Archive for the ‘Machine Learning’ Category

Teslas acquisition of DeepScale starts to pay off with new IP in machine learning – Electrek

Teslas acquisition of machine-learning startup DeepScale is starting to pay off, with the team hired through the acquisition starting to deliver new IP for the automaker.

Late last year, it was revealed that Tesla acquired DeepScale, a Bay Area-based startup that focuses on Deep Neural Network (DNN) for self-driving vehicles, for an undisclosed amount.

They specialized in computing power-efficient deep learning systems, which is also an area of focus for Tesla, who decided to design its own computer chip to power its self-driving software.

There was speculation that Tesla acquired the small startup team in order to accelerate its machine learning development.

Now we are seeing some of that teams work, thanks to a new patent application.

Just days after Tesla acquired the startup in October 2019, the automaker applied for a new patent with three members of DeepScale listed as inventors: Matthew Cooper, Paras Jain, and Harsimran Singh Sidhu.

The patent application called Systems and Methods for Training Machine Models with Augmented Data was published yesterday.

Tesla writes about it in the application:

Systems and methods for training machine models with augmented data. An example method includes identifying a set of images captured by a set of cameras while affixed to one or more image collection systems. For each image in the set of images, a training output for the image is identified. For one or more images in the set of images, an augmented image for a set of augmented images is generated. Generating an augmented image includes modifying the image with an image manipulation function that maintains camera properties of the image. The augmented training image is associated with the training output of the image. A set of parameters of the predictive computer model are trained to predict the training output based on an image training set including the images and the set of augmented images.

The system that the DeepScale team, now working under Tesla, is trying to patent here is related to training a neural net using data from several different sensors observing scenes, like the eight cameras in Teslas Autopilot sensor array.

They write about the difficulties of such a situation in the patent application:

In typical machine learning applications, data may be augmented in various ways to avoid overfitting the model to the characteristics of the capture equipment used to obtain the training data. For example, in typical sets of images used for training computer models, the images may represent objects captured with many different capture environments having varying sensor characteristics with respect to the objects being captured. For example, such images may be captured by various sensor characteristics, such as various scales (e.g., significantly different distances within the image), with various focal lengths, by various lens types, with various pre- or post-processing, different software environments, sensor array hardware, and so forth. These sensors may also differ with respect to different extrinsic parameters, such as the position and orientation of the imaging sensors with respect to the environment as the image is captured. All of these different types of sensor characteristics can cause the captured images to present differently and variously throughout the different images in the image set and make it more difficult to properly train a computer model.

Here they summarize their solution to the problem:

One embodiment is a method for training a set of parameters of a predictive computer model. This embodiment may include: identifying a set of images captured by a set of cameras while affixed to one or more image collection systems; for each image in the set of images, identifying a training output for the image; for one or more images in the set of images, generating an augmented image for a set of augmented images by: generating an augmented image for a set of augmented images by modifying the image with an image manipulation function that maintains camera properties of the image, and associating the augmented training image with the training output of the image; training the set of parameters of the predictive computer model to predict the training output based on an image training set including the images and the set of augmented images.

An additional embodiment may include a system having one or more processors and non-transitory computer storage media storing instructions that when executed by the one or more processors, cause the processors to perform operations comprising: identifying a set of images captured by a set of cameras while affixed to one or more image collection systems; for each image in the set of images, identifying a training output for the image; for one or more images in the set of images, generating an augmented image for a set of augmented images by: generating an augmented image for a set of augmented images by modifying the image with an image manipulation function that maintains camera properties of the image, and associating the augmented training image with the training output of the image; training the set of parameters of the predictive computer model to predict the training output based on an image training set including the images and the set of augmented images.

Another embodiment may include a non-transitory computer-readable medium having instructions for execution by a processor, the instructions when executed by the processor causing the processor to: identify a set of images captured by a set of cameras while affixed to one or more image collection systems; for each image in the set of images, identify a training output for the image; for one or more images in the set of images, generate an augmented image for a set of augmented images by: generate an augmented image for a set of augmented images by modifying the image with an image manipulation function that maintains camera properties of the image, and associate the augmented training image with the training output of the image; train the computer model to learn to predict the training output based on an image training set including the images and the set of augmented images.

