Login with your Social Account

artificial intelligence human brain

The brain inspires new type of AI algorithms

Machine learning which was developed 70 years ago is based on learning dynamics in the human brain. Deep learning algorithms have been able to generate results equivalent to human specialists in various areas with the help of fast and large-scale processing computers and giant data sets. However, they produce results distinct from the present knowledge of learning in neuroscience.

A team of scientists at Bar-Ilan University in Israel has illustrated a new kind of high-speed artificial intelligence algorithms which are based on the slow brain dynamics exceeding the learning rates attained to date by state-of-the-art learning algorithms using advanced experiments on neuronal cultures and simulations. The paper has been published in The Scientific Reports.

The research lead author, Prof. Ido Kanter, of Bar-Ilan University’s Department of Physics and Gonda (Goldschmied) Multidisciplinary Brain Research Center, said that till now it has been considered that neurobiology and machine learning are separate disciplines that progressed separately and the absence of likely reciprocal influence is puzzling.

He added that the data processing speed of the brain is slower than the first computer invented over 70 years ago because the number of neurons in a brain is less than the number of bits in a usual disc size of modern computers. Prof. Kanter, whose research team includes Herut Uzan, Shira Sardi, Amir Goldental and Roni Vardi also added that learning rates of the brain are very complex and isolated from the principles of learning in artificial intelligence algorithms. Since the biological system has to deal with asynchronous inputs, brain dynamics do not follow a well-defined clock synchronized for the nerve cells.

A key difference between artificial intelligence algorithms and the human brain is the nature of inputs handled. The human brain deals with asynchronous inputs, where the relative position of objects and the temporal ordering in the input are important such as identifying cars, pedestrians, other road signs while driving. On the other hand, AI algorithms deal with synchronous inputs where relative timing is ignored.

Recent studies have found that ultrafast learning rates are unexpectedly identical for small and large networks. So, the disadvantage of the complicated brain’s learning system is indeed an advantage. Another important finding is that learning can occur without learning steps through self-adaptation according to asynchronous inputs. This type of learning-without-learning occurs in the dendrites, several terminals of each neuron, as was recently experimentally observed.

The concept of productive deep learning algorithms based on the very slow brain’s dynamics provides the possibility to execute an advance type of artificial intelligence based on fast computation bridging the gap between neurobiology and artificial intelligence. Researchers conclude that understandings of our brain’s principles have to be at the centre of artificial intelligence once again.

Journal Reference: The Scientific Reports

Automating artificial intelligence for medical decision-making

Automating artificial intelligence for medical decision-making

MIT computer scientists are hoping to accelerate the use of artificial intelligence to improve medical decision-making, by automating a key step that’s usually done by hand — and that’s becoming more laborious as certain datasets grow ever-larger.

The field of predictive analytics holds increasing promise for helping clinicians diagnose and treat patients. Machine-learning models can be trained to find patterns in patient data to aid in sepsis care, design safer chemotherapy regimens, and predict a patient’s risk of having breast cancer or dying in the ICU, to name just a few examples.

Typically, training datasets consist of many sick and healthy subjects, but with relatively little data for each subject. Experts must then find just those aspects — or “features” — in the datasets that will be important for making predictions.

This “feature engineering” can be a laborious and expensive process. But it’s becoming even more challenging with the rise of wearable sensors, because researchers can more easily monitor patients’ biometrics over long periods, tracking sleeping patterns, gait, and voice activity, for example. After only a week’s worth of monitoring, experts could have several billion data samples for each subject.

In a paper being presented at the Machine Learning for Healthcare conference this week, MIT researchers demonstrate a model that automatically learns features predictive of vocal cord disorders. The features come from a dataset of about 100 subjects, each with about a week’s worth of voice-monitoring data and several billion samples — in other words, a small number of subjects and a large amount of data per subject. The dataset contain signals captured from a little accelerometer sensor mounted on subjects’ necks.

