Discussing AI, complexity, machine learning, and the future of the field with real-world examples. It explores the benefits and challenges of machine learning in a friendly and engaging way.

Look Around

by

Joseph

AI Primer

The history of artificial intelligence (AI) is filled with claims and controversy. What constitutes true AI? Why is logistic regression considered AI but not a governor on a steam turbine? We will explore these questions and more, as we journey into the world of AI. From the basics of machine learning to the latest developments in the field, we’ll dive deep into this fascinating subject and explore how AI can transform our world.

What comes next will hopefully be a long and fruitful journey exploring what AI is, how we can use it and what we can do to nudge the future in our favor.

Broad Overview

Taxonomies

AI comes in many shapes and sizes, but irrespective of the media or the substrate it lives in, there are a few common taxonomies that can help us think about it.

  1. Weak AI :
    (AKA Narrow AI)
    Weak AI is like a hammer, a tool designed to do one thing and one thing only. Do not let the moniker fool you, this does not mean that the AI’s performance is weak, it only means that this type of learning algorithm is performant in only one domain. In many situations human level performance is achievable or surmountable using only these so called weak AIs. Simple Dog or not Dog classifiers fall under this category.
  2. Strong AI:
    (AKA General AI)
    General AI is more a kin to a multitool. An AI is said to start generalizing with increases in functionality. We have recently started seeing models especially those trained for generative tasks, being coaxed into performing functions they were not designed to do. A nice example of this is when large text to image models were found to be able to perform style transfer. Its interesting that the ability to create, seems to underlie intelligence in more than just philosophical musings.Academically, strong AI is usually foreseen to be the point at which an agent has reached the same intellectual level of a human across all domains. There is no hard and fast definition as yet and there is very little that can be said to create one. At the end of the day humans come with a magnificent variety of capabilities and pre-dispositions.
  3. Superintelligence
    A term popularized by Nick Bostrum in the book also named Superintelligence, refers to agents with an ability to perform tasks currently unknown and unknowable by the simple human colony. Superintelligence entities are fabled to surpass human intelligence in every way, be it learning, reasoning or decision making.Of course, such technology would have very profound implications for humanity. It is a highly speculative subject and one of much debate among some of the leading AI experts. Arguments vary greatly, some extoling its virtues like they are praying to a deity and others regarding it as a demon climbing from the abyss.Personally, I enjoy thinking about these issues through the principles of computational irreducibility and computational equivalence uncovered so eloquently by Stephen Wolfram.

Applications

It is a profoundly erroneous truism, repeated by all copy-books and by eminent people when they are making speeches, that we should cultivate the habit of thinking of what we are doing. The precise opposite is the case. Civilization advances by extending the number of important operations which we can perform without thinking about them. Operations of thought are like cavalry charges in a battle — they are strictly limited in number, they require fresh horses, and must only be made at decisive moments. - Alfred North Whitehead

AI in real life has started to become the tide lifting the proverbial boats and allowing many to do more with less. Just think about the countless human lives saved by spam filters. AI techniques, have been infiltrating every day life for a relatively long time. Here are some of the more niche applications that show how deeply this technology is already being applied.

  • Emotion AI (AKA Affective Computing) is a branch of AI that focuses on trying to understand human emotions. It tries to recognize emotions in speech, text, facial expressions and other human based markers. Companies have been built around this, offering services ranging from simple adaptive emoticons to healthcare plugins that gauge the emotional well being of patients. One such example is Affectiva a company that claims their AI can understand human emotion.
  • Predictive Maintenance, is a field of AI that uses machine learning and other analytical techniques to predict the optimal maintenance cycle for equipment. This reduces or eliminates unscheduled down time and unexpected costly repairs. Predictive maintenance can be naturally coupled with another very interesting field, known as digital twinning, to great effect.
  • Personalized Education, has been an increasingly popular field. It focuses on learning which modes of instruction bring out the fastest rate of improvement in students’ test scores. The thinking behind this is, test often, iterate and re-visit. Treating students in the same way that you would a black-box problem to be optimized by twiddling external parameters. Not very surprisingly, this application is purported to be very promising and I am sure we will be seeing many AI powered learning systems crop up in the not too distant future.
  • Generative Design. More of a science than an art, it involves using algorithms to generate and explore many different design alternatives automatically. Generative techniques are slowly creeping into most modes of production. In general these models try and construct a 3D model, image or other media format by evaluating options based on a set of criteria, such as fit, weight, or cost. Some generative models work using techniques like genetic algorithms and optimization, or by running subtractive or additive loops. Other models work through the use of GANs (Generative Adversarial Networks) that pit what in effect are two neural networks against each other. The Ted Talk by Maurice Conti is one that has aged very well and demonstrates the power of these techniques.