As we previously reported, Tesla is going through a significant foundational rewrite in the Tesla Autopilot. As part of the rewrite, CEO Elon Musk says that the neural net is absorbing more and more of the problem.

It will also include a more in-depth labeling system.

Musk described 3D labeling as a game-changer:

Its where the car goes into a scene with eight cameras, and kind of paint a path, and then you can label that path in 3D.

This new way to train machine learning systems with multiple cameras, like Teslas Autopilot, with augmented data could be part of this new Autopilot update.

Here are some drawings from the patent application:

Heres Teslas patent application in full:

FTC: We use income earning auto affiliate links. More.

Subscribe to Electrek on YouTube for exclusive videos and subscribe to the podcast.

Originally posted here:
Teslas acquisition of DeepScale starts to pay off with new IP in machine learning - Electrek

New AI improves itself through Darwinian-style evolution – Big Think

Machine learning has fundamentally changed how we engage with technology. Today, it's able to curate social media feeds, recognize complex images, drive cars down the interstate, and even diagnose medical conditions, to name a few tasks.

But while machine learning technology can do some things automatically, it still requires a lot of input from human engineers to set it up, and point it in the right direction. Inevitably, that means human biases and limitations are baked into the technology.

So, what if scientists could minimize their influence on the process by creating a system that generates its own machine-learning algorithms? Could it discover new solutions that humans never considered?

To answer these questions, a team of computer scientists at Google developed a project called AutoML-Zero, which is described in a preprint paper published on arXiv.

"Human-designed components bias the search results in favor of human-designed algorithms, possibly reducing the innovation potential of AutoML," the paper states. "Innovation is also limited by having fewer options: you cannot discover what you cannot search for."

Automatic machine learning (AutoML) is a fast-growing area of deep learning. In simple terms, AutoML seeks to automate the end-to-end process of applying machine learning to real-world problems. Unlike other machine-learning techniques, AutoML requires relatively little human effort, which means companies might soon be able to utilize it without having to hire a team of data scientists.

AutoML-Zero is unique because it uses simple mathematical concepts to generate algorithms "from scratch," as the paper states. Then, it selects the best ones, and mutates them through a process that's similar to Darwinian evolution.

AutoML-Zero first randomly generates 100 candidate algorithms, each of which then performs a task, like recognizing an image. The performance of these algorithms is compared to hand-designed algorithms. AutoML-Zero then selects the top-performing algorithm to be the "parent."

"This parent is then copied and mutated to produce a child algorithm that is added to the population, while the oldest algorithm in the population is removed," the paper states.

The system can create thousands of populations at once, which are mutated through random procedures. Over enough cycles, these self-generated algorithms get better at performing tasks.

"The nice thing about this kind of AI is that it can be left to its own devices without any pre-defined parameters, and is able to plug away 24/7 working on developing new algorithms," Ray Walsh, a computer expert and digital researcher at ProPrivacy, told Newsweek.

If computer scientists can scale up this kind of automated machine-learning to complete more complex tasks, it could usher in a new era of machine learning where systems are designed by machines instead of humans. This would likely make it much cheaper to reap the benefits of deep learning, while also leading to novel solutions to real-world problems.

Still, the recent paper was a small-scale proof of concept, and the researchers note that much more research is needed.

"Starting from empty component functions and using only basic mathematical operations, we evolved linear regressors, neural networks, gradient descent... multiplicative interactions. These results are promising, but there is still much work to be done," the scientists' preprint paper noted.

Related Articles Around the Web

Here is the original post:
New AI improves itself through Darwinian-style evolution - Big Think

Automated Machine Learning is the Future of Data Science – Analytics Insight

As the fuel that powers their progressing digital transformation endeavors, organizations wherever are searching for approaches to determine as much insight as could reasonably be expected from their data. The accompanying increased demand for advanced predictive and prescriptive analytics has, thus, prompted a call for more data scientists capable with the most recent artificial intelligence (AI) and machine learning (ML) tools.

