issue should not be invalidated because the use of a set of rules within the algorithm is not a must and other methods can be developed and used. Hiltons algorithm will inevitably preempt some AI developers from engaging with further development of DBNs technologies because this algorithm is a base algorithm, which made the DBNs plausible to implement so that it may be considered as a must. Hiltons algorithm enabled the implementation of image recognition technologies and some may argue based on McRO and Enfish that Hiltons algorithm patent would be preempting because it is impossible to implement image recognition technologies without this algorithm.

Even if an algorithm is a must-use for a technology, there is no reason to exclude it from patent protection. Patent law inevitably forecloses certain areas from further development by granting exclusive rights through patents. All patents foreclose competitors to some extent as a natural consequence of exclusive rights.

As stated in the Mayo judgment, exclusive rights provided by patents can impede the flow of information that might permit, indeed spur, invention, by, for example, raising the price of using the patented ideas once created, requiring potential users to conduct costly and time-consuming searches of existing patents and pending patent applications, and requiring the negotiation of complex licensing arrangements.

The exclusive right granted by a patents is only one side of the implicit agreement between the society and the inventor. In exchange for the benefit of the exclusivity, inventors are required to disclose their invention to the public so this knowledge becomes public, available for use in further research and for making new inventions building upon the previous one.

If inventors turn to trade secrets to protect their inventions due to the hostile approach of patent law to algorithmic inventions, the knowledge base in this field will narrow, making it harder to build upon previous technology. This may lead to the slow-down and even possible death of innovation in this industry.

The fact that an algorithm is a must-use, should not lead to the conclusion that it cannot be patented. Patent rights may even be granted for processes which have primary and even sole utility in research. Literally, a microscope is a basic tool for scientific work, but surely no one would assert that a new type of microscope lay beyond the scope of the patent system. Even if such a microscope is used widely and it is indispensable, it can still be given patent protection.

According to the approach embraced by McRO and BASCOM, while M.L. algorithms bringing a slight improvement, such as a higher accuracy and higher speed, can pass the eligibility test, algorithms paving the way for a whole new technology can be excluded from the benefits of patent protection simply because there are no alternatives to implement that revolutionary technology.

Considering that the goal of most AI inventions is to equip computers with new capabilities or bring qualitative improvements to abilities such as to see or to hear or even to make informed judgments without being fed complete information, most AI inventions would have the higher likelihood of being held patent ineligible. Applying this preemption test to M.L. algorithms would put such M.L. algorithms outside of patent protection.

Thus, a M.L. algorithm which increases accuracy by 1% may be eligible, while a ground-breaking M.L. algorithm which is a must-use because it covers all uses in that field may be excluded from patent protection. This would result in rewarding slight improvements with a patent but disregarding highly innovative and ground-breaking M.L. algorithms. Such a consequence is undesirable for the patent system.

This also may result in deterring the AI industry from bringing innovation in fundamental areas. As an undesired consequence, innovation efforts may shift to small improvements instead of innovations solving more complex problems.

Image Source: Author: nils.ackermann.gmail.com Image ID:102390038

Link:

Effects of the Alice Preemption Test on Machine Learning Algorithms - IPWatchdog.com

A new project called Keen is launching today from Googles in-house incubator for new ideas, Area 120, to help users track their interests. The app is like a modern rethinking of the Google Alerts service, which allows users to monitor the web for specific content. Except instead of sending emails about new Google Search results, Keen leverages a combination of machine learning techniques and human collaboration to help users curate content around a topic.

Each individual area of interest is called a keen a word often used to reference someone with an intellectual quickness.

The idea for the project came about after co-founder C.J. Adams realized he was spending too much time on his phone mindlessly browsing feeds and images to fill his downtime. He realized that time could be better spent learning more about a topic he was interested in perhaps something he always wanted to research more or a skill he wanted to learn.

To explore this idea, he and four colleagues at Google worked in collaboration with the companys People and AI Research (PAIR) team, which focuses on human-centered machine learning, to create what has now become Keen.

To use Keen, which is available both on the web and on Android, you first sign in with your Google account and enter in a topic you want to research. This could be something like learning to bake bread, bird watching or learning about typography, suggests Adams in an announcement about the new project.

Keen may suggest additional topics related to your interest. For example, type in dog training and Keen could suggest dog training classes, dog training books, dog training tricks, dog training videos and so on. Click on the suggestions you want to track and your keen is created.

When you return to the keen, youll find a pinboard of images linking to web content that matches your interests. In the dog training example, Keen found articles and YouTube videos, blog posts featuring curated lists of resources, an Amazon link to dog training treats and more.

For every collection, the service uses Google Search and machine learning to help discover more content related to the given interest. The more you add to a keen and organize it, the better these recommendations become.

Its like an automated version of Pinterest, in fact.

Once a keen is created, you can then optionally add to the collection, remove items you dont want and share the Keen with others to allow them to also add content. The resulting collection can be either public or private. Keen can also email you alerts when new content is available.

