iA


An Equivalenace of Guṇas & Box's Loop

by Karthik Dinakar. Average Reading Time: almost 5 minutes.

The sagacious Joost Bonsen once tried to outline his theory of multiplexing ideas, a chapter in his vast treasure-trove of constructs on engineering creativity in my first year at MIT. I found his elucidation to be coherent, almost like an Occam’s razor. But as with most constructs in life, what we deem fathomable in theory are but a train of shadows in practice. The vagaries of the real world, with its messiness and its formidable distribution of noise give us a dose of reality at best and encumber our optimism at their worst. I wanted to completely imbue Joost’s advice into my thoughts, but I found it too daring. I told him I was skeptical. But I was wrong and he was right.

Almost three years later, after many trysts with probabilistic graphical models, I find myself in the southern city of Atlanta, where I will be joining the extraordinary Matthew Nock in delivering a keynote address to the American Association of Suicidology. How in the world a skinny brown boy working on approximate posterior inference and tensor spectral methods can find himself in a community of advanced clinical psychologists studying the most devastating and dark mysteries of the human psyche – suicide and self-harm – is not at all very clear.

Of all complex constructs concerning the human psyche, suicide and self-harm are the most bewildering. How, in an ultimate act of self-protection, one commits self-annihilation is an area that is opaque and dark, laden with more questions than any conclusive answers. It makes me very sad whenever I think about it, and I can’t imagine how hard it must be for Matthew and researchers like him to be researching it every day of their lives.

Varṇam & Box's loop

Matthew Nock and Eric Horvitz

When I woke up today, I tried to listen to māthé malayadwaja sañjāthé, my favorite varṇam. As I started singing along, trying to collect overnight thoughts on posterior model checking and discrepancy functions, I felt a surge of deep gratitude as two very warm faces landed at the center of my thoughts. Matthew Nock and Eric Horvitz, two perspicacious human beings whose deep intellect is matched equally by the generosity of their spirits. It is hard to describe in words the connections and sūtras the mind finds itself capable of perceiving when it is cloaked in the kindness and warmth of other human beings.  I remembered what each of these two men had already done for me, as soon my thoughts were unable to differentiate their faces from the full import, the full impact and the full promise of their work, woven together by their magnetic personalities. I felt the presence of my guṇas viscerally, and when I sang māthé malayadwaja sañjāthé a second time, I was propelled into a synesthetic voyage across George Box’s loop, colorful and rich in granular detail as how we might ever devise a cognitive model of self-harm, one that is honed through meticulous observation and one whose translation into clinical manuals can save, even alter lives for the good.

pallavi:  We if are to implode a given complex phenomenon, into separate and distinct parts of problem specification, problem estimation and problem distribution as Fisher and Neyman once did, how to tackle a problem of sparsity and noise, of burstiness and highly contextual, as self-harm is? How can we break from our self-limiting notions of data cleaning, and choose a representation for a model?

anupallavi: Must we be forced to choose between two fettered choices of theory-driven inductive models and data-driven deductive ones? How can we imbue the selection of priors with our best understanding of self-harm so far, allowing data to correct and override them if it so happens?

śwara: Can we resist the temptation to merely sample, and derive variational updates that respect both the sparsity of data, yet give affordances for human-input during each iterative update?

jāthi: Instead of merely substituting an axe for a scalpel and use hackneyed, even forlorn measures of model performance, can we write a series of discrepancy functions in line with asking ‘what are some insights from the model we really want to know’?

sāhitya:  Can we discard hapless fights between the frequentist and Bayesian worlds, sample from our observed data directly, and also from what our generative model gives us, and apply the above discrepancy functions that allows to question and address ways in which our assumptions have failed us, or where the model produces insights that are mundanely too obvious? Can we truly criticize our model and go back to the drawing board as many times as it may take?

carṇam: Can we augment our model, devise a decision theoretic layer, overlay a vector interventions for folks engaging in self-harm, and adorn our model with embellishments crafted not only to ease and please, but has human-centered design as its core principle?

A balance of Guṇas

As these notions became clearer, I was now singing the carṇam, an improvised upward crescendo, with Matt and Eric still visibly at the center. I felt enormous gratitude and energy, as my thoughts now morphed into an image of Rajarajeshwari herself, complete with rich detail. As I glimpsed her from toe to head, a tapestry of such kindness and empathy, I realized her face was actually that of my great late grandmother. Any thought that leads to my grandmother is profoundly sacred to me, as I have neither seen nor experienced anything like her kindness. As the final notes of the word māthé left my tongue, tears of joy were falling thick and fast onto my lap, and I felt such peace and hope, that I am blessed to work with Matt and Eric, the perfect balance of guṇas.