What technologies are used in Artificial Intelligence systems?
- December 3, 2024
- 3 mins
In the media, likely due to the significant advancements made in recent years thanks to these techniques, Artificial Intelligence is often associated with Machine Learning and its most recent branch, Deep Learning. However, as its formal definition shows, AI encompasses a much broader range of techniques, of which Machine Learning is indeed a part.
Designing an Artificial Intelligence system actually involves combining these various techniques to create a system capable of performing one or more of the three main tasks that constitute the human decision-making process: Perceive, Understand, and Decide.
From a computational perspective, ‘Perceive’ corresponds to data collection and storage. ‘Understand’ involves structuring information from this data to make it intelligible. Finally, ‘Decide’ corresponds to using this information to find the best possible answer to a given problem.
In this article, we will explore which techniques are used at each of these stages in designing AI-based solutions.
Collect and Store: The Essential Role of IoT and Big Data
The perception task involves making a machine collect (e.g., via IoT technologies) and store data in a format usable for future applications (i.e., all storage technologies). For specialists, this is not strictly considered ‘AI’ since there is no high-value task involved here. Nevertheless, it can be included in a definition of AI conceived as a complex system composed of a multitude of techniques, for three reasons:
- The fact that these are simple tasks today does not mean they have always been so. Those who worked in BI in the 1990s-2000s can attest to this. The democratization of data collection and storage technologies is recent and should not overshadow the underlying technical challenges, sometimes very complex.
- In humans, memory is readily associated with intelligence. ‘Having a good memory’ means being able to absorb and retain a lot of information. This is exactly what data collection and storage technologies do. From the perspective of higher-value tasks, they may seem purely utilitarian, but they are actually indispensable.
- It is also a way to emphasize the importance of data quality and the famous adage ‘garbage in, garbage out.’ While it is the understanding and decision-making stages that generate value, it is often poor quality or insufficient data that causes AI projects to fail.
Finding Nuggets in Data Through Machine Learning
The next two tasks, understanding and decision-making, are often mixed up. It’s hard to distinguish them because some technologies do a bit of both simultaneously (reinforcement learning and deep learning both use neural networks, for example).
Understanding involves processing data to make it intelligible, usually through more or less advanced statistical analyses. In this regard, Machine Learning is an advanced form of statistics that broadly aims to identify strong links between data and attempt to answer questions we might ask: what explains the peaks in consumption of a particular product? What are the locations where it is more difficult or time-consuming to deliver a package? This information does not exist in raw data, and the goal of Machine Learning algorithms is to identify these links in a more or less supervised way to predict this data in the future. Recently, Deep Learning has enabled impressive advancements in areas where machines were struggling, such as image recognition (where models recognize entities in photographs or videos) or natural language processing (which allows deducing the meaning of a sentence from its analysis). These applications have brought the term AI to the public’s attention because we see machines performing tasks that were thought to be efficiently done only by humans.
Understanding tasks thus allow extracting high-level information from collected data, but sometimes, they do not enable decision-making. Not all AI applications require this stage, but it is unfortunately common to see AI projects in the industry where this stage is particularly neglected because it is poorly identified. For example, in the last mile, we can predict (with Machine Learning) the future evolution of the number of parcels a delivery center will have to handle. We can also imagine that this same ML model could provide optimal tour purchase decisions from their providers: this is absolutely not the case. The model will simply estimate the future quantity of parcels, under certain assumptions, and that’s it. Optimal decision-making is more complex to achieve, and Machine Learning is not at all the technique to be used for that.
Decision Support: The Challenge of Operational Research and Reinforcement Learning
This is the goal of decision support technologies, which are often unknown. The oldest, Operational Research, deals with optimization problems found in all industries: financial investment optimization, factory planning optimization, delivery route optimization. It requires strong mathematical and algorithmic expertise, and it is generally in these applications that AI can truly do “better” than humans because it tackles problems with very high complexity. Operational Research allows and requires us to describe (i.e., model) the problem exactly so that optimization algorithms can explore all possible solutions to this problem and find the best one. This ability to “reason under constraint” is the great strength of this approach, allowing algorithms to simultaneously manage all industry-specific constraints. It also enables these algorithms to find completely “out-of-the-box” solutions even for an expert eye.
This strength of Operational Research is also its weakness. Indeed, the problem needs to be described and modeled in the form of “possible evaluated choices.” For example, in a route optimization problem, routes can be described as sequences of points evaluated by the kilometers traveled by all routes. However, this is not always possible, such as in certain games like chess, Go, or video games (see the recent successes of OpenAI in Dota or Starcraft). Here, evaluating a decision is particularly difficult! How to “mathematically” evaluate a move in Go or a “click” in Starcraft? The combinations are too numerous, and the underlying optimization model, if it exists, is likely unknowable. This is where reinforcement learning comes in: if it is not possible to explain to the machine which decision is the best, let’s teach it to find it by itself. These techniques rely on large-scale repetition, made possible by cloud computing, of the problem submitted to the machine, tirelessly, so that it finds its own strategies through a cycle of trial and error.
