The current state of artificial intelligence in 2019: Breakthroughs in machine learning, natural language processing, games and knowledge graphs

Artificial Intelligence is one of the fastest-growing fields today. Tracking and evaluating the development of AI requires not only continuous attention but also the ability to analyze and assess from multiple dimensions. This is exactly what Nathan Benaich, founder of Air Street Capital and RAAIS, AI angel investor, and Ian Hogarth, visiting professor at UCL’s IIPP, are doing.

Also: AI Created through Neuroscience — ZDNet YouTube

In their June 28 report titled “The State of AI Report 2019,” Benaich and Hogarth comprehensively covered various aspects of artificial intelligence with 136 slides: technological breakthroughs and capabilities; the supply, demand, and concentration of talent in the field; large platforms currently and potentially driving AI innovation; funding and application areas; AI politics; and AI in China.

Benaich and Hogarth are more than just venture capitalists — both have rich AI backgrounds, having been involved in numerous AI projects from research to startups. Additionally, they draw upon the expertise of renowned figures such as Google AI researcher and Keras deep learning framework lead François Chollet, AI thought leader and venture capitalist Kai-Fu Lee, and Facebook AI researcher Sebastian Riedel.

This collaborative work condenses a wealth of expertise, experience, and knowledge. After discovering and reading the report, we reached out to Benaich for an in-depth Q&A. We distilled the report and his insights into a two-part series: first exploring technological breakthroughs and capabilities, then their impact and the political implications of AI.

---

Interpreting Artificial Intelligence

If you are interested in AI, this is probably not the first AI report you have encountered. Many are familiar with FirstMark’s Data and AI Landscape (authored by Matt Turck and Lisa Xu) and MMC Ventures’ State of AI: Divergence. These three reports were updated almost simultaneously. While this can cause some confusion due to obvious overlaps, there are also differences in content, approach, and format.

FirstMark’s report is more detailed, covering participants across various fields from data infrastructure to AI. It has evolved from a big data landscape to a data and AI landscape. As we noted earlier, the evolution from big data to AI is a natural process. MMC Ventures offers a different perspective — it is more abstract and may be better suited for executives. The three reports present different viewpoints and are not in opposition — each has its own strengths.

First, we asked Benaich why they did this: Why share undoubtedly valuable knowledge, put in extra effort, and seemingly for free?

Benaich stated they believe AI will become a multiplier for technological progress in our increasingly digital and data-driven world. This is because everything around us today, from culture to consumer goods, is a product of intelligence:

“We believe people increasingly need accessible, detailed, and accurate information about the current state of AI in multiple domains (research, industry, talent, politics, and China). Our goal in publishing the report is to drive in-depth discussions about AI development and its future impact.”

This report meets the goal Benaich set in his reply. The first 40 pages, presented as slides, focus on progress in AI research — technological breakthroughs and capabilities. Key areas include reinforcement learning, gaming applications and future directions, natural language processing breakthroughs, deep learning in medicine, and AutoML.

Also: Will Alien AI Visit Us? — ZDNet YouTube

---

Reinforcement Learning, Gaming, and Real-World Learning

Reinforcement learning (RL) is a branch of machine learning that has attracted widespread research attention over the past decade. Benaich and Hogarth define it as:

“A software agent that learns goal-directed behavior through repeated trial-and-error in an environment, which provides rewards or penalties based on the actions (called policies) the agent takes to achieve the goal.”

Much of the progress in RL has been tied to training AI to play games, reaching or surpassing human-level performance. StarCraft II, Quake III Arena, and Montezuma’s Revenge are just a few examples.

More important than the sensationalism of “AI beating humans” is how RL achieves these results: game-driven learning, the combination of simulation and the real world, and curiosity-driven exploration. Can we train AI by playing games?

In childhood, we learn and practice complex skills and behaviors through low-risk methods like play. Researchers have leveraged the concept of supervised gameplay to teach robots control skills, enabling them to better handle disturbances compared to training with expert demonstrations.

In RL, agents learn tasks through repeated trials. They must balance exploration (trying new behaviors) and exploitation (repeating effective behaviors). In the real world, rewards are hard to explicitly encode. A viable solution is to have the RL agent store observations of its environment in memory and reward it when it encounters observations “not in memory.”

