Home » Nova »

Deep Learning Transformation Capabilities for Industries

 

Inspired by the depth structure of the brain, Deep Learning architectures have revolutionized the approach to data analysis. Deep learning networks have won a large number of hard machine learning contests, from voice recognition to image classification to natural language processing to time-series prediction—sometimes by a large margin. Traditionally, Artifficial Intelligence (AI) has relied on heavily handcrafted features. For instance, to get decent results in image classification, several preprocessing techniques have to be applied, such as filters, edge detection, and so on. The beauty of Deep Learning is that most, if not all, features can be learned automatically from the data—provided that enough (sometimes million) training data examples are available. Deep models have feature detector units at each layer (level) that gradually extract more sophisticated and invariant features from the original raw input signals. Lower layers aim to extract simple features that are then clumped into higher layers, which in turn detect more complex features.

The falling costs of computation and the ease of accessing cloud-managed clusters have democratized AI in a way we’ve never seen before. In the past, building a computer cluster to train a deep neural network was prohibitively expensive. Today, it’s possible to run a cluster overnight to experiment with new algorithms for a few hundred dollars a month with a competent GPU-equipped PC.
AI has emerged from the labs and entered firmly into the business world with a tremendous impact on the automation of processes and services. For instance, an AI-powered CRM system could feed leads to sales reps in real time using algorithms designed to maximize the likelihood of a sale, based on breaking information about the customer, their company, and the sales rep.
Companies are pressed to build their own AI capabilities and teams and not rely on third-party consultants for this critical competency. AI cannot be seen as a one-shot process but rather a vital component in the strategy of business.
Deep Learning will affect profoundly every sector, including the automobile industry, robotics, drones, biotechnology, finance, or agriculture. According to ARK Invest’s research, companies founded on deep learning will unlock trillions of dollars in productivity gains and add $17 trillion in market capitalization to global equities during the next two decades;
• $17 trillion in market capitalization creation from deep learning companies by 2036
• $6 trillion in revenue from autonomous on-demand transportation by 2027
• $6 billion in revenue for deep learning processors in the data center by 2022, growing more than tenfold over 5 years
• $16 billion addressable market for diagnostic radiology
• $100 to $170 billion in savings and profit from improved credit scoring
• $12 trillion in real GDP growth in the US from automation by 2035

Processor performance has improved roughly five orders of magnitude since Intel’s original Pentium processor. But the performance of deep learning programs also depends on the amount of data used for training. Thanks to the Internet’s size and scale, deep learning has thrived with access to very large data sets at a minimal cost. While the 1990 LeCun handwriting reader used approximately 10,000 samples collected from the U.S. Postal Service, the 2009 ImageNet data set contains more than 10 million examples of high-resolution photographs. Also, Baidu’s DeepSpeech is trained upon more than 10,000 hours of audio data compared to a few hundred hours in legacy data sets.
Neural nets themselves have become larger and more sophisticated, as measured by their number of free “parameters.” Networks with a billion parameters are common nowadays. Larger networks allow for a more expressive capability to capture relations in the data. Today’s deep learning networks have roughly ten million parameters, or four orders of magnitude more than LeCun’s original handwriting reader.

Deep Learning Opportunity
Deep learning–powered AI is already transforming most industries. AI will fundamentally change and automate numerous functions within companies, from pricing, budget allocation, fraud detection, and security to marketing optimization. But for an organization to take full advantage of AI, it needs to be fully integrated across all different departments and functions; this will enable organizations to truly become customer-centric.

Deep learning is well suited for data-intensive activities such as advertising and click-through information. Most of the data will be collected by mobile phones, and a myriad of devices will deliver real-time georeferenced information. Multimodal learning will allow companies to integrate text, images, video, and sound with a unified representation.
The implications of Deep Learning technology applied to certain areas, like self-driving cars, are obvious, and its consequences could revolutionize transportation systems and car ownership. In other areas, the impact may not seem so obvious and immediate; however, as Deep Learning technology progress, many more industries will also be at risk of being disrupted. Some will be enumerated.