In experiments, the model used features automatically extracted from these data to classify, with high accuracy, patients with and without vocal cord nodules. These are lesions that develop in the larynx, often because of patterns of voice misuse such as belting out songs or yelling. Importantly, the model accomplished this task without a large set of hand-labeled data.

“It’s becoming increasing easy to collect long time-series datasets. But you have physicians that need to apply their knowledge to labeling the dataset,” says lead author Jose Javier Gonzalez Ortiz, a PhD student in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). “We want to remove that manual part for the experts and offload all feature engineering to a machine-learning model.”

The model can be adapted to learn patterns of any disease or condition. But the ability to detect the daily voice-usage patterns associated with vocal cord nodules is an important step in developing improved methods to prevent, diagnose, and treat the disorder, the researchers say. That could include designing new ways to identify and alert people to potentially damaging vocal behaviors.

Joining Gonzalez Ortiz on the paper is John Guttag, the Dugald C. Jackson Professor of Computer Science and Electrical Engineering and head of CSAIL’s Data Driven Inference Group; Robert Hillman, Jarrad Van Stan, and Daryush Mehta, all of Massachusetts General Hospital’s Center for Laryngeal Surgery and Voice Rehabilitation; and Marzyeh Ghassemi, an assistant professor of computer science and medicine at the University of Toronto.

Forced feature-learning

For years, the MIT researchers have worked with the Center for Laryngeal Surgery and Voice Rehabilitation to develop and analyze data from a sensor to track subject voice usage during all waking hours. The sensor is an accelerometer with a node that sticks to the neck and is connected to a smartphone. As the person talks, the smartphone gathers data from the displacements in the accelerometer.

In their work, the researchers collected a week’s worth of this data — called “time-series” data — from 104 subjects, half of whom were diagnosed with vocal cord nodules. For each patient, there was also a matching control, meaning a healthy subject of similar age, sex, occupation, and other factors.

Traditionally, experts would need to manually identify features that may be useful for a model to detect various diseases or conditions. That helps prevent a common machine-learning problem in health care: overfitting. That’s when, in training, a model “memorizes” subject data instead of learning just the clinically relevant features. In testing, those models often fail to discern similar patterns in previously unseen subjects.

“Instead of learning features that are clinically significant, a model sees patterns and says, ‘This is Sarah, and I know Sarah is healthy, and this is Peter, who has a vocal cord nodule.’ So, it’s just memorizing patterns of subjects. Then, when it sees data from Andrew, which has a new vocal usage pattern, it can’t figure out if those patterns match a classification,” Gonzalez Ortiz says.

The main challenge, then, was preventing overfitting while automating manual feature engineering. To that end, the researchers forced the model to learn features without subject information. For their task, that meant capturing all moments when subjects speak and the intensity of their voices.

As their model crawls through a subject’s data, it’s programmed to locate voicing segments, which comprise only roughly 10 percent of the data. For each of these voicing windows, the model computes a spectrogram, a visual representation of the spectrum of frequencies varying over time, which is often used for speech processing tasks. The spectrograms are then stored as large matrices of thousands of values.

But those matrices are huge and difficult to process. So, an autoencoder — a neural network optimized to generate efficient data encodings from large amounts of data — first compresses the spectrogram into an encoding of 30 values. It then decompresses that encoding into a separate spectrogram.

Basically, the model must ensure that the decompressed spectrogram closely resembles the original spectrogram input. In doing so, it’s forced to learn the compressed representation of every spectrogram segment input over each subject’s entire time-series data. The compressed representations are the features that help train machine-learning models to make predictions.

Mapping normal and abnormal features

In training, the model learns to map those features to “patients” or “controls.” Patients will have more voicing patterns than will controls. In testing on previously unseen subjects, the model similarly condenses all spectrogram segments into a reduced set of features. Then, it’s majority rules: If the subject has mostly abnormal voicing segments, they’re classified as patients; if they have mostly normal ones, they’re classified as controls.

