In this special guest feature, Yonatan Geifman, CEO & co-founder of Deci, discusses how automated machine learning (or AutoML) can “democratize data science” by gradually implementing different levels of automation. Yonatan co-founded Deci after completing his PhD in computer science at the Technion-Israel Institute of Technology. His research focused on making Deep Neural Networks (DNNs) more applicable for mission-critical tasks. During his studies, Yonatan was also a member of Google AI’s MorphNet team. His research has been published and presented at leading conferences across the globe, including the Conference on Neural Information Processing Systems (NeurIPS), the International Conference on Machine Learning (ICML), and the International Conference on Learning Representations (ICLR).
Automated machine learning, or AutoML, has generated plenty of excitement as a pathway to “democratizing data science,” and has also encountered its fair share of skepticism from data science’s gatekeepers.
Complicating the conversation even further is that there is no standard definition of AutoML, which can make the debate incredibly difficult to follow, even for those well-versed.
The goal is straightforward enough: By embracing a new AI mindset and automating key elements of algorithm design, AutoML can make machine learning more accessible to users of various stripes, including individuals, small startups, and large enterprises. More specifically, AutoML can make deep learning, machine learning’s more complex subset, accessible to data scientists, despite its more complicated nature.
That’s an attractive value proposition – one that promises unprecedented efficiency, cost savings, and revenue creation opportunities. It also helps explain why the AutoML market will grow astronomically in the coming years. According to a forecast by P&S Intelligence, the global AutoML market is on pace to grow from $269.6 million in 2019 to more than $14.5 billion by 2030 – advancing at a CAGR of over 40%.
But what will fall under the umbrella of this rapidly surging market? There is no neat-and-simple answer. Instead, to better understand AutoML, we should examine it as a spectrum, and not as black and white as either fully autonomous or fully manual. Consider the automotive industry, where autonomy is split into different levels, ranging from Level 1 (driver assistance technology) to Level 5 (fully self-driving cars, which remain a far-off prospect). Thinking of AutoML in this way serves as a useful reminder – building automated AI models isn’t an all-or-nothing proposition. Here’s a closer look at the graduated scale that is redefining the AI pipeline.
Level 0: No Automation
By definition, fully manual deep learning processes rely on the skillsets of data scientists and other specialists to carry out key processes, including programming neural networks, handling data, conducting architecture searches, and so on.
The level of skill required to execute these tasks is formidable, which helps explain why deep learning (along with the expensive talent required to implement it) has proven elusive to many organizations.
Level 1: High-Level DL Frameworks
While manually implementing DL from scratch poses many obstacles, the accumulated work of DL programmers and data scientists has led to the creation of high-level frameworks like Caffe, TensorFlow, and PyTorch, which provide DL models and pipelines for users to write their own networks and more.
Just as Level 1 autonomous driving, which encompasses Advanced Driver Assistance Systems (ADAS), has delivered substantial benefits to ordinary drivers while still being a far way off from full autonomy, these high-level DL libraries are making DL pipelines simpler and more efficient. Implementation still requires a high degree of expertise in programming and DL, but it doesn’t require PhD-level data science expertise, making it much more accessible to many organizations.
Level 2: Solving Set Tasks
Leveraging pre-trained models alongside transfer learning yields an even more automated process.
This level of automation builds upon the availability of trained models like open source repositories and labeled data, which are then fine-tuned to solve a given problem. Again, this stage does not obviate the need for data expertise; it relies upon engineers to pre-process data and tweak the model according to the task at hand.
Level 3: AutoML with NAS
Neural architecture search (NAS) is an emerging field in which algorithms scan thousands of available AI models to yield an appropriate algorithm for a given task. Put another way, AI is brought to bear to build even better AI.
While NAS has been the exclusive preserve of Big Tech companies like Google, Facebook, and major academic institutions like Stanford University, further innovation in the field will beget greater scalability and affordability, opening a range of highly valuable applications – including more sophisticated analysis of medical images, for instance.
Level 4: Full Automation
When the deep learning pipeline is fully automated, meta-models will set the parameters needed for a given task. Given training data, the meta-model can invent the right architectures needed for the task at hand, as well as offer prior knowledge on the architecture hyper-parameters.
Although full automation is still several years off, working toward meta-models will deliver vital gains in efficiency and sophistication even at the lower levels of autonomy (much as innovators working to ultimately develop self-driving cars have already rolled out improvements to automotive technology). Because each level of autonomy builds on the other, NAS models will play an important role in finding the right meta-models to run in each use case.
Making Deep Learning Accessible to a broader audience
Precisely how long it takes for full automation to become feasible remains to be seen. But with more accessible deep learning as the lodestar, AI developers are embracing a new mindset and paving the way to a future in which many of the most cumbersome and costly tasks of the AI lifecycle will become obsolete – unleashing a new generation of AI progress.
At the end of the day, regardless of how we define AutoML, the goal is to make deep learning more accessible to those who need it. This means simplifying its use so that any company, big or small, can productize AI.
Sign up for the free insideBIGDATA newsletter.
Join us on Twitter: @InsideBigData1 – https://twitter.com/InsideBigData1