The revolution that autonomous vehicles are expected to bring about in transportation, infrastructure, town-planning and energy distribution, has led to growing partnerships between auto and tech companies. PwC has proven its capabilities in the field, with its work in partnership with a Fortune 100 automaker earning it the Alconics Award for Best AI Application in the Enterprise, beating better known finalists.
The innovative solution developed by PwC used simulation and reinforcement learning, instead of more conventional machine learning or natural language processing methods, to address strategic questions related to launching a successful ridesharing network. The project teaches individual vehicles to perform collaboratively as a fleet, making decisions taking into account factors like city maps, customer wait times and charging locations.
The Strategic Implications of AI for Enterprises
The PwC project worked because, in partnership with a client that understood the nuances of AI, appropriate methodologies were used that were grounded in business use cases. Asking simply how to use AI is the wrong question; the right query is how best to approach existing issues, and choosing machine learning over traditional models if data is available and forecasting is most important.
Absence of this pragmatism raises the risk of disruption from a competitor offering the same service with lower cost and higher quality using the right AI solutions. As AI becomes a standard factor across the board, strategic advantage will come from prioritizing initiatives that provide value to the consumer. AI can be used to simulate markets or competitive dynamics, especially in completely new domains where historical data is unavailable, to improve decision making. Even in old domains, changing consumer usage patterns can inform the creation of better product and service offerings with the help of AI.
This works like what the tech community calls the ‘data flywheel’ – a positive feedback loop where customer-generated data, when analyzed by AI, increases value and greater value then increases usage. In the right conditions, this can generate explosive growth, and AI will be most important in markets where the flywheel can be strengthened.
Data Strategy and Architecture: The Most Important Ingredients
Beyond talent, data – in large volumes, of high quality and aligned to consumer priorities – is probably the most important ingredient for an AI solution, helping companies stave off competition. Such data requires deep domain expertise. Projects can fail to deliver if a solution, while technically sound, does not effectively address the business question. But, equally importantly, a clear vision without the support of talent, techniques and data doesn’t work either.
An effective, baseline architecture should include data management tools to maintain data sets as well as intermediate analysis results. It also needs the infrastructure to run models at scale and tools that provide data scientists with the flexibility to experiment with multiple techniques. Finally, there needs to be a visualization layer to enable data scientists to effectively communicate results.
AI’s Growing Role
AI technologies, driven by machine learning, are gradually transforming every industry, right from customer inquiries managed by natural language processing systems to robots in manufacturing. This is aided by exponential increases in computer processing power and volume of data creation. These effects are expected to only magnify in the coming decade as organizations transform their core processes and business models to take advantage of machine learning.
Moderate Adoption Rate
But a recent MIT Sloan study of business leaders has shown a substantial difference in numbers between those who believe in AI’s potential for change and those who have actually incorporated it in business processes. One reason for this is the confusion and hype surrounding AI. Most current investments come from R&D divisions of companies like Amazon, Baidu and Google, while elsewhere one mostly finds projects in discussion or pilot phases.
CIOs, AI and Data
Another reason is the added layer of complexity brought on by AI for CIOs already grappling with the disruption caused by digital transformation. Of course, AI can aid digital transformation, provided the data strategy from CIOs is robust. The key to effective use of AI is data that is accurate as well as meaningful and high-quality. Machine learning processes rely on the quality of data and metadata available to train the AI. So, leaders will have to invest in talent and information infrastructures, while laggards struggle with analytics expertise and easy access to data.
An important point to note is that, in most cases, AI for enterprises won’t mean entirely new applications. Machine learning techniques will be incorporated into platforms, products and services already in use in order to improve analytical power, data management and overall performance. So, a firm’s data analysts will not be replaced by AI systems, but will see their productivity and effectiveness improve.
