The talk of Artificial Intelligence (AI) seems to be everywhere these days: in businesses across all industries and across all types of media we constantly hear about the latest promises of AI. The prospect of self-driving cars comes with the vision that we can be productive (or relax) while driving from A to B. Increasingly intelligent machines and robots promise to make manufacturing processes ever more efficient or provide AI-based diagnostics in health care.
New, AI-based tools may certainly open the doors to new, exciting opportunities across all industries and across all business functions. Yet, based on my work over the past two years, including interviews with dozens of AI experts and senior executives for several projects, I found that there is still a very limited understanding of how AI works and its potential implications on a business. Similarly, there are not even clear definitions for commonly used terms such as Data Science or Big Data.
Since the opportunities seem limitless but also given the current hype, it is key for every business leader to understand what AI can and cannot do for their business in the more immediate future. In this context, we want to look at the notions of ‘Top-down AI’ vs. ‘Bottom-up AI’, understand the differences, and explore which type of AI may be most appropriate to support business growth or create business process efficiencies. Along these lines, I have found that ‘Top-down AI’ often holds considerably more promise for short-term business improvements. I will explain why and share relevant examples.
Top-down AI vs. Bottom-up AI
The notions of ‘Top-down AI’ and ‘Bottom-up AI’ go back to Alan Turing and the ‘Manifesto’ he wrote in the late 1940s. Generally, when facing a problem or equation, you can find a solution in one of two ways. You can either process the problem in your head by matching it to previous information or you can let the equation present its variables to you without adding any context. The first describes a top-down approach, used by those who prefer applying previous knowledge to educate their perception. The opposite approach, bottom-up, is based on the belief that development should part from a stimulus. In other words, what drives our perception is that which we sense.
In simple terms and in the context of AI, it is probably easiest to imagine ‘Top-down AI’ to be based on a decision tree. For example, a call center chat bot is based on a defined set of options and, depending on the answers, it guides the caller through a tree of options. What we typically refer to as AI these days – for applications such as self-driving cars or diagnostic systems in health care – would be defined as ‘Bottom-up AI’ and is based on machine learning (ML) or deep learning (DL). These are applications of AI that provide systems with the ability to automatically learn and improve from experience without being explicitly programmed.
Generally, both Top-down and Bottom-up AI can offer good results. The question is rather when to use the one and when to use the other. You should always be mindful of the problem you are trying to solve and the data you have available when deciding what tool or methodology to apply. A great example in this regard was shared with me by an expert who formerly worked for one of the leading hard disk manufacturers. They were running a cloud-based, big data research program. They decided to set-up their own Hadoop cluster and then spent a lot of time trying to figure out how to tune the system. Apparently, this process was extremely cumbersome, so in parallel they also tried a top-down approach based on genetic algorithms. It turned out that the results of top-down vs. bottom-up were very comparable, yet the top-down approach was much more efficient.
The application of bottom-up AI may encounter 3 typical issues: (1) the necessary data may not be available, (2) the necessary data may be too expensive to generate, and (3) edge cases or unknowns. An example for the latter is autonomous cars and the rare scenarios that bottom-up AI cannot handle. For example, when your autonomous car encounters a snowplow for the first time it will simply not know what to do and you better switch over to manual or stop. To counter this issue, developers are trying to capture as many miles as possible to increase learning (weather, traffic scenarios, …).
Data availability and cost are the key drivers in deciding whether to go for top-down vs. bottom-up AI. Bottom-up systems require lots of (often costly) data, while top-down AI is data-efficient which may give it an edge over data-hungry bottom-up AI. A good measure for whether bottom-up AI makes sense is the ratio ‘number of samples / number of features’. If this ratio is high, it is better to apply non-parametric models (bottom-up AI). Once the ratio gets very small, you want to go for parametric methods (top-down AI).