However, such highly-skilled data scientists are costly and hard to find. Truth be told, theyre such a valuable asset, that the phenomenon of the citizen data scientist has of late emerged to help close the skills gap. A corresponding role, as opposed to an immediate substitution, citizen data scientists need explicit advanced data science expertise. However, they are fit for producing models utilizing best in class diagnostic and predictive analytics. Furthermore, this ability is incomplete because of the appearance of accessible new technologies, for example, automated machine learning (AutoML) that currently automate a significant number of the tasks once performed by data scientists.

The objective of autoML is to abbreviate the pattern of trial and error and experimentation. It burns through an enormous number of models and the hyperparameters used to design those models to decide the best model available for the data introduced. This is a dull and tedious activity for any human data scientist, regardless of whether the individual in question is exceptionally talented. AutoML platforms can play out this dreary task all the more rapidly and thoroughly to arrive at a solution faster and effectively.

A definitive estimation of the autoML tools isnt to supplant data scientists however to offload their routine work and streamline their procedure to free them and their teams to concentrate their energy and consideration on different parts of the procedure that require a more significant level of reasoning and creativity. As their needs change, it is significant for data scientists to comprehend the full life cycle so they can move their energy to higher-value tasks and sharpen their abilities to additionally hoist their value to their companies.

At Airbnb, they continually scan for approaches to improve their data science workflow. A decent amount of their data science ventures include machine learning and numerous pieces of this workflow are tedious. At Airbnb, they use machine learning to build customer lifetime value models (LTV) for guests and hosts. These models permit the company to improve its decision making and interactions with the community.

Likewise, they have seen AML tools as generally valuable for regression and classification problems involving tabular datasets, anyway, the condition of this area is rapidly progressing. In outline, it is accepted that in specific cases AML can immensely increase a data scientists productivity, often by an order of magnitude. They have used AML in many ways.

Unbiased presentation of challenger models: AML can rapidly introduce a plethora of challenger models utilizing a similar training set as your incumbent model. This can help the data scientist in picking the best model family. Identifying Target Leakage: In light of the fact that AML builds candidate models amazingly fast in an automated way, we can distinguish data leakage earlier in the modeling lifecycle. Diagnostics: As referenced prior, canonical diagnostics can be automatically created, for example, learning curves, partial dependence plots, feature importances, etc. Tasks like exploratory data analysis, pre-processing of data, hyper-parameter tuning, model selection and putting models into creation can be automated to some degree with an Automated Machine Learning system.

Companies have moved towards enhancing predictive power by coupling huge data with complex automated machine learning. AutoML, which uses machine learning to create better AI, is publicized as affording opportunities to democratise machine learning by permitting firms with constrained data science expertise to create analytical pipelines equipped for taking care of refined business issues.

Including a lot of algorithms that automate that writing of other ML algorithms, AutoML automates the end-to-end process of applying ML to real-world problems. By method for representation, a standard ML pipeline consists of the following: data pre-processing, feature extraction, feature selection, feature engineering, algorithm selection, and hyper-parameter tuning. In any case, the significant ability and time it takes to execute these strides imply theres a high barrier to entry.

In an article distributed on Forbes, Ryohei Fujimaki, the organizer and CEO of dotData contends that the discussion is lost if the emphasis on AutoML systems is on supplanting or decreasing the role of the data scientist. All things considered, the longest and most challenging part of a typical data science workflow revolves around feature engineering. This involves interfacing data sources against a rundown of wanted features that are assessed against different Machine Learning algorithms.

Success with feature engineering requires an elevated level of domain aptitude to recognize the ideal highlights through a tedious iterative procedure. Automation on this front permits even citizen data scientists to make streamlined use cases by utilizing their domain expertise. More or less, this democratization of the data science process makes the way for new classes of developers, offering organizations a competitive advantage with minimum investments.