Radiology and Medical Imagery
Deep learning is making rapid advances in diagnostic radiology. The ARK report estimates that the total global addressable market for computer- aided diagnostics software could be worth $16 billion. From revenues of $1 billion today, the growth in medical software companies and imaging device manufacturers could average 20 percent to 35 percent per year as deep learning enhances their productivity and creates new products and services during the next 10 to 15 years.
Diagnostic radiology is essential to modern healthcare; yet the visual interpretation of medical images is a laborious and error-prone process. Historically the average diagnosis error rate among radiologists is around 30 percent, according to https://www.ncbi.nlm.nih.gov/pmc/articles/ PMC1955762/. Because of rudimentary technology, lung cancer nodules are routinely missed, especially at earlier stages of development, and 8 percent to 10 percent of bone fractures are missed or misdiagnosed. Initially, radiologists miss roughly two-thirds of breast cancers in mammograms that are visible in retrospective reviews.
Intelligent software powered by deep learning has the potential to change the status quo. Early results are promising: the latest deep learning systems already outperform radiologists and existing algorithms in a variety of diagnostic tasks.
Early diagnosis is key to successful treatment. Each year more than 2 million people worldwide die from lung and breast cancers according to Cancer Research UK. If 10 percent of later-stage cases could be caught at stage 1 with Computer Aided Design (CAD), ARK estimates it would save 150,000 life years. Valuing human life at $50,000 per year, 51 breast or lung diagnoses at stage 1 would equate to $7.6 billion of life value saved. Impacting a wide range of radiology problems from bone fractures to Alzheimer’s disease, the value of deep learning would be orders of magnitude greater.

Self-Driving Cars
Considering that 94 percent of car accidents originate from human error and that, on average, a driver in Europe spends six hours a week in traffic jams, it’s not difficult to accept that one of the most transformative applications of deep learning is self-driving cars. By some estimates, self- driving cars can reduce the traffic in cities by as much as 90 percent and increase free space, presently devoted to parking, by as much.
Without deep learning, fully autonomous vehicles would be unconceivable. Navigating a vehicle through streets, weather conditions, and unpredictable traffic is an open-ended problem that learning algorithms such as deep learning can solve. ARK believes that deep learning is a fundamental requirement for level 4 or higher autonomous driving (level 5 corresponds to fully autonomous vehicles).
Deep learning solves two key problems facing autonomous driving: sensing and path planning. Neural nets allow a computer to segment the world into drivable and nondrivable paths, detect obstacles, interpret road signs, and respond to traffic lights. Additionally, with reinforcement learning, neural nets can learn how to change lanes, use roundabouts, and navigate around complex traffic conditions.
While self-driving systems have yet to reach the level required for autonomous driving, the observed rate of progress from Google and others suggests that self-driving technology will be available by the end of this decade. Fully deployed, self-driving technology will reduce the cost of transport and bring to life mobility-as-a-service (MaaS). Based on ARK’s research, by 2020 not only will most cars have autonomous driving capabilities but the cost of travel will fall to $0.35 per mile, roughly one-tenth the cost of human-driven taxis. As a result, transportation will transition primarily to an on-demand model, introducing a flood of new consumers to the point-to-point mobility market. The number of autonomous miles driven will rise dramatically from de minimis to 18 trillion per year by 2027. At $0.35 per mile, the market for autonomous on-demand transport will approximate a $6 trillion market in ten years.

Building value with Deep Learning
Deep Learning is associated either with startups or with big companies like Google, Amazon, or Baidu. However, traditional business can also profit from this transformative technology that is fast leveraging the competitive landscape.
From a business perspective, it’s important to have a solid grounding in the fundamentals of data science and the algorithms behind deep learning to grasp its far-reaching strategic implications within an organization and not just go with the hype. The implications of having a data-centric business culture are not only useful for a specific problem but are unfolding a set of forces that will lead to the application of similar methodologies in different departments.
The customer-centric view requires the collection of vast amounts of data and the capabilities to learn robustly on unstructured data. DL provides the tools for such an approach that could provide substantial uplift, for instance for targeting the right customers, over traditional marketing campaigns.
These ideas diffused to the online advertising industry and online advertising to incorporate the data of online social connections. Companies consider how they can obtain a competitive advantage from their data and their data science capabilities. Data is a strategic asset, but you need to think carefully as to how data and data science can provide value in the context of your business strategy and also whether it would do the same in the context of your competitors’ strategies.

Sometimes is not the data nor the algorithms that create the strategic value but how the extracted insights are implemented in improving products, customer service, and, most important, reorganizing business processes to transform the business. The effectiveness of a predictive model may depend critically on the problem engineering, the attributes created, the combining of different models, and so on. Even if algorithms are published, many implementation details may be critical for getting a solution that works in the lab to work in production.
Success may also depend on intangible assets such as a company culture—a culture that embraces business experimentation is completely different from one that doesn’t. The criteria of success is not the accuracy of the model that data scientists design; it’s the value created from what the business implements.

 

Related Posts

  • No Related Posts