Excellent, Elementary and Beyond
February has brought an interesting twist from the pages of Conan Doyle as well as delivered a glimpse of how the real brain can inspire not only a digital neural network but also a pure hardware chip design.
Machines still struggling with commonsense reasoning
“Excellent!” I cried. “Elementary,” said he.
Contrary to popular belief, Sherlock Holmes never said: “Elementary, dear Watson” and, in fact, used abductive rather than deductive reasoning in his thought processes. This manner of reasoning comes with some struggle for many people, and it once appeared to be impossible for AI.
Commonsense reasoning, or abductive reasoning, is the art of extracting unobvious information from incomplete data. An example of this can be seen in the knowledge that if you see a road sign with the distance in miles and brightly reading “Las Vegas”, then you’ll conclude that the photo was taken in the US.
At a basic level, this is bread and butter for the human brain, filtering out both the obvious and less obvious information from the surrounding world, both simple and entirely natural. Yet machines struggle with this task, even despite recent advances in the AI field.
To tackle the challenge, the Allen Institute for Artificial Intelligence, the University of California, Berkeley, and MIT-IBM Watson AI researchers have created a dataset called “Sherlock.” It contains a collection of over 100,000 images paired with clues and hints the observer could use to form a hypothesis.
The dataset is aimed to test an AI’s commonsense reasoning, which is a harsh challenge for machines, as it requires a great deal of knowledge, not excluding all the information people collect during their life experience.
More details regarding the dataset and the way it measures the commonsense of machine learning models can be found in this Arxiv paper.
Studying the brain can provide us with better ML hardware
Contrary to AI models, when the human brain learns something new, it naturally adapts instead of simply forgetting old information. This is possible due to the constant rewiring of the neural network hidden inside the human skull.
This stands in vast contrast to static hardware that, once manufactured, remains unchanged to the day of it being scrapped in favor of something new. Thus, making new connections and accommodating new information is basically impossible.
Researchers from Purdue University proposed a new architecture for chips that can be altered using electric pulses, which are used to modify the amount of hydrogen near the chip’s center. Depending on the amount of hydrogen, the chip can further serve as either a neuron or a synapse in an artificial brain.
More on the matter can be read in ScienceDaily.
Modern talking better than websites
Chatbots are digital assistants that are being used to automate multiple areas of customer service, be that in delivering requested information or supporting purchases.
Apparently, this manner of gathering information proves to be more effective than traditional web browsing. According to a study delivered by Botco.ai, 82% of consumers have used chatbots to engage with the company. Also, 64% of users are more likely to ask a chatbot than to scroll through the website in search of information.
More about the ways chatbots can be and are being used can be found in this botco.ai report.
Using Reinforcement Learning for finding treatments in high-risk states
Reinforcement learning is an approach where a machine learns from the experience gathered from its environment – be that a simulated city, the stock exchange, or basically anything else. This approach has been successfully applied in building RL agents that have beaten human champions in Chess, Go, and StarCraft III.
Researchers from Microsoft have found a way to use this approach to deliver a neural network that suggests medical treatments based on the current state of the patient. The aim of the RL agent is not to find the best cure, which can be extremely challenging, instead, it checks if the suggested treatment comes with a high risk of unexpected outcomes and warns against them.
This approach is aimed towards use in intensive care, where limited time is a key factor and physicians work under extreme pressure.
More about the model and the approach can be found in this research paper.