Machine Learning

Machine learning is a subset of artificial intelligence that involves using algorithms to learn from data and make predictions or decisions without being explicitly programmed. The basic categories of machine learning are supervised, unsupervised and reinforcement learning.

  • Supervised learning is like learning to drive with a driving instructor telling you explicitly what each road artifact is and how to act in every situation.
  • Unsupervised learning is like learning to drive without an instructor and figuring things out on your own given pre set routes and driving experiences. In unsupervised learning you would need to intuit the meaning of road signs, which light colors mean stop and go but also the lack of relevant meaning in graffiti.
  • Reinforcement learning is like playing a driving simulator where you get better with each time you drive. In reinforcement learning you have a measure of success, as an example consider you get more points the closer you get to a pastizzi dive. You have sensory information, but you start knowing nothing about what that means or what your possible outputs do. You time-out, loose points beyond a threshold or crash and burn and get to respawn and try again until you learn how to maximize that measure.

This is by no means a discreet list. In fact most useful models use combinations of these concepts in their entirety or merged together in creative ways. As an introduction however, this will provide us with enough of a foundation to build upon.

Here are some real-world examples to help build some intuition around these models:

Table Plugin

Pros & Cons

Machine Learning Models offer advantages in both the qualitative and quantitative realms. Improved accuracy, automation, scalability, mass personalization and cost saving, to name a few. Benefits, that are real and worth considering seriously as they can result in a wonderful future for all of life on earth and beyond. However we must exert caution. The flip side, even baring the worst case consciousness ending scenarios, comes with over reliance and a lack of transparency and interpretability. These, when coupled with innate biases and an unequal access can spell disaster.

The hope with this and other game changing technologies is always that there are many more willing to direct the vector towards the globally beneficial rather than the nihilistic, the beautiful rather than the profane, and the creative rather than destructive.

Trends

For the most part, the main trends that we are seeing develop can fit into two main categories, cautious, and gung-ho.

On the cautious side of things we see an ever increasing interest in explainable AI, ensuring data privacy and understanding and mitigating algorithmic bias. The threat to many jobs and power structures has never been more real, and we do well to take it seriously. Some believe that the way to reduce the risks is by aggressively democratizing this technology. There has been an acceleration towards training larger and larger models on ever larger datasets and making them publicly available for free or for cheap.

At the time of writing, this trend is still in its infancy, but there have already been significant advancements. For instance, Chat GPT set a world record by achieving one million users in just five days. Additionally, stable diffusion and other generative models are breaking new ground in prompt generated art and 3d models. Not to be overlooked, Whisper, a speech-to-text model also developed by Open AI, is delivering best-in-class performance despite receiving little attention.

Only time will tell which approaches or trends will lead us into the next stages of our journey, but it is definitely a wonderful time to be alive.

Summation

Artificial intelligence is a field that involves algorithms learning from data to make predictions and decisions. With applications ranging from image classification to gaming, the potential benefits are vast. However, we must also be cautious of the risks and work to ensure transparency, privacy, and fairness.

Let’s embrace the positive trends while remaining vigilant and responsible, and continue to explore the potential of this transformative technology.

For The Interested Reader

AI for Social Good by Professor Yoshua Bengio
Anatomy of an AI system by Professor Kate Crawford
Living with Artificial Intelligence by Professor Stuart Russell