Go here to read the rest:
Automated Machine Learning is the Future of Data Science - Analytics Insight

Research Team Uses Machine Learning to Track COVID-19 Spread in Communities and Predict Patient Outcomes – The Ritz Herald

The COVID-19 pandemic is raising critical questions regarding the dynamics of the disease, its risk factors, and the best approach to address it in healthcare systems. MIT Sloan School of Management Prof. Dimitris Bertsimas and nearly two dozen doctoral students are using machine learning and optimization to find answers. Their effort is summarized in the COVIDanalytics platform where their models are generating accurate real-time insight into the pandemic. The group is focusing on four main directions; predicting disease progression, optimizing resource allocation, uncovering clinically important insights, and assisting in the development of COVID-19 testing.

The backbone for each of these analytics projects is data, which weve extracted from public registries, clinical Electronic Health Records, as well as over 120 research papers that we compiled in a new database. Were testing our models against incoming data to determine if it makes good predictions, and we continue to add new data and use machine-learning to make the models more accurate, says Bertsimas.

The first project addresses dilemmas at the front line, such as the need for more supplies and equipment. Protective gear must go to healthcare workers and ventilators to critically ill patients. The researchers developed an epidemiological model to track the progression of COVID-19 in a community, so hospitals can predict surges and determine how to allocate resources.

The team quickly realized that the dynamics of the pandemic differ from one state to another, creating opportunities to mitigate shortages by pooling some of the ventilator supply across states. Thus, they employed optimization to see how ventilators could be shared among the states and created an interactive application that can help both the federal and state governments.

Different regions will hit their peak number of cases at different times, meaning their need for supplies will fluctuate over the course of weeks. This model could be helpful in shaping future public policy, notes Bertsimas.

Recently, the researchers connected with long-time collaborators at Hartford HealthCare to deploy the model, helping the network of seven campuses to assess their needs. Coupling county level data with the patient records, they are rethinking the way resources are allocated across the different clinics to minimize potential shortages.

The third project focuses on building a mortality and disease progression calculator to predict whether someone has the virus, and whether they need hospitalization or even more intensive care. He points out that current advice for patients is at best based on age, and perhaps some symptoms. As data about individual patients is limited, their model uses machine learning based on symptoms, demographics, comorbidities, lab test results as well as a simulation model to generate patient data. Data from new studies is continually added to the model as it becomes available.

We started with data published in Wuhan, Italy, and the U.S., including infection and death rate as well as data coming from patients in the ICU and the effects of social isolation. We enriched them with clinical records from a major hospital in Lombardy which was severely impacted by the spread of the virus. Through that process, we created a new model that is quite accurate. Its power comes from its ability to learn from the data, says Bertsimas.

By probing the severity of the disease in a patient, it can actually guide clinicians in congested areas in a much better way, says Bertsimas.

Their fourth project involves creating a convenient test for COVID-19. Using data from about 100 samples from Morocco, the group is using machine-learning to augment a test previously designed at the Mohammed VI Polytechnic University to come up with more precise results. The model can accurately detect the virus in patients around 90% of the time, while false positives are low.

The team is currently working on expanding the epidemiological model to a global scale, creating more accurate and informed clinical risk calculators, and identifying potential ways that would allow us to go back to normality.

We have released all our source code and made the public database available for other people too. We will continue to do our own analysis, but if other people have better ideas, we welcome them, says Bertsimas.

More:
Research Team Uses Machine Learning to Track COVID-19 Spread in Communities and Predict Patient Outcomes - The Ritz Herald

Model quantifies the impact of quarantine measures on Covid-19’s spread – MIT News

The research described in this article has been published on a preprint server but has not yet been peer-reviewed by scientific or medical experts.

Every day for the past few weeks, charts and graphs plotting the projected apex of Covid-19 infections have been splashed across newspapers and cable news. Many of these models have been built using data from studies on previous outbreaks like SARS or MERS. Now, a team of engineers at MIT has developed a model that uses data from the Covid-19 pandemic in conjunction with a neural network to determine the efficacy of quarantine measures and better predict the spread of the virus.

