“How do we infer data without any unconscious human biases?” a corporate executive asked.
“Well, we speak the truth as we see them” an equally erudite professional replied.
“How do I see it without your version of truth?” the executive retorted.
As business functions and public policy domains embed machines with human-like intelligence, ever growing number of corporate leaders and managers are increasingly making choices, drawing inferences and examining future consequences under conditions of uncertainty, driven largely by data, without introducing unconscious cognitive human biases. This enables corporate leaders to appropriately intervene the machines intelligence for decisions knowing fully the expected future consequences. The situation has probably been well explained in a scene from the movie Searching for Bobby Fischer where Bruce Pandolfini, the teacher, tells his ten-year old prodigy Josh Waitzkin, “don’t move until you figure it out in your head,” and then sweeps the pieces off the board forcing his student to go without his crutch of having to see the pieces, just to get him to visualize his strategy before he acts. However, these unconscious cognitive biases that often lead to sub-optimal decisions are spread far and wide in most organizations (see here and here). The topic that we are passionate about is how do we train machines to infer dynamic data, aided and unaided as with and without, human biases in (im)perfect and (in)complete information conditions, to assist corporate leaders better “manage the probable” for future economic conditions and “lead the possibilities” for business decisions and their expected consequences.
Although, corporate leaders who routinely handle investment and operational decisions under conditions of uncertainty require new thought leadership that juxtaposes between human and machine intelligence, yet also importantly oversees “human-loop” intelligence across the value-chain to address spin-gaps that shape real-world economic activities. We will explain the spin-gap and its associated costs later in the article. Perhaps, it is also important to recognize that, although some intelligence methods are combinatorial and, therefore, handle perfect and complete information conditions, the vast majority of business decisions, as opposed to facts, are largely applied to imperfect and incomplete information conditions. Some of these methods include econometrics, statistical (machine) learning, operations research or management science, linear, integer and dynamic programming, Markov decisions, game theory, optimal control, etc. that permit computers to find optimal or “satisficing” solutions by imposing a strong mathematical structure on the decision-making problem (Simon 1978). And, other methods include network science, which includes artificial neural networks, graph theory, statistical physics, quantum statistical mechanics, behavioral economics, etc. that emulates human’s deep (sometimes hidden) information structure to find “why” and “how” individuals go about solving problems and making decisions that are stochastic and, therefore, not often optimal. The question is what in interaction structures of economic or knowledge networks accounts for intelligence, if any, to bridge a potential spin-gap and the efficiency of decision-making effectiveness.
The term “spin-gap”, taken from physical science, is characterized by a separation of two-states – the excited state from the ground state – by a significant amount of energy. Since, this is not extensively used in business terminology, as part of the preamble, let me explain the term for the reader. To put it simply, if my virtual assistant (machine intelligence) autonomously sets up a meeting on a particular day and time with your virtual assistant without my and your (human intelligence) prior knowledge, this situation will create what we call “spin-gap” of intelligence. And, if either of us have other prior engagement at that time, in order to fulfill the machine intelligence, the human intelligence will probably incur opportunity costs or spin-gap costs. Well, as difficult and disjointed as it is to give a comprehensive explanation of the term, the investment trader who is exposed to decision making under uncertainty faces an incoherence of choice and the challenge of deciding whether to spend time statistical programing, comprehending and deciding on multi-dimensional dynamic and age-dependent data allied to human-created insights, inferences and metrics gleaned from his memory for knowledge and intelligent strategic actions, or using autonomously machine-generated choice options on various assets, sourcing ‘multiple-layered’ content and saving it in the “deep network” for future access, as needed, which would probably posit a completely different knowledge and decisions as the spin-gap of intelligence. This could be relevant for other senior executives in marketing, finance, manufacturing and operations as well. If these functional executives expect that there is not a significant knowledge-gap to fill for decisions, then they will have to invest the necessary time to set-up metrics and accountabilities for decisions and give managers and subordinates the time to communicate them across the organization for future effectiveness. However, if these executives are willing to fill the gap in their knowledge, skills, capability, etc., for decision effectiveness by using machines to create and retain such knowledge for future access, then such investment of time – to set up and follow ups on metrics and accountabilities – by the executives is highly unlikely.
