Welcome to the second part of our series of articles on machine learning.
In part one we talked a bit about why understanding machine learning technologies is incredibly important in any field, let alone one that is as digitally facing as modern marketing.
This time we’re going to talk about what it can actually do for you. Let’s face it there is no shortage of commentary right now on how AI is going to doom us all (1), or save us all (2), or chase us all to Mars (3), without being terribly specific as to how.
So I’m going to talk about the types of tasks that machine learning is good at, and how this may be of interest to those of us who are, for whatever reasons, inclined towards marketing.
Strengths and weaknesses of machine learning
As we have already discussed, machine learning is an approach to AI that involves a computer system that develops its own instructions and decision making strategies, rather than relying on an existing list of instructions being provided. Machine learning is AI, but not all AI is machine learning.
- Within focused areas can be better at a task than a human being (sometimes dramatically)
- Can perform tasks far faster than a human being
- Can find solutions that humans wouldn’t consider
- Infinitely replicable
- Reliable and tireless
- Don’t require a salary
- Unlikely to unionize
- Restricted to tightly defined task or tasks
- Require large amounts of properly labelled data to train
- Will internalize biases that exist in that data
- Likely to be resource intensive or require constant cloud connection
- Unable to leverage wider understanding of context when solving problems. E.g – they can find solutions that a human wouldn’t consider for very good reasons
- Can be slow to respond to changing circumstances
- No understanding of cultural, legal, or ethical boundaries
- Supervision may require expensive and hard to recruit staff
- May be impossible for humans to understand or optimize resulting models
- May be prone to sabotage or manipulation by humans
- May be difficult to determine value of outcomes at a granular level, or spot mistakes
Just because a system was developed using machine learning, doesn’t mean that it is still actively learning. In many applications, a pre-trained machine learning model can be used (4), which will no longer respond to changing circumstances, but which will be less resource intensive. Commercial cloud applications will usually have been trained on publicly available or pooled client data. These could be continually learning in response to incoming data, but they could also be using a regularly updated pre-trained model.
Cloud sourced software and hardware resources that allow an organisation to create and maintain their own models, are readily available, but the skills required to get started are going to be considerably more demanding than using a neatly packaged product.
Machines, what are they good for?
Many of the tasks that are being performed by machine learning systems have been the target of automation for a long time, but in most cases they have lacked flexibility, sensitivity to context, or ability to respond to changing circumstances. Machine learning systems however, while still mostly weaker in these areas than humans, represent real progress in closing the gap, while retaining the innate advantages of digital problem solving.
So what do machine learning systems do best?
Enhancing the ability of humans, by predicting a likely next action and then providing needed information. A great example here would be predictive texting, where the system anticipates the language a user is going to use next, and speeds up their data entry. The end result here doesn’t have to be perfect, only good enough that it regularly provides useful information, but this is still a demanding task, requiring sensitivity to the context provided by previous language use to make suggestions, rather than just a regurgitation of commonly used words.
With machine learning, systems can adjust themselves to predict future needs, even on an individual user level, responding to which information is marked correct, and which advice is acted upon. Such a system can surface relevant information resources, suggest contacts, provide language suggestion for correspondence, and warn against actions that have previously proved negative.
AI systems excel in roles that allow them to perform the busy work adjacent to more complicated tasks. Automating form completion, prioritising incoming emails, recording outcomes, or accessing resources that will be needed for a specific task.
This frees up experienced staff for tasks that actually require their experience. The ability to draft correspondence, pre-personalized for a given recipient, can be of tremendous value for direct marketing campaigns, potentially reducing human input to a simple approval to proceed.
The web makes a massive amount of information available to automated research systems, but it’s not the only part of the internet that is accessible to AI. Many data collection organisations make API’s (application program interface) available that programs can interact with directly, reducing the need to try and “scrape” data directly from web pages.
More complicated applications can start to perform research tasks in a way that moves beyond simple cut and paste search, becoming much more sophisticated in how they parse information requests, identifying relevant information, weighing reliability across sources, and presenting the data in the most relevant format for the user. IBM’s Watson famously started out as a system designed to win at the quiz show, Jeopardy.