A good example of the application of bottom-up AI goes back to the 1990’s when post offices introduced machines that could sort envelopes based on zip codes. To learn this capability, it was only necessary to learn the 10 digits. Assuming you would have 1 million training samples for the 10 digits, the ratio is 1,000,000 / 10 = 100,000 – a fine machine learning task.
For a clinical trial, however, a typical data set can easily cost around $20k. If you want data for 500 people, you already need to come up with $10m. But then you may also need to sequence the data, etc., so for a Phase III trial, you are typically paying $30-50k per data set. When you look at the data set for a typical patient then you can easily find 50,000 gene expressions, plus imaging data, lab data, and demographics data … Now the ratio ‘number of samples(=patients) / number of features’ becomes say 500 / 50,000 = 1 / 100. Bottom line in this case: you can easily spend tens of millions of dollars on data, but you would not learn much using bottom-up AI.
In terms of data cost, several experts I interviewed highlighted that even companies that are not necessarily ‘cash deprived’ prefer data efficient and, hence, more cost-efficient methods, also because they can be better understood (more on this aspect below).
In this context, it needs to be highlighted that a major fraction of data cost is often associated with data cleansing and labeling. One of my interviewees told me that AI professor Michael Jordan from Berkeley was once asked: ‘If someone would give you $1b – what would you use it for?’. His answer was: ‘I would hire more people to label data.’ This quote clearly illustrates the fact that most of the time in bottom-up AI is used to clean and label data, not for coding.
As we will discuss below, in addition to considerations on data availability and cost, properly framing the problem is key. Only once you have described your objectives, anticipated cost and effort required, and the expected outcome, you can make the best possible decision on whether to go for top-down or bottom-up AI (or maybe even an approach that is not AI-based).
‘Bottom-up AI’ = ‘Black Box AI’
An often-ignored issue with bottom-up AI systems is the fact that the resulting algorithms are inherently a ‘black box’ for any human operator. In a business context this is often a major issue. For businesses of any kind, the key theme in most decision-making processes is ‘we need to understand it as well as possible’.
As a result, trust in bottom-up AI is often limited – even if the results are positive. This challenge is already relevant during the development stage. A data scientist who had worked in the area of bottom-up AI for quite some time told me: ‘Too many times I could not make sense of the outcomes – there was too much noise and no clear answers. I tried different methods, I looked for shortcuts, but with only very limited success.’ One of my interviewees from the finance sector shared an example where a fund would commit 1% of the fund value to be managed by a bottom-up AI system. Even when the system performed very successfully, the managers were still scared rather than thrilled – saying: ‘But you can’t explain this to me …’. Apparently, many in the finance sector still remember a ‘flash crash’ example from 2012 where a fund did not understand their model as well as they thought and lost 60% of their value in just one day – so people in this sector remain cautious.
I heard similar examples and opinions from across many sectors. ‘Black boxes’ can frustrate the developers and are mostly not considered acceptable in a business context. Instead, there is a strong preference that a human operator understands what’s going on, so that he / she can follow the decision-making process – which is a key value proposition of the top-down approach. This is particularly true in strongly regulated environments, such as Finance and Health Care. Here, transparency of decision-making is typically a must which often excludes bottom-up AI. At the same time, this opens the door for data-efficient, top-down AI which typically provides at least a ‘grey box’ or even a ‘white box’. Hence, a human operator can very much make sense of the results, can understand the decision-making process, and can also interfere, if necessary.
‘Top-down AI’ = ‘Data-efficient AI’
Apart from the value proposition of ‘making sense’, another key advantage of top-down AI is that common-sense reasoning, decision trees, expert systems, or statistical methods (such as Gaussian regressions) can be ‘baked’ into its architecture. Hence, ‘Top-down AI’ systems are also ‘data-efficient’ since they require less data to make decisions with confidence. In addition, even if a top-down approach is algorithmically difficult and you run it in the cloud, the required computing performance is often equal to that of a laptop. The key prerequisite for data-efficient AI, however, is the availability of high-quality data.