Google, to some extent, already uses similar techniques to power its news feed in the Google app. The feed, in that case, uses a combination of items from your Google Search history and topics you explicitly follow to find news and information it can deliver to you directly on the Google apps home screen. Keen, however, isnt tapping into your search history. Its only pulling content based on interests you directly input.

And unlike the news feed, a keen isnt necessarily focused only on recent items. Any sort of informative, helpful information about the topic can be returned. This can include relevant websites, events, videos and even products.

But as a Google project and one that asks you to authenticate with your Google login the data it collects is shared with Google. Keen, like anything else at Google, is governed by the companys privacy policy.

Though Keen today is a small project inside a big company, it represents another step toward the continued personalization of the web. Tech companies long since realized that connecting users with more of the content that interests them increases their engagement, session length, retention and their positive sentiment for the service in question.

But personalization, unchecked, limits users exposure to new information or dissenting opinions. It narrows a persons worldview. It creates filter bubbles and echo chambers. Algorithmic-based recommendations can send users searching for fringe content further down dangerous rabbit holes, even radicalizing them over time. And in extreme cases, radicalized individuals become terrorists.

Keen would be a better idea if it were pairing machine-learning with topical experts. But it doesnt add a layer of human expertise on top of its tech, beyond those friends and family you specifically invite to collaborate, if you even choose to. That leaves the system wanting for better human editorial curation, and perhaps the need for a narrower focus to start.

Read the original here:

Googles latest experiment is Keen, an automated, machine-learning based version of Pinterest - TechCrunch

According to PwC, AIs potential global economic impact will reach USD 15.7 trillion by 2030. However, the enterprises who look to deploy AI are often hampered by the lack of time, trust and talent. Especially, with the highly regulated sectors such as healthcare and finance, convincing the customers to imbibe AI methodologies is an uphill task.

Of late, the AI community has seen a sporadic shift in AI adoption with the advent of AutoML tools and introduction of customised hardware to cater to the needs of the algorithms. One of the most widely used AutoML tools in the industry is H2O Driverless AI. And, when it comes to hardware Intel has been consistently updating its tool stack to meet the high computational demands of the AI workflows.

Now H2O.ai and Intel, two companies who have been spearheading the democratisation of the AI movement, join hands to develop solutions that leverage software and hardware capabilities respectively.

AI and machine-learning workflows are complex and enterprises need more confidence in the validity of their AI models than a typical black-box environment can provide. The inexplicability and the complexity of feature engineering can be daunting to the non-experts. So far AutoML has proven to be the one stop solution to all these problems. These tools have alleviated the challenges by providing automated workflows, code ready deployable models and many more.

H2O.ai especially, has pioneered in the AutoML segment. They have developed an open source, distributed in-memory machine learning platform with linear scalability that includes a module called H2OAutoML, which can be used for automating the machine learning workflow, that includes automatic training and tuning of many models within a user-specified time-limit.

Whereas, H2O.ais flagship product Driverless AI can be used to fully automate some of the most challenging and productive tasks in applied data science such as feature engineering, model tuning, model ensembling and model deployment.

But, for these AI based tools to work seamlessly, they need the backing of hardware that is dedicated to handle the computational intensity of machine learning operations.

Intel has been at the forefront of digital revolution for over half a century. Today, Intel flaunts a wide range of technologies, including its Xeon Scalable processors, Optane Solid State Drives and optimized Intel software libraries that bring in a much needed mix of enhanced performance, AI inference, network functions, persistent memory bandwidth, and security.

Integrating H2O.ais software portfolio with hardware and software technologies from Intel has resulted in solutions that can handle almost all the woes of an AI enterprise from automated workflows to explainability to production ready code that can be deployed anywhere.

For example, H2O Driverless AI, an automatic machine-learning platform enables data science experts and beginners to streamline their AI tasks within minutes that usually take months. Today, more than 18,000 companies use open source H2O in mission-critical use cases for finance, insurance, healthcare, retail, telco, sales, and marketing.

The software capabilities of H2O.ai combined with hardware infrastructure of Intel, that includes 2nd Generation Xeon Scalable processors, Optane Solid State Drives and Ethernet Network Adapters, can empower enterprises to optimize performance and accelerate deployment.

Enterprises that are looking for increasing productivity while increasing the business value of to enjoy the competitive advantages of AI innovation no longer have to wait thanks to hardware backed AutoML solutions.

comments

Read more from the original source:

Deploying Machine Learning Has Never Been This Easy - Analytics India Magazine

Patrick Opie, founder of Scout9, and his dog, Orin. (Photo courtesy of Patrick Opie)

After adopting his first dog last year, Patrick Opie was struggling with figuring out what Orin, his mini Australian shepherd, needed and when.

The struggle went beyond coping with normal puppy stuff, like when a dog chews up a favorite pair of shoes or pees where hes not supposed to. Opie was buying products that were irrelevant or unfit for his dog and he was spending too much time researching what to get each month.

Those things add up, Opie said. Thats where I realized I really wished there was a product or something that could help navigate or work with you to help you find what you need to get going.

Opies new adventures in dog parenthood led him to create Scout9, a Seattle startup that offers an intuitive and economical way for new dog owners to prepare for each step of their dogs development through the use of an autonomous Personal Pocket Scout.

