everywhere you go you take yourself

Not to be that person but just thought I’d give people a heads up, just in case you’d hate for your face to be a profile pic for some Russian bot account on social media....

here’s an article about FaceApp for more

Diagnoses are a matter of classification and categorization (Foucault 1982; Madigan 1992). BPD is currently understood through an arrangement of human behavior that classifies like individuals into typologies of deficit. A personality disorder diagnosis declares the deficit to be a fundamental feature of a person rather than a transient state. When a clinician, armed with this model and definition, makes a diagnosis of BPD, for instance, the power to classify derived from this knowledge can influence how individuals view themselves in relation to societal standards. In Foucault’s (1982) sense, the client may therefore internalize the problem discourse and come to understand themselves as deficient and that deficiency as a fundamental quality.

The etiology of BPD is a highly studied field by researchers, and even critics of BPD have adopted a causal model that names childhood abuse as a risk factor for BPD (Shaw and Proctor 2005). The public comes to think that BPD is the understandable and inescapable result of a stressor, when in fact it is a diagnosis dependent on the mere judgment of a clinician. This is to say, “there is no disorder … unless somebody with authority applies a psychiatric conceptualization” (Burstow 2005, 1299).

Of importance here is that a BPD diagnosis is situated within the dominant Western discourse on identity, a conception of selfhood that values autonomy and goal-directed behavior. These characteristics are closely tied to cultural norms of self-provision through work. In order for members of society to be self-sufficient and goal-directed, personality and identity must be conceptualized as relatively stable, inherent aspects of oneself that emerge through behaviors, traits, and other external manifestations (White 1999; Bradley and Drew 2006). In traditional treatment, clinicians decode and interpret these manifestations in relation to their deviation from society’s norms for behavior (Madigan 1992).

For example, self-injury and suicidal behaviors—two diagnostic criteria of BPD—are seen as pathological actions that undermine the valued sense of selfhood. Disrupting the dominant narrative of goal-directed behavior, the self-directed injury is seen as an inability to be an agentic, goal-directed individual. Some types of self-harm—such as overworking at one’s place of employment to the point of causing physical ailments, neglect of interpersonal relationships, and loss of sleep—are not seen as pathological because these acts resonate with cultural values, such as self-sacrifice for a greater goal. But because the self-directed nature of self-injury cannot be reconciled with other cultural norms, self-injury is seen as a manifestation of severe pathology; the person must be viewed as disordered for such an action to make sense (Madigan 1992).

Studies of BPD offer us reasons to rethink these dominant conceptions of pathological behavior and the supposed stability of identity. We know now that BPD symptoms diminish over time such that “after about 10 years, as many as half of the individuals no longer have a pattern of behavior that meets full criteria” for BPD (American Psychiatric Association 2013). Another study showed that among an adult cohort, 73 percent were in remission from symptoms after six years (Zanarini et al. 2003), which undermines the narrative that personality is largely unvarying. Furthermore, many symptoms of BPD are normative during adolescence, such as chaotic relationships, recklessness, and extreme emotional shifts, but deemed unacceptable in adulthood.

Feminist critics of BPD offer an alternative perspective, generally viewing the diagnosis of BPD as pathologizing the ways that women respond to gendered abuse and oppression. Shaw and Proctor (2005) theorize the diagnosis as a form of social control: “[BPD] can be applied to women who fail to live up to their gender role because they express anger and aggression. Conversely, the diagnosis is also given to women who conform ‘too strongly,’ by internalizing anger, and expressing this through self-focused behavior such as self-injury” (485). They show how the diagnosis of BPD presents a double-bind: women with BPD who engage in behaviors that are not stereotypically feminine—self-injury, multiple sexual partners, external expressions of anger—are cast in the archetype of the overemotional hysterical woman.

Charlotte Moorman, Performance wearing artist Nam June Paik’s “TV Cello” and “TV Glasses,” New York, 1971

So we apply softmax to the output to make sure everything adds up to 100% (the difference between softmax and vanilla normalization is that we exponentiate the values first to magnify the differences). And that sounds entirely reasonable if your idea of reasonable is “whatever conforms to probability theory,” but beyond that it’s a jarringly external operation. It’s not like the neurons themselves are organized to give a softmaxed output, because the neurons in one layer don’t know about the values of the neurons in the same layer. Sure, you could probably add more layers responsible for normalizing the output of exaggerated differences to try to stick to a minimal set of rules for the heck of it, but why? if you take a step back you have to notice that human brains simply don’t do this. Human brains for the most part stick to incoherent outputs that don’t sum to 100% (unless the brain is made to think really hard and double check the options to unnaturally force a consistent output) .

And, I don’t know, I feel like there’s something important there but it’s difficult to think about without getting into a masturbatory and implausible treatise on Godel / completeness vs consistency / non-monotic logic.  

My broader point is (to plagiarize Hinton) that the biggest obstacle deep learning currently faces is that it works. It’s developed tools that are effective but clearly wrong, and this leads to weird situations where researchers are sort of resistant to fresh, completely unproven approaches. This is sort of understandable, we don’t know that what we have is definitely broke, so why fix it? But like, if your definition of not broke is restricted to the parts that work then …I don’t know that sounds a lot like it’s broke? Regardless, almost everyone (except for OpenAI, apparently) agrees we can’t keep relying on algos restricted by the designs of GPUs and TPUs. LIke yeah, the algorithms scale fine, but with such a large scalar value, the power requirements are absurd.