In experiments, the model performed as accurately as state-of-the-art models that require manual feature engineering. Importantly, the researchers’ model performed accurately in both training and testing, indicating it’s learning clinically relevant patterns from the data, not subject-specific information.

Next, the researchers want to monitor how various treatments — such as surgery and vocal therapy — impact vocal behavior. If patients’ behaviors move form abnormal to normal over time, they’re most likely improving. They also hope to use a similar technique on electrocardiogram data, which is used to track muscular functions of the heart.

Materials provided by Massachusetts Institute of Technology

New AI programming language goes beyond deep learning

New AI programming language goes beyond deep learning

A team of MIT researchers is making it easier for novices to get their feet wet with artificial intelligence, while also helping experts advance the field.

In a paper presented at the Programming Language Design and Implementation conference this week, the researchers describe a novel probabilistic-programming system named “Gen.” Users write models and algorithms from multiple fields where AI techniques are applied — such as computer vision, robotics, and statistics — without having to deal with equations or manually write high-performance code. Gen also lets expert researchers write sophisticated models and inference algorithms — used for prediction tasks — that were previously infeasible.

In their paper, for instance, the researchers demonstrate that a short Gen program can infer 3-D body poses, a difficult computer-vision inference task that has applications in autonomous systems, human-machine interactions, and augmented reality. Behind the scenes, this program includes components that perform graphics rendering, deep-learning, and types of probability simulations. The combination of these diverse techniques leads to better accuracy and speed on this task than earlier systems developed by some of the researchers.

Due to its simplicity — and, in some use cases, automation — the researchers say Gen can be used easily by anyone, from novices to experts. “One motivation of this work is to make automated AI more accessible to people with less expertise in computer science or math,” says first author Marco Cusumano-Towner, a Ph.D student in the Department of Electrical Engineering and Computer Science. “We also want to increase productivity, which means making it easier for experts to rapidly iterate and prototype their AI systems.”

The researchers also demonstrated Gen’s ability to simplify data analytics by using another Gen program that automatically generates sophisticated statistical models typically used by experts to analyze, interpret, and predict underlying patterns in data. That builds on the researchers’ previous work that let users write a few lines of code to uncover insights into financial trends, air travel, voting patterns, and the spread of disease, among other trends. This is different from earlier systems, which required a lot of hand coding for accurate predictions.

“Gen is the first system that’s flexible, automated, and efficient enough to cover those very different types of examples in computer vision and data science and give state of-the-art performance,” says Vikash K. Mansinghka ’05, MEng ’09, PhD ’09, a researcher in the Department of Brain and Cognitive Sciences who runs the Probabilistic Computing Project.

Joining Cusumano-Towner and Mansinghka on the paper are Feras Saad and Alexander K. Lew, both CSAIL graduate students and members of the Probabilistic Computing Project.

Best of all worlds

In 2015, Google released TensorFlow, an open-source library of application programming interfaces (APIs) that helps beginners and experts automatically generate machine-learning systems without doing much math. Now widely used, the platform is helping democratize some aspects of AI. But, although it’s automated and efficient, it’s narrowly focused on deep-learning models which are both costly and limited compared to the broader promise of AI in general.

But there are plenty of other AI techniques available today, such as statistical and probabilistic models, and simulation engines. Some other probabilistic programming systems are flexible enough to cover several kinds of AI techniques, but they run inefficiently.

The researchers sought to combine the best of all worlds — automation, flexibility, and speed — into one. “If we do that, maybe we can help democratize this much broader collection of modeling and inference algorithms, like TensorFlow did for deep learning,” Mansinghka says.

In probabilistic AI, inference algorithms perform operations on data and continuously readjust probabilities based on new data to make predictions. Doing so eventually produces a model that describes how to make predictions on new data.

Building off concepts used in their earlier probabilistic-programming system, Church, the researchers incorporate several custom modeling languages into Julia, a general-purpose programming language that was also developed at MIT. Each modeling language is optimized for a different type of AI modeling approach, making it more all-purpose. Gen also provides high-level infrastructure for inference tasks, using diverse approaches such as optimization, variational inference, certain probabilistic methods, and deep learning. On top of that, the researchers added some tweaks to make the implementations run efficiently.

