Perspectives on AI Heading into 2019
I know this is not breaking news, but I finally got around to reading the December 2018 edition of WIRED (and yes, I still enjoy the print version). It was an interesting issue, with a cover theme and articles on AI — Less Artificial, More Intelligent.
I have an avid interest in AI/ML, and have been working over the last year with a very promising start-up in the ML for IIoT space (Lecida — https://lecida.com/), and I thought the topic was very well represented and I would recommend it as a reading. There were a couple of key elements that I thought were worth sharing.
First is from the article/interview with Fei-Fei Li, “The Human in the Machine.”. She is an amazing force behind the many of the advancements in AI over the last decade, working from the intersection of neuroscience and computer science. From assistant professor at Princeton, to professor and Director of the AI Lab at Stanford, to chief AI scientist at Google Cloud, to the only woman invited to speak at the US House Committee on Science, Space, and Technology hearing on “Artificial Intelligence — With Great Power Comes Great Responsibility” last June, her accomplishments in the field are arguably second to none.
Fei-Fei is the mind and energy behind ImageNet, the AI toolset and database that has revolutionized deep learning and is behind many of the major image recognition tools. As a woman and immigrant from China, she is particularly sensitive to the fact that most of the current researchers in AI are white males.
As the last speaker at the hearing, her message was perhaps not what the committee expected, but in my opinion, what they needed to hear. She began with “There’s nothing artificial about AI. It’s inspired by people, it’s created by people, and — most importantly — it impacts people. It is a powerful tool that we are only beginning to understand, and that is a profound responsibility.”
While some very prominent, powerful (and male) tech leaders have been warning about a future where AI can become an existential threat to humans (a la Skynet from Terminator), Fei-Fei is much more concerned about near term questions on how AI will impact the way people work and live. As AI is still very much guided by human direction, and the majority of those humans are currently white males, the concern comes down to “bias in, bias out.” Even the most well-intentioned researchers and developers cannot avoid the inherent bias of their social upbringing. With AI being applied to so many processes that affect people, from applicant screening for hiring and loan applications to facial, image and voice recognition that is being used for “security” and other state and corporate policy monitoring, it is easy to see how there could be negative social impacts.
Examples of major failings or controversy already exist. Just last October Amazon revealed that they had scrapped an AI recruiting too that had been under development since 2015 because it had taught itself to be biased against women. This cannot be viewed as a fault of AI tech itself, but of faults within the algorithms and training data used. Even under Fei-Fei’s tenure at Google, the company came under major fire from not only the public but from their own employees related to Project Maven, a contract with the US Defense Department to provide AI to interpret video images that could be used to target drone strikes. Google subsequently announced it would not renew the Maven contract and also went on to hire Shannon Vallor as a consulting ethicist for Google Cloud. Li had championed Vallor’s appointment and quoted her at the Congressional hearing, saying “There are no independent machine values. Machine values are human values.”
Fei-Fei’s message is that we need to drive researchers to think as ethicists, who are guided by principle over profit, and informed by a diverse set of backgrounds. Part of the solution is to try to encourage and ensure diversity in the community of researchers and developers. The other is to ensure that companies developing AI solutions adopt clear ethical guidelines and enforce them. Let’s all do what we can to make this happen!
Second is from the article/interview with Karl Friston, “The Man Who Explained Everything.” Friston is considered one of the world’s leading neuroscientists and is the scientific director of University College London’s Functional Imaging Laboratory. His early work developed revolutionary tools for brain imaging that has enabled researchers worldwide to better understand and map the functions of the human brain. With over 1,000 academic papers published since 2000, he is also one of the most prolific and most cited scholars in history.
So, what do Friston and his work have to do with AI? As it turns out, maybe everything. Over the last decade, he has focused much of his efforts in developing the “free energy principle.” However, as stated in the WIRED article, the principle is “maddeningly difficult to understand.” Indeed, there is a Twitter account with over 3,000 followers dedicated to mocking its opacity.
Perhaps it is my background and interest in physics and mechanical engineering, perhaps the way my own brain processes information and seeks logic, but I believe I have at least a basic understanding of what Friston has come up with and will try to summarize it here.