The perspectives cited in the report seem equally excellent and natural. So, could these views represent the future direction of AI development? Benaich notes that games are fertile ground for training, evaluating, and improving various learning algorithms, but he also raised some doubts:

“Data generated in virtual environments is often cheaper and more accessible, which is beneficial for experimentation. Furthermore, the complexity of game environments can be adjusted during model development according to experimental goals. However, most games do not accurately simulate the real world and its rich subtleties. This means they are a good starting point but not the end.”

Also: Understanding AI in the Supply Chain — ZDNet YouTube

---

Natural Language Processing and Commonsense Reasoning

As Benaich and Hogarth pointed out, this has been a landmark year for NLP: Google AI’s BERT and Transformer, the Allen Institute’s ELMo, OpenAI’s Transformer, Ruder and Howard’s ULMFiT, and Microsoft’s MT-DNN demonstrated that pre-trained language models can significantly improve performance on various NLP tasks.

Learning high-level and low-level features through pre-trained models has revolutionized computer vision, thanks mainly to ImageNet — a dataset with over 20,000 categories. A typical category, such as “balloon” or “strawberry,” contains hundreds of annotated images.

Since 2010, the ImageNet project has annually hosted the ImageNet Large Scale Visual Recognition Challenge (ILSVRC), where competing software programs classify and detect objects and scenes. The challenge uses a “pruned” list of 1,000 non-overlapping categories and has been a driving force behind the progressive refinement of computer vision technology.

ImageNet.jpg
ImageNet is a carefully curated computer vision training dataset that has driven state-of-the-art advancements. Image source: Nvidia

Last year, similar breakthroughs were made in pre-training language models on large text corpora to learn high- and low-level language features. Unlike ImageNet, these language models are typically trained on large amounts of publicly available, unlabeled text from the web.

This approach can be further scaled to yield benefits in NLP tasks and unlock many new commercial applications, just as transfer learning from ImageNet pushed computer vision into more industrial uses.

Benaich and Hogarth highlighted the GLUE benchmark, which provides a single metric for evaluating NLP systems across a range of tasks involving logic, commonsense understanding, and lexical semantics.

To demonstrate the rapid pace of NLP progress, they added that the state-of-the-art GLUE score rose from 69 to 88 points in 13 months, while the human baseline is 87. Progress exceeded expectations, and a new benchmark, SuperGLUE, has been introduced.

However, language is special in human cognition. It is closely related to commonsense reasoning, which has also advanced. Recently, Salesforce raised the technical bar by 10%.

Researchers at New York University showed that neural models can gain simple commonsense abilities and reason about previously unseen events through generative training on reasoning knowledge datasets. This approach extends efforts such as the Cyc knowledge base project, which began in the 1980s and is the longest-running AI project in the world.

Also: Trends in Machine Learning and AI — ZDNet YouTube

---

The Way Forward: Combining Deep Learning and Domain Knowledge?

We asked Benaich about the approach of combining deep learning with domain knowledge for NLP, as experts like David Talbot from Yandex see this as promising. Benaich agreed, saying combining deep learning with domain knowledge is a fruitful exploration:

“Especially when AI projects aim to solve real-world problems rather than building a general intelligent agent expected to learn to solve ‘whiteboard problems.’ Domain knowledge can effectively help deep learning systems guide their understanding of problems by encoding primitives, rather than forcing models to learn these problems from scratch using (potentially expensive and scarce) data.”

Benaich also emphasized the importance of knowledge graphs for commonsense reasoning in NLP tasks. Cyc is a famous knowledge graph, or knowledge base (the original term). However, he added that commonsense reasoning is unlikely to be solved solely through text as the only modality.

Other highlights from the report include advances in data privacy, such as federated learning, Google’s TensorFlow Privacy, and Dropout Labs’ TF-Encrypted, as well as numerous use cases of deep learning in medicine — some seemingly science-fictional feats like decoding EEG brain waves and restoring motor control in disabled patients.

To interpret all content in the report—such as AutoML, GANs, and progress in speech synthesis deepfakes (which we predicted years ago)—requires very deep study. Simply browsing the report takes time, but it is certainly full of valuable insights.