Our model is the first which uses data from the coronavirus itself and integrates two fields: machine learning and standard epidemiology, explains Raj Dandekar, a PhD candidate studying civil and environmental engineering. Together with George Barbastathis, professor of mechanical engineering, Dandekar has spent the past few months developing the model as part of the final project in class 2.168 (Learning Machines).

Most models used to predict the spread of a disease follow what is known as the SEIR model, which groups people into susceptible, exposed, infected, and recovered. Dandekar and Barbastathis enhanced the SEIR model by training a neural network to capture the number of infected individuals who are under quarantine, and therefore no longer spreading the infection to others.

The model finds that in places like South Korea, where there was immediate government intervention in implementing strong quarantine measures, the virus spread plateaued more quickly. In places that were slower to implement government interventions, like Italy and the United States, the effective reproduction number of Covid-19 remains greater than one, meaning the virus has continued to spread exponentially.

The machine learning algorithm shows that with the current quarantine measures in place, the plateau for both Italy and the United States will arrive somewhere between April 15-20. This prediction is similar to other projections like that of the Institute for Health Metrics and Evaluation.

Our model shows that quarantine restrictions are successful in getting the effective reproduction number from larger than one to smaller than one, says Barbastathis. That corresponds to the point where we can flatten the curve and start seeing fewer infections.

Quantifying the impact of quarantine

In early February, as news of the virus troubling infection rate started dominating headlines, Barbastathis proposed a project to students in class 2.168. At the end of each semester, students in the class are tasked with developing a physical model for a problem in the real world and developing a machine learning algorithm to address it. He proposed that a team of students work on mapping the spread of what was then simply known as the coronavirus.

Students jumped at the opportunity to work on the coronavirus, immediately wanting to tackle a topical problem in typical MIT fashion, adds Barbastathis.

One of those students was Dandekar. The project really interested me because I got to apply this new field of scientific machine learning to a very pressing problem, he says.

As Covid-19 started to spread across the globe, the scope of the project expanded. What had originally started as a project looking just at spread within Wuhan, China grew to also include the spread in Italy, South Korea, and the United States.

The duo started modeling the spread of the virus in each of these four regions after the 500th case was recorded. That milestone marked a clear delineation in how different governments implemented quarantine orders.

Armed with precise data from each of these countries, the research team took the standard SEIR model and augmented it with a neural network that learns how infected individuals under quarantine impact the rate of infection. They trained the neural network through 500 iterations so it could then teach itself how to predict patterns in the infection spread.

Using this model, the research team was able to draw a direct correlation between quarantine measures and a reduction in the effective reproduction number of the virus.

The neural network is learning what we are calling the quarantine control strength function, explains Dandekar. In South Korea, where strong measures were implemented quickly, the quarantine control strength function has been effective in reducing the number of new infections. In the United States, where quarantine measures have been slowly rolled out since mid-March, it has been more difficult to stop the spread of the virus.

Predicting the plateau

As the number of cases in a particular country decreases, the forecasting model transitions from an exponential regime to a linear one. Italy began entering this linear regime in early April, with the U.S. not far behind it.

The machine learning algorithm Dandekar and Barbastathis have developed predictedthat the United States will start to shift from an exponential regime to a linear regime in the first week of April, with a stagnation in the infected case count likely betweenApril 15 and April20. It also suggests that the infection count will reach 600,000 in the United States before the rate of infection starts to stagnate.

This is a really crucial moment of time. If we relax quarantine measures, it could lead to disaster, says Barbastathis.

According to Barbastathis, one only has to look to Singapore to see the dangers that could stem from relaxing quarantine measures too quickly. While the team didnt study Singapores Covid-19 cases in their research, the second wave of infection this country is currently experiencing reflects their models finding about the correlation between quarantine measures and infection rate.

If the U.S. were to follow the same policy of relaxing quarantine measures too soon, we have predicted that the consequences would be far more catastrophic, Barbastathis adds.

The team plans to share the model with other researchers in the hopes that it can help inform Covid-19 quarantine strategies that can successfully slow the rate of infection.

Original post:
Model quantifies the impact of quarantine measures on Covid-19's spread - MIT News