Jeff Dean (Google Senior Fellow) – O’Reilly Conference (Nov 2019)


Chapters

00:00:02 Machine Learning for the Next 100 Years
00:04:58 Machine Learning Applications in Autonomous Vehicles and Health Informatics
00:08:40 Recent Advances in Machine Learning: From Language Understanding to AutoML
00:20:09 AutoML-Based Solutions Surpass Human-Guided Experimentation in Machine Learning
00:22:15 A Vision for the Future of Machine Learning

Abstract

Revolutionizing the World: The Transformative Impact of Machine Learning in Multiple Domains

Machine learning has fundamentally transformed numerous fields, from healthcare and language processing to autonomous vehicles and urban infrastructure. This article delves into the substantial growth of TensorFlow and machine learning, highlighting their pivotal role in advancing computer vision, speech recognition, and various other applications. Emphasizing Google’s innovations, such as the AutoML technology and ethical AI principles, the article explores the profound implications of these technologies in addressing some of the grand challenges identified by the US National Academy of Engineering, like healthcare and infrastructure improvement.

Growth of Machine Learning and TensorFlow:

Since its open-source release, TensorFlow has experienced substantial growth, reflecting the worldwide surge in machine learning interest. Machine learning research is expanding at a rate surpassing even the historical exponential growth of computational power, as outlined by Moore’s Law. AutoML systems, a significant breakthrough, perform automated experimentation to optimize machine learning models. These systems have demonstrated superior accuracy and computational efficiency compared to human-guided experimentation, showing versatility across various tasks such as image recognition, language translation, and autonomous vehicle computer vision.

Impact on Computer Vision and Speech Recognition:

The realm of computer vision has been revolutionized by deep neural networks, enabling precise object recognition and classification directly from raw pixel data. Similarly, significant advancements in speech recognition technology have enabled models to transcribe audio to text with high accuracy. Moreover, multimodal models have emerged, capable of synthesizing both visual and auditory data to produce enhanced image descriptions.

Progress in Computer Vision:

Illustrating the rapid advancements in this field, the ImageNet contest revealed a dramatic drop in error rates for image classification, plummeting from 26% in 2011 to just 3% in 2016.

14 Challenges for Science and Engineering:

Jeff Dean’s talk highlights 14 challenges identified by the US National Academy of Engineering, earmarked as crucial for the next 100 years. These challenges span a diverse range of fields and offer opportunities for significant improvements in human life.

Restoring and Improving Urban Infrastructure:

In addressing the challenges of aging infrastructure, AI emerges as a crucial tool. It enables predictive maintenance to prevent failures, optimized traffic management to alleviate congestion, and improved strategies for energy distribution and conservation.

Autonomous Vehicles:

Machine learning and computer vision synergistically process raw sensor data to provide a comprehensive understanding of a car’s surroundings. Alphabet’s Waymo subsidiary, for example, has conducted real-passenger tests in Phoenix without safety drivers, illustrating the commercial viability of autonomous vehicles due to machine learning advancements.

Advanced Health Informatics:

Machine learning has the potential to democratize healthcare expertise, making it accessible worldwide. A notable application is in the screening for diabetic retinopathy, a preventable cause of blindness often hampered by a shortage of specialists. In India, for instance, 45% of patients with diabetic retinopathy experience vision loss before diagnosis due to the scarcity of eye doctors. Machine learning models offer a viable solution to effectively screen for this condition, potentially reducing instances of blindness.

Potential Solution for Diabetic Retinopathy:

Google Health’s machine learning model diagnoses diabetic retinopathy with accuracy comparable to that of specialists, offering a significant advantage in regions with limited access to specialists.

Diabetic Retinopathy Screening:

Deep learning models can diagnose diabetic retinopathy by analyzing retinal images. Training these models requires high-quality labeled data from multiple ophthalmologists. Additionally, explainability techniques are being developed to help clinicians understand the predictions made by these models.

Language Understanding with Transformers:

The introduction of Transformer models, which process text in parallel, has enhanced efficiency in language understanding. BERT (Bi-directional Encoder Representations from Transformers), for instance, uses context from all around the text and is pre-trained on vast text datasets before being fine-tuned for specific tasks.

Automating Machine Learning – AutoML:

AutoML focuses on automating the experimental process in machine learning. It includes neural architecture search, which generates and evaluates various neural network architectures. Models created through AutoML have shown accuracy comparable to those designed by human experts.

Broad Applications of Machine Learning:

The versatility of machine learning is evident in its applications, ranging from scientific discovery to social impact. TensorFlow’s widespread adoption is a testament to its utility across various domains.

AutoML Outperforms Human-Guided Experimentation:

In terms of accuracy and efficiency, AutoML systems, exemplified by EfficientNet models, have surpassed human-designed models.

Desired Properties of Machine Learning Models:

Ideal machine learning models should be capable of multitasking, adapt easily to new tasks, resist adversarial attacks, and be interpretable. They should also demonstrate fairness in predictions. Models like sparsely activated models and neural pathway search are examples of advancements in these areas.

Thoughtful Use of AI in Society:

Google’s AI principles emphasize responsible and ethical use of AI, focusing on reducing bias, promoting interpretability, and ensuring fairness. Privacy-preserving techniques like federated learning are being explored to protect user data. Efforts are also underway to develop interpretable models that aid clinicians in understanding machine learning predictions.



Deep neural networks and machine learning are driving progress across various domains, offering solutions to complex challenges. Google’s commitment to ethical AI ensures that these technologies not only push the boundaries of innovation but also adhere to principles that prioritize societal welfare and responsible usage.


Notes by: ChannelCapacity999