Today's Deep Learning "AI" Technologies Is Machine Learning Not Magic

By Kimberly Cook | Dec 5, 2018 | 10608 Views

To the general public, today's "AI" technologies are nothing short of magic. Algorithms that can eerily understand video, images, speech, and text, translate between languages with uncanny accuracy, drive cars, play video games, find cancer and even best humans at complex strategy games by developing novel moves that no human had ever devised. Groundbreaking new milestones are crossed almost daily. Computer science programs are flooded with students eager to become AI experts and companies can't hire enough AI programmers. It would seem the era of AI has truly arrived. In reality, today's AI algorithms are nothing more than traditional machine learning algorithms. Even the most powerful systems are more artistry than science, require large amounts of carefully curated data from which they still inherit significant biases, are largely unable to generalize beyond their training domain, are black boxes that even their creators don't fully understand and are for all intents and purposes little more advanced in terms of construction than the tools they replace.

In the half-decade since the deep learning renaissance leaped into the public consciousness with groundbreaking feats that had hereto eluded all previous efforts, AI has come to hold an almost mythical reverence. Academics hold "intelligent systems" conferences, marketing brochures tout an AI solution to every problem, media coverage breathlessly proclaims every advance the end of human superiority and the general public worries about the AI revolution. Not a week goes by that an academic paper or corporate research blog doesn't document yet another algorithmic or application breakthrough. Major companies perform wholesale replacements for their existing algorithmic infrastructure with new deep learning systems.

It seems the future is here.
As powerful and capable as current deep learning systems are, they are still only rote pattern extractors. A computer vision system can take a pile of cat photographs and "learn" to recognize cats. Transfer learning can be used to teach it to recognize dogs with a much smaller pile of training images. However, the underlying algorithm is not reasoning about what it is seeing, it is merely breaking the image into distinct colors, patterns, and shapes and associating specific visual cues with a textual label. It cannot generalize from what it sees to autonomously expand its vocabulary to new mammals or understand the concept of "fur" or "paws" even as it associates a particular covering texture and four rectangularly distributed shapes with the images it has seen.

This lack of higher order reasoning is the reason that AI algorithms are so easily fooled. Make a few subtle changes to an image and you can easily make an image of a dog appear as an image of a bus or a stop sign be returned as background noise. It is the reason that algorithms are so easily skewed by biased input data, blindly "learning" all of the human biases and faults that we turned to AI to rid ourselves of. Though, at least in AI form it is possible to quantitatively document and corrects those biases if we are creative enough to see them.

The bigger problem is that creating the deep learning models of today is far more artistry than science. Contrary to popular perception, building a neural network isn't as simple as clicking s "Build Model" button and point the system to a pile of annotated imagery. There are myriad decisions to be made and parameters to be tuned. From how the input data is prepared, balanced and ordered, to the model construction and components it relies upon to the tuning of that model, building a modern deep learning model is a cross between brute force experimentation and the fine artistry of experience and guesswork guiding the process towards expected starting points.

Making matters worse, the underlying toolkits and algorithms are moving so fast that a workflow that yields useful results this month may break next month with a new release. Recommended approaches and algorithms change with dizzying speed. Convene a panel of ten seasoned experts and you may get ten wildly different approaches. Even looking to AI thought leaders can yield confusing and contradictory advice as experts clash over the best approaches. Meanwhile, AI incubators churn out papers that rely on approaches that the incubators' parent companies have publicly depreciated and begun to remove from their software frameworks.

In short, the field is advancing so fast that even the companies pioneering it can no longer keep up. The research-centric nascent state of AI means the field is still being invented even as it is being put into production service.

AI algorithms are beginning to make life-or-death decisions for us even as their creators have little idea how they are making those decisions or where they might go wrong.

The democratization of AI has made it easy to create models, but it is still hard to create good models.

Building a basic deep learning model in a framework like TensorFlow is fairly straightforward. The copious example code and tutorials make it possible for a reasonably skilled programmer to fairly rapidly builds a basic model. The problem is that there is still an exponential leap between this basic "hello world" model and the accuracy needed of a production system. At the most advanced end, skill is less important than experience in knowing just which components and parameters to chose for each dataset and application problem.

Cloud companies are solving this by offering libraries of prebuilt state-of-the-art models that can be used right out of the box, offer the best general domain accuracy available today and are regularly updated. Enabling customization of these models, tools like AutoML leverage transfer learning to enable non-technical users to rapidly construct their own customer-specific models without having to understand any of the underlying deep learning principles. For those customers with their own deep learning teams, libraries of prebuilt models and components can be used in plug-and-play fashion to build state-of-the-art bespoke systems.

Putting this all together, stepping back from all of the hype and hyperbole, building a modern deep learning pipeline is in many ways little different than building any kind of traditional machine learning system. Data cleaning, sample preparation, sample ordering, algorithm selection, and parameter tuning are all the bread and butter of machine learning, long predating the deep learning revolution. Deep learning approaches are capable of achieving far greater accuracy than previous approaches, but they are ultimately merely pattern extraction systems. They identify and codify basic patterns, rather than generalize from their inputs into an abstract mental model of the world around them.

At the end of the day, the deep learning systems of today are less "AI" than fancy pattern extractors. Like any machine learning system, they are able to blindly identify the underlying patterns in their training data and apply those patterns as-is to future data. They cannot reason about their input data or generalize to higher order abstractions that would allow them to more completely and robustly understand their data. In short, while they can perform impressive feats, deep learning systems are still extraordinarily limited, with brittleness that can manifest in highly unexpected ways.

After all, the "AI" of today's deep learning revolution is still just machine learning, not magic.

Source: HOB