The ability to rapidly evaluate and look for patterns within large quantities of data has always been a key feature of computing. Where AI shine is in making this process far more streamlined, allowing data to be identified and pulled automatically into an analytical model, and the resulting insights surfaced to the user or integrated into the software’s own learning process. As with research applications the ability to transform data for appropriate analysis or comparison is tremendously important in increasing usefulness.
In marketing applications, the ability to track and link up profiles of users across different channels has tremendous value, as does automatic evaluation of programmatic ad locations. AI can also help identify the best timing for ad campaigns and social media updates. In general, in an industry in which there is tremendous value within data, AI systems can help expand the scope and maximise the value of existing data.
For many, voice recognition is now practically synonymous with the concept of an AI, as increasing numbers of people have access to virtual assistants such as Siri, Alexa, and Cortana. Being able to parse sentences, recognise faces & behavioural cues, or simply observe the environment around us are massively complicated tasks, and large portions of our own brains are devoted to the task. But massive amounts of research, and access to truly enormous datasets (5) have allowed fields like natural language processing and computer vision to make massive strides, to the point where software can exceed the ability of humans to parse speech, translate languages, recognize faces, and many other related tasks.
Probably the most obvious application here is as an interface, people can often work faster through spoken or written language than a traditional control panel, such a natural language interface can be less intimidating, especially for those who lack familiarity with technology.
Language sentiment analysis and image recognition can help brands identify discussion about a product or brand online and across social channels, as well as flag and prioritise negative conversation. Automatic transcription of phone conversations (6) & written documents or accurate translation of foreign language can be an immense boost to productivity for human marketers, especially in outward facing roles such as PR. Facial and behavioural recognition has tremendous potential to impact out of home advertising, once advertisers put more thought into making these campaigns feel more like marketing and less like Orwellian surveillance.
Great web content is tremendously important to marketing, so naturally there is a lot of interest in using AI to help produce content. Machine learning systems are capable of reviewing large amount of media and identifying the patterns that make it successful. They can then use these patterns to produce new media according to those rules. Machine learning systems have been built that can produce new images in the style of old masters, edit video, write screenplays, publish articles, create photorealistic images based on descriptions, or convincingly fake video or audio.
What is missing from the picture right now is originality. A machine learning system has no understanding of what it is copying, just an ability to remix or replicate media that fits the same pattern. For example, a machine learning systems might be great at writing news updates on sports results, which are constructed according to basic rules according to data input, but would absolutely not be able to write an meaningful article about what the hiring of a new manager means for a football team.
Where machine learning currently has the most value is, again, enhancing the productivity of existing workers, and making the acquisition of artistic skills less of a barrier to creating content. An AI system can do the heavy lifting, supervised by a human who can inject genuine insight and originality into the end result.
Bringing it together
Most individual machine learning systems are very focused on specific tasks, but as computing power increases, it becomes easier to tie multiple processes together and build systems that can respond flexibly to changing circumstances and complicated tasks. In fact, relatively mature applications such as speech recognition are already reliant on many separate AI processes in order to work.
This is where cloud computing becomes critical to advanced AI. Complicated machine learning systems will require many different, resource intensive processes. Positioning these processes in the cloud means that the resources can be shared between many different users, and only tapped into as required. So, for example rather than an Amazon echo box being forced to host all of the resources for every task it might need to perform, only rarely using them, the box can instead share a small proportion of a wide library of processes through the cloud, tying up resources only as required.
When considering any machine learning software deployment, you are going to need to ensure that you understand.
- Exactly when and how the model decision making will update in response to new data.
- Where the data used by the model is coming from.
- What happens to any data that you provide the system, and who else will benefit from it?
- Who else might have access to, and be competing to leverage the same insights.
- That adequate steps are being taken with regard to meeting your responsibilities under data protection regulations.
- The relative merits of prebuilt, user friendly, products Vs building your own system.
Keep in mind that different learning or decision making models are suitable for different tasks. There isn’t a single best approach to AI and, in turn, there isn’t a single best approach to machine learning. The most complicated deep learning will be outperformed by the simplest pre-programmed algorithms for many basic tasks. Ultimately you will want to judge performance on outcomes and usability, as with any other service.
Artificial Intelligence of all types represents an unprecedented opportunity to transform the way that marketing works as an industry, but just like any technology it can be overhyped and oversold. In the next article we will take a closer look at how to identify and avoid some of the common pitfalls encountered when using AI systems.