Here we would also like to make a distinction between precision engineering and intelligence. One of the fundamental principles in precision engineering is that of determinism. System behavior is fully predictable, even to nanometer-scale motions. To do the job efficiently and correctly one also requires models and algorithms, where the basic idea is that machine obeys cause and effect relationships that are within human ability to understand and control and that there is nothing random or probabilistic about their behavior. Further, the causes are not esoteric and uncontrollable, but can be explained in terms of familiar and precise engineering principles. Intelligence, on the other hand, is a process that gives humans (and in case of machines it is termed “artificial intelligence”) the ability to learn to handle difficult multivariate decision problems, sometimes involving uncertainty. Intelligence, as opposed to fact, is stochastic in nature. It finds optimal solutions, derives reasons, infers actions, recognizes patterns, comprehends ideas, solves problems and uses language to communicate, drawing from both (im)perfect and (in)complete information conditions.
Why Unconscious Biases Are So Rampant?
First, most economic system are not designed, instead they have evolved over time. Human biases and heuristics are probably the genesis of creativity, imaginations and creative interpretations of data. They often lack an interdisciplinary branch that deals with the behavior of dynamical systems with inputs (see here), and how their behaviors and clouded accountabilities are modified by feedback for future business outcomes. We would like to clarify that although this does not conflict with the strategic planning function, since most strategy planning is an outside view for organizational growth, the models in machine intelligence provide an inside view of organic growth (See here Daniel Kahneman on Outer and Inner view) for the future.
Second, enterprise intelligence, as defined by IBM (see here), in which an organization analyzes “disparate data sources to identify the vulnerabilities, fraud and threats across your enterprise” builds hidden patterns that looks out into ‘future’ probabilities and probability distributions under (im)perfect, (in)complete information conditions. More often, the cross-functional models in machine intelligence identifies greater opportunities for revenue growth, cost reduction and risk mitigation and disinters the underlying causal factors and drivers for expectation maximization and maximum likelihoods.
Third, business intelligence, in which a set of pre-determined ‘measurements’ create a hierarchy of performance metrics and benchmarking based on pre-defined-pre-determined ‘fixed’ models for progress towards business goals, generates reports that support ex-post executive decisions and collaborates on programs that work together through data sharing across the organization. Instead, the models in machine intelligence leverage various methods, tools and techniques (here we advocate that such models should not be developed based on historical data or patterns thereof, instead these models needs to be designed for machine to learn from historical data and calibrated for future business goals and outcomes) autonomously derive new metrics for the opportunities and outcomes that the business is seeking at some future point in time.
Fourth, corporate intelligence, defined as understanding and learning what is happening in the world outside the business so that the business can be as competitive as possible, skews information with early identification of risks and predispositions inside the organization before they may or may not become obvious. However, the models that train machine intelligence dynamically define and re-shape the organizations that manage those metrics to stay competitive in the marketplace.
So, What is Cognitive Unaided Elint?
The Cognitive Unaided Elint, or CUE, is a multi-layered multi-dimensional mechanism on neural network structures that machines learn to determine signals, uncover hidden states, prescribe likely actions, predict responses and optimize (+ or -) rewards on a defined decision problem – be it financial, marketing, operations or service engineering – using enterprise, extended enterprise and external data without intervention of human intelligence and analyze the same with expected bounded rationalities largely caused by humans. CUE enables corporate leaders and managers to perform dynamic planning, forecasting and optimizing future market/business conditions consistently in age-dependent decisions with non-parametric non-linear systemic use of unbiased insights, inferences and impact thereof. This also enables the corporate leadership to re-shape organizations to manage those metrics that are relevant to future competitiveness.