In the series of interviews that I conducted various experts shared the opinion that we will hear much more about data-efficient AI in the coming years. A former VP at Google stated that he was expecting that in the immediate future many will be adjusting their expectations and understand that not everything is a deep learning or bottom-up AI problem. Another expert stated: ‘I believe that we are slowlymoving beyond the hype phase when it comes to bottom-up AI. After all, decision trees can still do wonderful things today.’
In terms of practical applications of top-down AI, the Internet-of-Things (IoT) was frequently mentioned. IoT devices are typically tiny (and hence very limited in terms of processing power), yet in many cases you want to incorporate the capability to learn. Here, data-efficient AI can help to make them ‘smarter’. As an example, an investor and advisor of a personal medical devices start-up mentioned a ‘data-efficient AI’-powered Fitbit-type device that could provide users with specific recommendations, rather than only indicating whether or not the 10,000 steps per day goal has been achieved.
Looking at the characteristics and capabilities of both top-down AI and bottom-up AI, it may not come as a surprise that many experts mentioned the huge potential of hybrid systems. These combine key business logic elements of top-down AI with the machine or deep learning capabilities of bottom-up AI to facilitate decision-making. One expert described it as ‘combining the capabilities of the cloud with the capabilities of your laptop’.
A very ‘tangible’ example of a hybrid system is the combined use of decision trees and chatbots at call centers. Today’s chatbots, while using bottom-up machine learning (e.g. for speech recognition), are far from being bottom-up intelligent – we still need to tell them what to do when using top-down logic. Applying top-down AI (as simple as a decision tree) may result, for example, in fewer escalations to supervisors. Today, any case with a bit of complexity is often escalated and then, ultimately, customer wait time increases, along with customer frustration.
Another example from the commercial space that one interviewee shared was from a shipping company that, amongst other things, often had to ship thousands of pipes. Counting all the pipes manually would be an extremely time-consuming effort. A simple answer to the problem seemed to be to simply take photos and use bottom-up AI to develop an algorithm that counts the pipes. The problem, however, was that there were no standard sizes and dimensions for the pipes. So even if you captured all shapes and size today, there might be new types of pipes tomorrow. Also, obtaining the necessary training data would have been relatively expensive. Hence, the company successfully chose a hybrid approach, combining the capabilities of the bottom-up algorithm with some top-down rules and logic.
Finally, a more exotic example would be driving a car on Mars which we cannot drive remotely controlled from Earth due to the communication delay. While we have a lot of data for autonomous driving cars on Earth, conditions on Mars are very different. We can now overlay different other concepts (e.g., low gravity environment or no atmospheric pressure) top-down on the bottom-up models created for terrestrial cars to enable autonomous driving on Mars.
Humans in the Loop
Both top-down and bottom-up AI systems can also benefit from a human in the loop. One of my interviewees shared an example of a paper that was published a few years ago with the claim that a machine learning-based AI could diagnose a certain cancer with 99.9% accuracy, , while doctors only detected about 93% of the cases under the same conditions. Victory was declared! However, later someone noticed that in all the images for the cancer cases that were used to train the algorithm, a tube was in the image somewhere – which was essentially a property of all the images with the cancer. It became clear that the algorithm had not really learned much more than to detect the artifact ‘tube’.
Another interviewee shared a story that happened at one of the leading hard disk manufacturers. They were trying to figure out issues in the manufacturing process applying both bottom-up and top-down AI methods. The various parts for the hard disks were procured from different suppliers. Since a certain part of supplier A may not work well with a part from supplier B, potential issues needed to be figured out using a bottom-up approach. The key objective: quickly identify the issue, stop the factory, fix the issue, continue production. At some point there were a lot of failures with a certain type of drive. It was initially tried to figure out the issue bottom-up, then top-down (using genetic algorithms), but without success. Ultimately it was found that there seemed to be an issue with the screws from two of the five suppliers, but the root cause of the issue remained a mystery. It was then found that a particular factory in China received screws only from the two identified suppliers. Realizing the need to dig even deeper, a team went to the factory in China and met with the head of supply chain, the SVP manufacturing, and the director of the factory. It ultimately turned out that the screws had nothing to do with the issue. The actual issue was that one specific worker always put on a specific part, a spinner, upside down. This could only be found through human observation. Bottom line: despite the application of AI, human(s) in the loop can make a unique difference!
