It’s Eternal Spring For AI

The field of Artificial Intelligence (AI) has had many ups and downs largely due to unrealistic expectations created by everyone involved including researchers, sponsors, developers, and even consumers. The “reemergence” of AI has lot to do with recent developments in supporting technologies and fields such as sensors, computing at macro and micro scales, communication networks and progress in deep learning and other reasoning methods. What is enabling the reemergence of AI as a major force is the ability of AI algorithms to directly interface with sensors and data (microphones, cameras, medical imaging scanners, bio-signals, financial and medical records, and the like).

The early days – AI tools at the end of the pipeline
In its early days, AI was mainly focused on developing strategies for searching over a large space of solutions. The solution space was often far removed from data, separated by methods that were designed to develop appropriate representations that formed the solution space. Early successes with board games—and more recent ones, demonstrated on the game show Jeopardy—point to this general trend. AI also benefitted by logical and probabilistic reasoning methods. One of the features of these approaches was that the representations or attributes extracted from the sensor data were handcrafted by scientists and engineers. This approach dominated the fields of computer vision, speech, and speaker recognition. As an example, in computer vision—which aims to describe a scene, objects, and their interactions in the scene using images or video collected by cameras—the researcher would come up with representations such as edges, linear features, texture, motion, and depth. A reasoning algorithm would then process these features to answer questions on who, what, where, why, and so on. In this traditional scenario, AI tools were invoked at the end of the pipeline.

There were many issues in the traditional pipeline described above. Due to the many intermediate steps between data and AI, real-time decisions were hard to realize. Front-end algorithms that extracted attributes or features were prone to errors requiring sophisticated reasoning methods for decision making. These mechanisms were ill suited for handling large-scale problems in computer vision. While there were some pioneering demonstrations of autonomous driving on the autobahns and Pittsburgh, Pennsylvania to San Diego, California, these systems did not lead to large-scale adoption. Unsuccessful attempts to design and build large-scale AI systems can be traced to the burden on the AI algorithms due to the multiple steps between sensors and AI, lending credence to the common saying, “There is many a slip between the cup and the lip.” The situation with speech and speaker identification and verification was much better, probably due to reduced dimensions (1-D versus 2-D or 3-D), rather close proximity between sensors and subjects and a better understanding of mechanisms that generate speech signals.

What has changed? The sensor-to-reasoning direct pipeline
Decision-making or reasoning algorithms are now directly talking to sensors and data, completely sidestepping the burdensome tasks of generating the hand-crafted intermediate representations upon which AI methods were designed to operate on. Since 2012, the sensor-to-reasoning direct pipeline has shown remarkable performance improvements in many computer vision tasks such as object and face detection and verification, and the floodgates have opened. The availability of large amounts of labeled data collected by sensors is one of the factors in the reemergence of AI as a pervasive technology.

Well, performance does matter. DCNNs are creating new state-of-the art results in almost every problem that lends itself to learnable data-to-decision mapping. A significant number of papers in computer vision and other conferences use DCNNs and their variations. It’s like we have all settled down for one type of ice-cream!

What comes next: Cold winter or eternal spring?
Next comes the question, will AI face another winter? I don’t believe so. The representations created by the data to decision paradigm are so much better than most hand-crafted features, I believe this methodology is here to say. In fact, the sustainability or the eternal spring for AI looks promising as one can apply either the direct approach when it makes sense or integrate the AI reasoning methods from yester years with data-driven representations generated by DCNNs.

AI will continue to make progress in sensor or data-rich applications such as self-driving cars, medicine, finance and education. Self-driving cars have cameras, 3D depth sensors, and geo-location devices, and most of us are already enjoying the benefits of increased automation and safety (blind spot warning, smart cruising, lane change warnings, automatic braking, etc.). We expect more autonomy leading to fully self-driving vehicles and trucks to fill our roads and highways in the foreseeable future.

Likewise, there is much opportunity for multi-modal data in medicine. Various imaging devices such as X-rays, CT, MRI and ultrasound spew tremendous amount of data that can be analyzed using AI techniques. Pathologists and radiologists will have many AI-based diagnostic tools to quickly triage the voluminous data they have to analyze. Another area where AI will be helpful is in predicting undesirable outcomes (such as, will the blood pressure of a patient under anesthesia reach dangerously low levels or will a patient in ICU suffer from delirium?)

Education is an area where the benefits of AI haven’t yet been fully explored. One can imagine having one teacher who is physically present and 30 virtual teachers in a classroom of 30 students, with each student learning from his/her own virtual teacher at a pace that is suitable to him or her. Individually-paced learning and education using AI agents will have profound impact on disseminating information to the next generation worldwide.

AI will have a significant impact on the semiconductor industry and is poised to redefine the business over the next decade. The complexity of chip development continues to increase, and chip makers are turning to AI to improve their design and manufacturing processes. The role of computing in AI cannot be overemphasized. The demands posed by DCNN training algorithms necessitated the development of powerful GPUs. Somewhat paradoxically, these days, the strength of an AI group is based on how many GPUs the group has access to. The semiconductor industry will continue to play a key role in the development of high-caliber GPU servers and workstations; it will also enable the operation of much smaller AI devices by providing smaller GPUs that enable computing at edge or mobile computing with less power. As a result, AI can be housed in IoT devices, household appliances, smart phones, tablets and other portable devices such as health-monitoring units.

Is AI safe and ethical?
Concerns have been raised regarding the vulnerability of AI system due to adversarial attacks, whether the AI applications are safe and ethical? Answers to these questions are not straightforward and need philosophical and moral perspectives. The vulnerability of AI systems to adversarial attacks is real and is being addressed by numerous research groups worldwide. Many defenses to adversarial attacks are being developed. To paraphrase the statistician George Box’s comments on models, “All defenses can and will be broken; some may outlive others.” This is an ongoing cat-and-mouse game not unlike the cyber threats we face every day.

As of now, it appears that systems that are robust to adversarial attacks do not perform at a level needed for deployment. This situation can be improved by embedding the AI software in secure hardware. The semiconductor industry will have a significant role to play in developing secure hardware for AI systems. Since the current performance achievements are largely driven by the availability of data, data bias can adversely affect the performance of the AI systems. A well-known example is the poor performance of some face recognition systems when applied to certain groups of individuals causing legitimate concerns about the fairness or ethical use of AI systems. Another concern is the privacy of data being used to design the AI systems. These concerns can be addressed by involving ethicists in the design of AI systems so that bias mitigation steps and careful privacy-protecting data collection protocols can be incorporated. Thus, there is an increased need to bring AI methods, as they are deployed in society, under a public policy framework, as has been done with other technological innovations in the past.

As we prepare for the integration of AI systems in our daily lives, we need to focus not only on the technical issues, but also on moral, ethical, societal and philosophical issues. This will require the involvement of all groups drawn from humanity so that a better, prosperous, harmonious and peaceful world can emerge and sustain.