At a basic level, it allows corporate decision makers to view the trade-offs between machine intelligence and human intelligence where automation of tasks, if not the job, is necessary vis-à-vis augmentation of human performance. To illustrate, let me cite the service market where the provider is often not sure about the choice between human intelligence and machine intelligence or intervention of human intelligence into the machine intelligence or vice-a-versa, nor the efficiency and productivity thereof, at the time of delivering the service. An international airline, for example, uses the online ticketing system to collect profile, travel preferences and travel pattern of travelers. This view on bookings and confirmation systems give airlines instant visibility of the market and enables them to identify opportunities to maximize revenue. However, since airlines are always not sure if the customer is better served with only the online system, or booking over phone with human assistance, or intervention of human assistance for the online system, they allocate costs of both human and machine intelligence systems, without knowing a priori their contributions to the overall efficiency and productivity. However, the customer is generally aware of all the procedures of travel and rules of booking and check-in, including price, seat allocation, cancellation, change fees, baggage limits, check-in time, etc. at the time of booking. The machine intelligence system is generally combinatorial in nature and, therefore, optimizes procedures and rules that are in complete information conditions to a satisficing and aspiring level (Simon 1996). Some observables, alterable elements as attributes, with spin-gap characteristics attached to a customer, are subject to manipulation and, therefore, require humans to intervene and interact with the machine. However, the CUE applies methods for a deeper learning of such underlying latent variables to identify how individuals interact in different “cognitive” systems on the neural network structures using algebraic topology – a system used to describe networks with constantly changing spaces and structures. These, so called interaction mechanisms, are paramount in understanding the emergent nature of collective decisions, as they often lead to common features and probe the same patterns within the network structure in more detail (See here). And that shines light on the “completeness” of information, and remains “salient” if information is “incomplete” and therefore “dissatisfied” until the information becomes “complete” and “satisfied” at some future time.
The spin-gap further expands, and probably becomes even larger where there is “imperfect” or asymmetric information that requires “random” search for decisions. Imagine a scenario, for example, where a one set of machine intelligence screens (screener system) the travelers’ luggage and body, while another set of machine intelligence, replacing human intelligence intervention, learns to observe and randomly search (search system) the screener system to find something malicious and noteworthy features that are either anomalous or unknown to the search system. The trade-offs between speed and accuracy of the screener system, on the one hand, and risk and uncertainty of the search system, on the other, has also been an important focus to improve checkpoint inspection performance which highlights the need for both effective detection (minimizing “false negatives”) and exposure of the false alarm (minimizing “false positives” or mitigating their negative consequences). A large body of evidence shows that human choices, as in this case the traveler’s choices, are not consistent and transitive as they would be if a utility function existed (Kahneman and Tversky 1973). Therefore, additionally, we want the machine intelligence for search to be capable of detecting and learning more types of anomalies and abnormal human behaviors, in this case traveler behaviors, besides performing tasks such as evaluating baggage content, mis-association of travelers with their carry-on bags and body-scans by leveraging the image data from the continuously panning images.
Ok. That’s enough for now.
There is indeed at some point in the decomposition of machine (hardware, software, services, etc.) and division of labor (management, managers, officers, etc.) that such machine intelligence can legitimately be asked to refrain from supplying the values to the organization. But, in so far as the machine intelligence can make the issues involved clearer, they may seem to have a function to perform in this field as well. The point really is that we are not interested in today’s valuation of tomorrow’s satisfaction but in tomorrow’s satisfaction itself. The more distant the future, the lesser the likelihood of human being interested in (transferring his knowledge into the machine merely to replace self) his inheritor than in his own. Thus, the future may indeed mean less to him and some sort of time discount is very understandable. While no precise calculation of diminishing marginal utility or of uncertainty discount is possible, this may help humans, and therefore corporate leadership, to roughly fix the planning horizon to their advantage.
Moreover, since each person gathers his or her own set of competencies, humans may choose either not to or may not know how to transfer these competencies as input, as the situation or environment in each case may lead to different results. Since every machine will learn and infer each situation differently and progressively develop different learning maturity, irrespective of their inherent personalities, they will always lack aspects of human intelligence, creativity and imaginations that require decisions formulated as a heuristic and, therefore, remain uncertain.
After all, human creativity begins where machine intelligence ends.
(Authors: Prabir Sen and Chandrashekhar Gopinath)
 Kahneman, D., Slovic, P., & Tversky, A. (Eds.). (1982). Judgment under uncertainty: Heuristics and biases. New York : Cambridge University Press.
 Simon, H.A (1978): On How to Decide What to Do, The Bell Journal of Economics, 9 (1978)
 Kahneman, D (2011): Beware the ‘inside view’ McKinsey Quarterly, November, 2011
 Simon, H.A (1996): The Sciences of the Artificial, The MIT Press 1996
 Reimann, M.W, Nolte, M, Scolamiero, M, Turner, K, Perin, R, Chindemi, G, Dłotko, P, Levi, R, Hess, K and Markram, H (2017): Cliques of Neurons Bound into Cavities Provide a Missing Link between Structure and Function, Front. Comput. Neurosci., 12 June 2017
 Kahneman, D and Tversky, A. (1973): On The Psychology of Prediction, Psychological Review, 80 (1973) 237-251
First published on LinkedIn