A coalition of AI researchers, data scientists, and sociologists has called on the academic world to stop publishing studies that claim to predict an individuals criminality using algorithms trained on data like facial scans and criminal statistics.

Such work is not only scientifically illiterate, says the Coalition for Critical Technology, but perpetuates a cycle of prejudice against Black people and people of color. Numerous studies show the justice system treats these groups more harshly than white people, so any software trained on this data simply amplifies and entrenches societal bias and racism.

Lets be clear: there is no way to develop a system that can predict or identify criminality that is not racially biased because the category of criminality itself is racially biased, write the group. Research of this nature and its accompanying claims to accuracy rest on the assumption that data regarding criminal arrest and conviction can serve as reliable, neutral indicators of underlying criminal activity. Yet these records are far from neutral.

An open letter written by the Coalition was drafted in response to news that Springer, the worlds largest publisher of academic books, planned to publish just such a study. The letter, which has now been signed by 1,700 experts, calls on Springer to rescind the paper and for other academic publishers to refrain from publishing similar work in the future.

At a time when the legitimacy of the carceral state, and policing in particular, is being challenged on fundamental grounds in the United States, there is high demand in law enforcement for research of this nature, write the group. The circulation of this work by a major publisher like Springer would represent a significant step towards the legitimation and application of repeatedly debunked, socially harmful research in the real world.

In the study in question, titled A Deep Neural Network Model to Predict Criminality Using Image Processing, researchers claimed to have created a facial recognition system that was capable of predicting whether someone is likely going to be a criminal ... with 80 percent accuracy and no racial bias, according to a now-deleted press release. The papers authors included Phd student and former NYPD police officer Jonathan W. Korn.

In response to the open letter, Springer said it would not publish the paper, according to MIT Technology Review. The paper you are referring to was submitted to a forthcoming conference for which Springer had planned to publish the proceedings, said the company. After a thorough peer review process the paper was rejected.

However, as the Coalition for Critical Technology makes clear, this incident is only one example in a wider trend within data science and machine learning, where researchers use socially-contingent data to try and predict or classify complex human behavior.

In one notable example from 2016, researchers from Shanghai Jiao Tong University claimed to have created an algorithm that could also predict criminality from facial features. The study was criticized and refuted, with researchers from Google and Princeton publishing a lengthy rebuttal warning that AI researchers were revisiting the pseudoscience of physiognomy. This was a discipline was founded in the 19th century by Cesare Lombroso, who claimed he could identify born criminals by measuring the dimensions of their faces.

When put into practice, the pseudoscience of physiognomy becomes the pseudoscience of scientific racism, wrote the researchers. Rapid developments in artificial intelligence and machine learning have enabled scientific racism to enter a new era, in which machine-learned models embed biases present in the human behavior used for model development.

The 2016 paper also demonstrated how easy it is for AI practitioners to fool themselves into thinking theyve found an objective system of measuring criminality. The researchers from Google and Princeton noted that, based on the data shared in the paper, all the non-criminals appeared to be smiling and wearing collared shirts and suits, while none of the (frowning) criminals were. Its possible this simple and misleading visual tell was guiding the algorithms supposed sophisticated analysis.

The Coalition for Critical Technologys letter comes at a time when movements around the world are highlighting issues of racial justice, triggered by the killing of George Floyd by law enforcement. These protests have also seen major tech companies pull back on their use of facial recognition systems, which research by Black academics has shown is racially biased.

The letters authors and signatories call on the AI community to reconsider how it evaluates the goodness of its work thinking not just about metrics like accuracy and precision, but about the social affect such technology can have on the world. If machine learning is to bring about the social good touted in grant proposals and press releases, researchers in this space must actively reflect on the power structures (and the attendant oppressions) that make their work possible, write the authors.

Read more from the original source:
AI experts say research into algorithms that claim to predict criminality must end - The Verge

What opportunities have we missed by not having these procedures in place?

Its easy to overhype whats possible, and AI was probably never going to play a huge role in this crisis. Machine-learning systems are not mature enough.

But there are a handful of cases in which AI is being tested for medical diagnosis or for resource allocation across hospitals. We might have been able to use those sorts of systems more widely, reducing some of the load on health care, had they been designed from the start with ethics in mind.

With resource allocation in particular, you are deciding which patients are highest priority. You need an ethical framework built in before you use AI to help with those kinds of decisions.

So is ethics for urgency simply a call to make existing AI ethics better?

Thats part of it. The fact that we dont have robust, practical processes for AI ethics makes things more difficult in a crisis scenario. But in times like this you also have greater need for transparency. People talk a lot about the lack of transparency with machine-learning systems as black boxes. But there is another kind of transparency, concerning how the systems are used.

This is especially important in a crisis, when governments and organizations are making urgent decisions that involve trade-offs. Whose health do you prioritize? How do you save lives without destroying the economy? If an AI is being used in public decision-making, transparency is more important than ever.

What needs to change?

We need to think about ethics differently. It shouldnt be something that happens on the side or afterwardssomething that slows you down. It should simply be part of how we build these systems in the first place: ethics by design.

I sometimes feel ethics is the wrong word. What were saying is that machine-learning researchers and engineers need to be trained to think through the implications of what theyre building, whether theyre doing fundamental research like designing a new reinforcement-learning algorithm or something more practical like developing a health-care application. If their work finds its way into real-world products and services, what might that look like? What kinds of issues might it raise?

Some of this has started already. We are working with some early-career AI researchers, talking to them about how to bring this way of thinking to their work. Its a bit of an experiment, to see what happens. But even NeurIPS [a leading AI conference] now asks researchers to include a statement at the end of their papers outlining potential societal impacts of their work.

Youve said that we need people with technical expertise at all levels of AI design and use. Why is that?

Im not saying that technical expertise is the be-all and end-all of ethics, but its a perspective that needs to be represented. And I dont want to sound like Im saying all the responsibility is on researchers, because a lot of the important decisions about how AI gets used are made further up the chain, by industry or by governments.

But I worry that the people who are making those decisions dont always fully understand the ways it might go wrong. So you need to involve people with technical expertise. Our intuitions about what AI can and cant do are not very reliable.

What you need at all levels of AI development are people who really understand the details of machine learning to work with people who really understand ethics. Interdisciplinary collaboration is hard, however. People with different areas of expertise often talk about things in different ways. What a machine-learning researcher means by privacy may be very different from what a lawyer means by privacy, and you can end up with people talking past each other. Thats why its important for these different groups to get used to working together.

Youre pushing for a pretty big institutional and cultural overhaul. What makes you think people will want to do this rather than set up ethics boards or oversight committeeswhich always make me sigh a bit because they tend to be toothless?

Yeah, I also sigh. But I think this crisis is forcing people to see the importance of practical solutions. Maybe instead of saying, Oh, lets have this oversight board and that oversight board, people will be saying, We need to get this done, and we need to get it done properly.

Read more:
If AI is going to help us in a crisis, we need a new kind of ethics - MIT Technology Review

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.comImage ID:102390038

Read this article:
Effects of the Alice Preemption Test on Machine Learning Algorithms - IPWatchdog.com

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

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.

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