Unlocking Machine Learning with Human Centered Design
Human Centered Design (HCD) is at its core, a process for eliciting the practice of human values and embedding those new values into a new artifact, process, or service. It is perhaps challenging to consider how HCD may have a role within something as quantitative as Machine Learning, yet HCD can be an important component to the formation downstream benefit for machine learning services. The value to injecting HCD into the design and use of machine learning solutions is valuable to contribute to unknown opportunities and risk the threat of future-tense machine learning decisions.
Quick Intro to Machine Learning
Machine learning is best described as a methodology for a computer program to learn to write new programs without human assistance. There are two primary domains of machine learning - supervised and unsupervised. Unsupervised machine learning utilizes various statistical functions to identify patterns in information and to extrapolate decisions from that information. The reference to Big Data is often a use of unsupervised algorithms that are, at their core, a regression analysis. Supervised Machine Learning also leverages statistics, but to do so, relies upon a body of data curated by an individual. The computer program builds an understanding of the data to replicate it or to use that understanding as a filter on other data. For example, if 2000 images of a tree are shared with the program, it will use those 2000 images to build a concept of a tree, and will then be able to identify a tree from a suite of images which may or may not contain trees.
Of Machine Models and Human Cognition
To build an HCD approach to Machine Learning, it is important to first distinguish how machine learning is similar and different from Human Learning. First, both the human brain and the ML program do have a core similarity - they are massive engines of statistical pattern recognition. Our brains engage and understand the world through pattern recognition. To see an apple, know it is an apple, and understand what an apple does is in fact a massively complex undertaking. The connection between sight and concept requires our brain to internalize and make sense of about 8,000 different points of information... how light bounces off the surface into our eyes to inform geometry, color, shape, texture, and subtle visual implications of weight, density, and so on that are difficult to articulate. When our brain has sufficiently identified enough information bits to create a pattern - the apple pattern - it has built a conceptual model. Our brains are massive repositories of conceptual models, and we use these models to discover new ones. "That is not an apple but it looks like an apple."
Machine learning programs are similar (as most are based on our brain's neural architecture). They use many different methods to analyze information and identify a pattern, like "tree." Yet their methods are various different from ours, such as rapid quantitative measurements between corners, measurements of curvature, measurement of gaps - and thus generate a very different kind of model. The ML model is not interpretable to the human - nor does it need to be. The result is the total information required by a computer to make inferences about other possible models. A computer, also, has limits and it is difficult to make inferences. Thus a human can rapidly intuit a new situation "apple by a tree suggests apple tree," the computer will only see an apple and see a tree, unless enough information is presented about apple trees to make the leap.
The Learning and Reproduction of Human Values
Machine Learning asserts a transactional value by using found pattern within a set of information to predict how that pattern is extended against a new inflow of information. We can use ML to facilitate market analysis, reduce risk in complex situations, and optimize organizations. Yet what does this ability inform... optimize how? Reduce risk to do what? HCD is a process that requires the designer also conducts pattern analysis - but a key distinction is in the domain of qualitative patterns. How does a user feel about a given situation over time? How can that situation shift over time to elicit other feelings? The intersection of opportunities between HCD and Machine Learning are vast when you consider the bigger context of the data.
Within supervised machine learning, the assembly of training data could be done in bulk or it can be curated. A designer can pay close attention to the data... what kind of trees are presented? What health are they in? Can multiple sets of training data be presented that contain healthy trees, sick trees, and variations in environmental conditions? By curating the data, the human agent can tune the specific types of patterns to be determined, and thus generate greater value later. Not all data is equal.
Likewise, as the outputs are generated from an ML program - such as new sets of data (often relayed in a dashboard or a customized suggestion), the human experience of that new data point can be catalogued and leveraged to tune the ML program. While there is such a thing of "Human in the Loop," to aid the precision of the algorithm, another way to thin about this is "Value in the Loop," to reward the program for generating particular kinds of experiences. While abstract, the program does not need to understand the human value - it will build its own conceptual model to make sense of the behavior. These values can tune the algorithm to new directions over time, wherein we can expect the algorithm to generate new data points very much unlike the original yet of great use.
Geographic Variance in Modelling
A critical aspect of being human is the variations that exists between humans for perception in the world. Within small groups, language can take on a range of meanings, while across geographies, the variance of meaning ascribed to concrete things can be quite large. Machine learning generates one kind of solution for one kind of person and then generate a new solution for another kind of person... it can be granular, like Netflix, to suggest customized outputs. Yet what about machine learning applications for broad data sets that are irregular?
Within the design community, I advocate close consideration of the role of place and data. There is a tendency for developers to look to large open data sets. Vast databases exist with training data (like this popular one at MIT). Yet consistent with my above comments, what are the values which drive the labelling of this data? For example, if you apply this training data to images of Mogadishu, it will generate results like "earthquake." Not only is this inaccurate - but the correct label will vary depending on who and where the labeler resides. A Somali in Nairobi will give a different answer than a design student in Boston to describe a set of data. Yet most Somalis in Nairobi will supply more similar labels.
Consequently, it is not enough to parse data by meaning. One must also ground the meaning, and to do this, I suggest the use of classic lat/long coordinates. If we can build rich databases of place-based data and place-based identification, we can do more than build intelligent softwares, we can build softwares that are flexible to the global shifts in meaning and identity which are traditionally at odds with the demands of computation. To fuse machine learning with the vast ocean of human value creation and reproduction is a great opportunity for the future.