Beyond the lab

External users are already finding ways to leverage Gen for their AI research. For example, Intel is collaborating with MIT to use Gen for 3-D pose estimation from its depth-sense cameras used in robotics and augmented-reality systems. MIT Lincoln Laboratory is also collaborating on applications for Gen in aerial robotics for humanitarian relief and disaster response.

Gen is beginning to be used on ambitious AI projects under the MIT Quest for Intelligence. For example, Gen is central to an MIT-IBM Watson AI Lab project, along with the U.S. Department of Defense’s Defense Advanced Research Projects Agency’s ongoing Machine Common Sense project, which aims to model human common sense at the level of an 18-month-old child. Mansinghka is one of the principal investigators on this project.

“With Gen, for the first time, it is easy for a researcher to integrate a bunch of different AI techniques. It’s going to be interesting to see what people discover is possible now,” Mansinghka says.

Zoubin Ghahramani, chief scientist and vice president of AI at Uber and a professor at Cambridge University, who was not involved in the research, says, “Probabilistic programming is one of most promising areas at the frontier of AI since the advent of deep learning. Gen represents a significant advance in this field and will contribute to scalable and practical implementations of AI systems based on probabilistic reasoning.”

Peter Norvig, director of research at Google, who also was not involved in this research, praised the work as well. “[Gen] allows a problem-solver to use probabilistic programming, and thus have a more principled approach to the problem, but not be limited by the choices made by the designers of the probabilistic programming system,” he says. “General-purpose programming languages … have been successful because they … make the task easier for a programmer, but also make it possible for a programmer to create something brand new to efficiently solve a new problem. Gen does the same for probabilistic programming.”

Gen’s source code is publicly available and is being presented at upcoming open-source developer conferences, including Strange Loop and JuliaCon. The work is supported, in part, by DARPA.

Materials provided by Massachusetts Institute of Technology

Godfathers of AI

The Godfathers of AI receive the prestigious Turing Award

The 2018 Turing Award, acknowledged as the “Nobel Prize of computing” has been awarded to a trio of researchers who have set the foundations for the current success in artificial intelligence.

ARTIFICIAL INTELLIGENCE:
The term artificial intelligence merely refers to the intelligence that is demonstrated by the computers. Artificial intelligence is renowned for its cycles of boom and bust, and the issue of hype is as old as the field itself. When the research fails to meet the inflated expectations, it generates a freeze in the funding and interest known as an “AI winter”. It was at the tail end of one such winter in the late 1980s that Bengio, Hinton, and LeCun began exchanging ideas and working on interconnected problems. These included neural networks. They are computer programs made from connected digital neurons that have become a key building block for contemporary and modern AI.

WINNERS:
Yoshua Bengio, Geoffrey Hinton, and Yann LeCun, who are often referred to as the ‘godfathers of AI’ have been recognized with the $1 million annual prize for their work evolving the AI subfield of deep learning. The techniques the trio developed in the 1990s and 2000s assisted huge breakthroughs in tasks like computer vision and speech recognition. Their work fortifies the current proliferation of AI technologies, from self-driving cars to automated medical diagnoses.

All three have ever since taken up prominent places in the Artificial intelligence research ecosystem, by straddling with the academic world and the tech industry. Hinton splits his time between Google and the University of Toronto; Bengio is an associate professor at the University of Montreal and has started an AI company called Element AI, while LeCun is Facebook’s chief AI scientist and an instructor and professor at NYU.

Google’s head of AI, Jeff Dean also praised the trio’s achievements. “Deep neural networks are accountable for some of the greatest advances in modern computer science,” said Dean in a statement.

Let us hope that people like this trio improve AI and all of us use AI in the right way. Tell us your view on AI and what do you think about the future of AI with a quick comment.