The first concept relates to a fundamental definition of life. To grasp this concept, it is first useful to understand the concept of entropy. From the classical thermodynamics definition in physics, entropy is a state measurement of a system, the change in entropy of a system is determined by its initial and final states, and systems tend to progress in the direction of equilibrium, or increasing entropy. The second law of thermodynamics states that for isolated systems, entropy never decreases. In reality, though there are few real examples of truly isolated systems. During a change in state, entropy can be transferred to the environment surrounding the system, but as most state changes are spontaneous or irreversible, the overall effect is an increase in dispersion and an overall increase of entropy in the universe.
In statistical mechanics, entropy is a measure of the number of ways in which a system may be arranged, often taken to be a measure of “disorder” (the higher the entropy, the higher the disorder). From both thermodynamics and statistical mechanics views, systems then tend towards dispersion of energy with a resulting increase of entropy. This leads to the concept of free entropy — in thermodynamics, this is also referred to as thermodynamic potential. Free entropy (or analogously free energy) is then related to the change in entropy that occurs when a system changes from one state to another. As previously noted, most state changes are cannot be directly reversed without the application of additional energy.
Which brings us back to Friston’s free energy principle and its definition of life. Each division of life, from a single cell to a human brain, to a human, to a society of humans, can be considered a system. One of the distinguishing characteristics of life is that of self-preservation. Just as a cell protects itself from the attack of pathogens, the brain protects itself from damage, humans protect themselves from harm, and societies protect themselves from attack. So, if life is defined as a biological system with a “prime imperative” of self-preservation, how can we describe that in a logical way and apply mathematical models?
In Friston’s view, the release of entropy (energy) from the biological system as it moves from one state to the next (one point in time to another), decreases the viability of the system, as it inherently disperses energy in an irreversible process. So, it would make sense that the system would tend towards reducing the amount of free energy that is at risk. According to Friston, this can be done by minimizing surprise, or in more logical terms, reducing prediction error.
The human brain is an inordinately complex biological system, constantly processing inputs from billions of sensors and attempting to analyze the data, generate hypotheses, and plan actions. Friston uses the term “active inference” to help describe the way intelligent systems (e.g. humans or AI) minimize prediction error, or free energy, as they navigate their environment. When your brain makes a prediction that isn’t immediately validated by your sensors, free energy can be minimized in two ways: you can revise your prediction or hypothesis, admitting error and updating the model, or you can act to affect the model and make the prediction true.
If we compare this to the statistical mechanics model of entropy, in which the number of possible states that a system could be in is a determining factor in the amount of entropy, by minimizing the error, or possible end states, we reduce the potential loss. This provides a well-defined set of models and equations that have been derived not only in statistical mechanics but in the related definitions of entropy in quantum mechanics and information theory. As it turns out, these can be applied in developing an active inference engine that can be used in AI systems.
Most current AI approaches are based on neural network concepts, which are designed to mimic the way our brains process data but cannot (yet) match the way we think and make inferences out of incomplete data. Pattern recognition AI generally requires massive amounts of data, as well as human guidance to “tag” the images or other artifacts to tell the system what they are. These systems are generally good at doing certain things but can be brittle when presented with data they cannot recognize. Deep learning neural nets let the system make its own conclusions based on the processing of even larger amounts of data, but still generally require human validation and can also result in illogical (to humans) output. Reinforcement learning AI has recently shown great success at winning games, like Go, chess and Breakout (did you know that Steve Jobs took on the task of developing Breakout while at Atari?), and does not require human labeling of large amounts of data, but rather defines a desired outcome, e.g. winning a game. These systems learn by iterating through the process (game) and optimizing to achieve the desired outcome. However, when faced with a change in rules (real life), they must start over.
Friston and his followers believe that the free energy principle, application of active inference, and use of related entropy equations offer a new path forward for AI. I tend to agree.
In another very significant application of the free energy principle that Friston is passionate about, he believes that it can be applied in advancing mental health and brain disorder research and repairing neurological disorders such as schizophrenia.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
If you have made it this far I applaud you… I did not intend this article to be this long, but I got emotionally and intellectually engaged. If I lost you in my explanation of Friston’s free energy principle I hope that at least I was able to share Fei-Fei Li’s message. I would love to get into a discussion of Markov blankets and the free energy principle — perhaps in a future article.