Jeff Dean (Google Senior Fellow) – Jeff Dean,Senior Fellow, Google (Nov 2018)


Chapters

00:01:32 Recent Advances in Deep Learning
00:04:41 Machine Learning for 21st Century Grand Engineering Challenges
00:10:23 Robotics Advancements Through Data, Reinforcement Learning, and Observational Skill Acquisition
00:13:18 Machine Learning in Healthcare: From Disease Diagnosis to Cardiovascular Risk Assessment
00:20:54 Machine Learning for Medical Record Prediction and Translation
00:23:03 Machine Learning in Medical Care and Chemistry
00:27:40 TensorFlow: Tools for Machine Learning Discovery and Production
00:31:43 Automating Machine Learning with Neural Architecture Search
00:42:05 Tensor Processing Units: From Supercomputers to Edge Devices
00:45:42 AI Principles and Research Advancements

Abstract

The Evolution of Machine Learning: Transforming Vision, Language, and Healthcare

Revolutionizing Technology and Healthcare with Advanced Machine Learning

The field of machine learning has experienced transformative advancements, significantly impacting various fields, including computer vision, speech recognition, healthcare, and engineering challenges. Key breakthroughs in image classification, speech transcription, machine translation, and image captioning demonstrate the rapid evolution of this technology. The healthcare sector, in particular, has benefitted from machine learning in diagnosing diabetic retinopathy and cardiovascular risks, and in analyzing electronic health records (EHRs). In the engineering domain, machine learning drives innovations in autonomous vehicles, robotic control, and drug discovery. Alphabet’s contributions, through TensorFlow and principles for ethical AI, highlight the broad implications of these advancements.

Computer Vision and Machine Learning Progress:

The realm of computer vision has seen remarkable advances due to machine learning. Robots have gained the ability to classify images by analyzing pixel colors and can distinguish between detailed categories like various monkey species with the proper training. The annual Stanford ImageNet Challenge illustrates this progress, where models classify millions of photos into thousands of categories. Notably, the error rate has drastically dropped since 2011 to a near-human level of 3% in 2016. Furthermore, machine learning algorithms have enabled the accurate diagnosis of diabetic retinopathy, matching or surpassing the expertise of average US-certified ophthalmologists. In India, these algorithms are addressing the shortage of ophthalmologists by diagnosing diabetic retinopathy, and cardiovascular risks are now identifiable from retinal images, indicating the potential for non-invasive monitoring methods.

Speech Recognition and Machine Translation:

Machine learning has revolutionized speech recognition systems and machine translation, enabling accurate transcription of audio clips and bridging language barriers. These systems, powered by extensive labeled data, can transcribe spoken words and grasp their meanings. Recently, TensorFlow, a second-generation system by Google, has emerged as a prominent tool in machine learning research and the productionization of machine learning workloads. TensorFlow’s design facilitates the expression of new machine learning ideas and their application in production environments.

Healthcare Applications:

In healthcare, machine learning algorithms effectively parse electronic health records (EHRs), aiding in disease diagnosis and treatment planning. These predictive models can analyze medical records to forecast future health risks, thereby enhancing patient outcomes. Additionally, machine learning models are now capable of predicting mortality risk earlier and with greater accuracy, leading to timely interventions. This advancement in technology is also evident in the improved accuracy of sequential prediction tests, crucial for analyzing EHR data.

Augmented Reality in Pathology:

Pathology has benefited from a prototype augmented reality microscope, which enhances traditional microscopes with real-time image analysis. This system overlays key features in the microscope’s eyepiece, assisting pathologists in identifying critical areas during tissue examination.

Electronic and Medical Record Analysis:

Machine learning’s application in parsing EHRs and electronic medical records (EMRs) has made significant strides in disease diagnosis and treatment. Predictive models have improved patient outcomes by anticipating future health risks and helping doctors to choose the most effective treatment plans. Furthermore, machine learning has refined sequential prediction tests, essential for EMR data analysis. An innovative approach for predicting output sequences based on input sequences has applications in various fields, such as replying to messages in Gmail and machine translation. These analytical methods enable healthcare professionals to predict diseases, identify high-risk patients, recommend personalized treatments, and ultimately reduce healthcare costs.

Autonomous Vehicles and Urban Infrastructure:

Autonomous vehicles, exemplified by Waymo’s self-driving cars, are transforming urban infrastructure, promising more efficient space utilization and improved safety. These vehicles, currently in the testing phase, utilize machine learning to integrate data from various sensors, forming a comprehensive understanding of their environment. Waymo has already begun testing these self-driving cars with passengers, without safety drivers, in Phoenix, Arizona. This represents a significant shift towards more efficient and safer urban infrastructure.

Advancements in Robotic Control:

Robotic control has seen a revolution with the advent of machine learning. Traditional robots, limited by hand-coded algorithms, struggled in complex environments. Machine learning now enables robots to learn, perceive the world, and decide on appropriate actions autonomously. This development has significantly improved their ability to grasp objects and perform other tasks through observation and shared experiences.

Engineering Better Medicines:

Machine learning has accelerated quantum chemistry simulations, crucial for drug discovery. It allows for rapid screening of molecules and prediction of their properties. For example, a neural network trained on quantum chemistry simulator data can predict outcomes 300,000 times faster than the simulator itself. This technology has opened new avenues for chemists to screen large numbers of molecules efficiently, a previously unattainable feat.

TensorFlow and AutoML:

TensorFlow, developed by Alphabet, has become a cornerstone in diverse machine learning applications. Alongside TensorFlow, AutoML represents a significant advancement in automating machine learning tasks, making it more accessible and efficient. TensorFlow’s eager mode and open-source licensing under the

Apache 2.0 license have contributed to its widespread adoption and community-driven improvements. The system has been employed in various innovative applications, such as analyzing sensor data from cows to detect health issues and developing mobile systems for diagnosing plant diseases even in the absence of network connectivity. These advancements in TensorFlow and AutoML are reshaping the landscape of machine learning.

TPUs: Pioneering Machine Learning Hardware:

Alphabet’s Tensor Processing Units (TPUs) stand at the forefront of machine learning hardware innovation. These specialized chips are designed for high-speed matrix multiplication, making them ideal for complex machine learning tasks. TPUs have evolved to handle both inference and training, with Edge TPUs extending these capabilities to mobile devices. The chips are part of larger configurations called pods, which deliver powerful computing capabilities, such as the 1.5 petaflops compute of TPU V2 pods. The third-generation TPUs feature similar architecture with improvements and liquid cooling, enabling even larger pods with over 100 petaflops of compute. These advancements are exemplified in Stanford’s DUNBENCH competition, where TPUs have excelled in accuracy and price-to-performance ratios. The accessibility of TPUs through Colab notebook-based interfaces has further facilitated experimentation and development in the field.

Ethical Considerations and Algorithm Fairness:

Alphabet’s commitment to responsible AI use and addressing algorithmic biases is a crucial aspect of their approach. By prioritizing fairness in algorithm development, they are setting a standard for ethical considerations in the rapidly advancing field of machine learning.



The impact of machine learning is both profound and wide-reaching. It has significantly enhanced our understanding and capabilities in areas like image and language processing, healthcare, and engineering. The innovations in hardware, such as TPUs, and software frameworks, like TensorFlow and AutoML, are key drivers of this growth. Alphabet’s focus on ethical AI practices ensures that these technological advancements contribute positively to society. The future of machine learning is poised to be integral in addressing some of the most challenging global issues, demonstrating its critical role in the evolution of technology and healthcare.


Notes by: datagram