Its a timely venture considering reports that the COVID-19 pandemic has led to a national surge in pet adoptions and fostering. As the pet industry heads toward $100 billion in annual spending, pet tech and web-based services are right in the mix, especially in dog friendly Seattle.

Opie was frustrated by his own mess-ups when it came to buying the right food and the right type of kennel as well as milestones he missed including when to start socialization and training for Orin.

Think of it like if Im Batman and I just got a dog, Opie said. I would want to have an Alfred who can kind of help me figure out the baseline: These are the things you need to think about, these are the things that I suggest you should do.'

Opies Alfred-the-butler vision is instead an online platform that relies on machine learning technology to create a dynamic timeline for milestones in the dogs life. Its not breed specific, but is instead based on some parameters given to the tool, such as the dogs initial age and size. Scout works by scouring the internet for relevant information and learning along the way what the human user accepts and rejects.

Scout will surface food choices, for instance, and do the shopping if given permission, by searching for the best available deals. The user has the ability to set their budget, so that Scout avoids overspending and gets the most out of the money it is allotted. Purchases can be automated so food shows up on time and Scout will learn and grow as your pet does.

A user can also take Scouts recommendations and go find food or other items on Amazon or somewhere else.

Scout9 will make money a couple different ways, either by collecting a commission from retailers whose affiliated links show up in the tool, or by charging users a service fee on transactions that are made by Scout on the users behalf.

Using Orin as a test case for the first year, Opie said he went from spending $1,700 on supplies down to $1,100 using his tool, for a 35 percent savings.

Opie, who is working on the new company with two friends, was previously a consultant at Boston Consulting Group and he spent more than three years at Accenture. He also worked as a developer at DevHub, and in April teamed with DevHub co-founder Mark Michael to create a virtual Gumwall to raise money for restaurant workers during the early days of the health crisis.

His goal is for dogs to be the jumping off point for Scout9 and the Personal Pocket Scout, and he envisions it being applied beyond raising puppies to such scenarios as raising a baby or buying a new house.

It definitely is an idea that will be across all life transitions, Opie said. My team all loves dogs. Weve been through that experience. Its easier for us to execute on that vision.

See the rest here:

This startup could be a dog owners best friend as it uses machine learning to help guide key decisions - GeekWire

Protesters and police in Seattle's Capitol Hill neighborhood. (Getty Images)

When incidents of police brutality occur, typically departments enact police reforms and fire bad cops, but machine learning could potentially predict when a police officer may go over the line.

Rayid Ghani is a professor at Carnegie Mellon and joined Seattles Morning News to discuss using machine learning in police reform. Hes working on tech that could predict not only which cops might not be suited to be cops, but which cops might be best for a particular call.

AI and technology and machine learning, and all these buzzwords, theyre not able to to fix racism or bad policing, they are a small but important tool that we can use to help, Ghani said. I was looking at the systems called early intervention systems that a lot of large police departments have. Theyre supposed to raise alerts, raise flags when a police officer is at risk of doing something that they shouldnt be doing, like excessive use of force.

What level of privacy can we expect online?

What we found when looking at data from several police departments is that these existing systems were mostly ineffective, he added. If theyve done three things in the last three months that raised the flag, well thats great. But at the same time, its not an early intervention. Its a late intervention.

So they built a system that works to potentially identify high risk officers before an incident happens, but how exactly do you predict how somebody is going to behave?

We build a predictive system that would identify high risk officers We took everything we know about a police officer from their HR data, from their dispatch history, from who they arrested , their internal affairs, the complaints that are coming against them, the investigations that have happened, Ghani said.

Can the medical system and patients afford coronavirus-related costs?

What we found were some of the obvious predictors were what you think is their historical behavior. But some of the other non-obvious ones were things like repeated dispatches to suicide attempts or repeated dispatches to domestic abuse cases, especially involving kids. Those types of dispatches put officers at high risk for the near future.

While this might suggest that officers who regularly dealt with traumatic dispatches might be the ones who are higher risk, the data doesnt explain why, it just identifies possibilities.

It doesnt necessarily allow us to figure out the why, it allows us to narrow down which officers are high risk, Ghani said. Lets say a call comes in to dispatch and the nearest officer is two minutes away, but is high risk of one of these types of incidents. The next nearest officer is maybe four minutes away and is not high risk. If this dispatch is not time critical for the two minutes extra it would take, could you dispatch the second officer?

So if an officer has been sent to a multiple child abuse cases in a row, it makes more sense to assign somebody else the next time.

Thats right, Ghani said. Thats what that were finding is they become high risk It looks like its a stress indicator or a trauma indicator, and they might need a cool-off period, they might need counseling.

But in this case, the useful thing to think about also is that they havent done anything yet, he added. This is preventative, this is proactive. And so the intervention is not punitive. You dont fire them. You give them the tools that they need.

Listen to Seattles Morning News weekday mornings from 5 9 a.m. on KIRO Radio, 97.3 FM. Subscribe to thepodcast here.

Visit link:

How machine learning could reduce police incidents of excessive force - MyNorthwest.com