Applying AI in your Business – Prospects & Limitations
Whatever AI-based use case or application you may have in mind for your business, first and foremost, you need to keep in mind that ‘well defined is half solved’. Once you have properly defined the (business) issue or problem you want to solve, you can decide whether a top-down AI and/or bottom-up AI solution makes sense. And there may be cases where a much more ‘low-tech’ approach yields a higher return on your investment.
The key drivers in the top-down AI vs. bottom-up AI decision process are certainly data availability and cost. In cases where data is cheap (e.g., IoT, advertising) bottom-up AI may often have an edge. If the available budget is not sufficient to allow for a bottom-up AI approach, you will likely move towards the data-efficient top-down approach. A practical example in this regard was shared by a former Google data scientist: ‘We did a lot of A-B Testing and we used a 3rd party A-B Testing platform for website optimization. Of course, when your customer is Amazon, you will have literally billions of observations. However, if you customer is Joe Simple who wants to decide whether a button should be black or blue, you will have only a very limited set of data. Hence, you will have to try and help Joe with a top-down approach.’
According to an AI start-up advisor and investor I talked to, larger commercial companies are increasingly (re-)discovering the value of the more traditional, top-down methods. This notion was confirmed by an independent data scientist I interviewed who stated that ‘in 90% of cases you will be more successful when presenting more traditional methods vs. deep learning-based approaches to customers because most customers want to understand the results or recommendations rather than having them come out of a black box.’
In summary, apart from the above-mentioned call center and IoT applications, three sectors may potentially benefit the most from the application of top-down AI:
- Health Care (e.g., diagnostics, assessment of insurance claims)
- Finance (e.g., security / trading; for example: ‘stop automated trade’, if outside ‘nominal’ range)
- Automotive (esp. self-driving cars; for example: ‘stop driving’ if something unknown happens)
In Health Care, apart from regulatory and data cost issues, limited data availability can also be a driver for top-down AI. For example, if you try to detect the outbreak of a new virus and you can count ~30 events per month, you need to step back and conclude that bottom-up is not the way to go.
The same is typically true for security-related problems in Finance. Here, the top-down approach is essentially your only option. In fact, providing bottom-up AI outputs to operators is pretty much a taboo since the operator needs to be able to correlate the data. In other words: ‘black boxes’ are not acceptable (as discussed above).
Whether the current enthusiasm regarding self-driving cars will prove to be warranted remains to be seen in the coming years. My interviewees expressed much more skepticism in this regard than I had expected.
In summary , it seems likely that data-efficient, top-down AI will find increased application wherever data is very limited or extremely expensive to collect. Of course, on the other end of the spectrum, bottom-up AI will continue to thrive in areas where top-down AI is less efficient (or simply makes no sense), such as speech detection or image analysis. In between, we are likely going to see an increase of hybrid applications and applications that enhance overall results by adding the human-in-the-human to the overall system.
Peter Eckart is an independent consultant and Managing Partner of Simplifier Inc., based in Palo Alto, CA. A rocket-scientist-turned-top-management-consultant-turned-senior-executive, Peter has been reporting into or advising CxO-level management of major corporations, SMBs, and start-ups for over 15 years. Being passionate about simplicity and creating impact, he has over 20 years of experience in business strategy, product development / innovation, and operations across numerous sectors (focus on High-Tech, Retail, Aerospace), functions, and geographies (US, Europe, Middle East). Peter holds a PhD in space engineering and a Masters in aerospace engineering from Technical University of Munich, Germany. He has been executing numerous projects for 10EQS since 2016.