In this environment, along with planning out their organization’s data strategy, CIOs will face two additional important tasks. This is the right time for them to educate the company leadership, and the organization in general, about the myths and realities of AI. And to complement the tools necessary to ensure quality data, they will have to put the right team in place to train AI algorithms.
So, as data-driven digital transformation continues to disrupt industries, CIOs will continue to play the role of agents of change. A data strategy that enables their organization to rapidly adopt AI will accelerate the pace of this change and ensure competitive differentiation.
Machine learning has previously been used to study brain scans (MRIs) and generate visualizations of human thought in case of simple geographic shapes or binary images. Past methods to reconstruct an image seen by a person have assumed that an image consists of pixels or simple shapes. But it is known that our brain processes visual information hierarchically extracting different levels of features or components of different complexities.
Deep Image Reconstruction
New research by four Kyoto University scientists uses neural networks as a proxy for this hierarchical human brain structure. Using the scientific platform BioRxiv and deep neural networks (DNN), the new technique lets thoughts involving sophisticated images to be decoded by a computer, producing images remarkably close to what a person is thinking.
Over 10 months, three subjects were shown natural images, artificial geometric shapes and letters for varying periods. Brain activity was measured either while the subject was looking at an image (Experiment 1) or later, when the subject was asked to think of an image shown earlier (Experiment 2).
Once the brain activity was scanned, a computer reverse-engineered the information to generate visualizations of the subject’s thoughts.
Interestingly, visual imagery could be reconstructed, albeit to a lower accuracy, even in Experiment 2. This is possibly because it is more difficult for a human to remember an image exactly as it was seen.
The mind-boggling range of possibilities, as the accuracy of this technology improves, includes:
· Making art by imagining something
· Visualization of dreams
· Visualizations of hallucinations of psychiatric patients
· Communication using thoughts
Other Research Activities
The Japanese researchers aren’t alone in this seemingly futuristic work to connect the brain with computing power. A former GoogleX-er is working to build a hat that could make telepathy possible within a decade, while another entrepreneur is working to build computer chips to implant in the brain to improve neurological functions.
The AI Index
Predicting when new developments in AI will take place has been famously difficult, but The AI Index from Stanford is an attempt to measure this, on some parameters at least.
The Approach – Measuring Hype or Actual Progress?
The AI Index tries to aggregate data across several ‘volume of activity’ metrics, like VC investments, academic conference attendance, papers published, etc. It also tracks creation of AI-related software at Github, interest in machine learning packages, and sentiment of AI-related news articles. But, this covers AI hype as much as it does progress, and the two might not always be correlated. Hype can also be cyclical in nature. To remedy this, the AI Index uses another metric.
Assessment of Progress of AI on Tasks
Measuring performance of AI systems on narrow tasks is useful. For instance, performance of computer vision in image annotation (great) or answering questions about images (not so great). But, it’s also very easy to measure – devise a metric that can be easily calculated, create a competition with a scoring system, or just compare new software with the old version.
It becomes more difficult to map narrow-task performances onto general intelligence. Computers are superhuman at chess now, or even Go, but does it mean we are any closer to general intelligence? The AI Index doesn’t attempt to offer a timeline for general intelligence because no one really knows how to measure progress. What it can do though is track the specialized performance of algorithms on tasks previously reserved for humans, like predicting skin cancer better than dermatologists. This shows that progress in AI over the next few years is likely to resemble a gradual rising tide, rather than a tsunami of general intelligence breakthrough.
Ethics of AI
Another challenge faced by the AI Index is to identify success measures by AI’s impact on people’s lives. These include the interactions between humans and AI systems; our ability to program values, ethics and oversight into these systems; and society’s flexibility in adapting to AI trends.
AI progress is a race for which we don’t know the endpoint or how to get there. This makes measuring it a daunting task. But the AI Index, as an annual collection of relevant information, is a good start.
Download the AI Index report at AIIndex.org
Here is a preview of a few of the slides of the presentation
Read the full presentation here Machine Learning 101