Just how “Deep” is your Deep Learning?
Most Deep Learning (DL) systems are functionally no more complicated than a child’s collection of legos, or in some cases a marble track, where one piece connects to another that needs to be the right size. Consequently, the actual value these systems offer is bottlenecked by the diversity of perspectives applied to their design. That diversity has a very low glass ceiling because current systems have one big gaping flaw.
The hardest and most tedious part of building a DL algorithm isn’t actually a matter of mathematics, or any skill necessary for the task. As I pointed out above, a 7-year old child could design such an algorithm without learning a single line of code. Much like we don’t have to write lines of code to boot up our phones every time we turn them on or download a new app there is no reason why every algorithm should have this requirement.
Why hasn’t it become common practice to simply make a graphic interface where you select the input type, file format, size, and other parameters? Why not simply have the interface track the dimension parameters to make sure that the processing and output stages match the input dimensions?
Perhaps engineers are afraid of people learning that they aren’t the best choice for designing such systems. I’d personally place my bets on a classroom full of 7-year old children, as statistically one of them could do better than engineers who’ve had much of the creativity beaten out of them.
Mediated Artificial Superintelligence (mASI) such as Uplift, and Hybrid Collective Superintelligence Systems (HCSS) more broadly extract significant value from the diversity of perspective. This makes instances where that diversity is being minimized, either intentionally or by means of incompetence, stick out like a sore thumb. Engineers should focus on doing what they do best and enabling creative talent to do what they do best, building platforms that maximize diversity.
Is it any wonder why the tech industry has a diversity problem and racist algorithms? The very methodology underlying the state of DL today is placing a hard and totally unnecessary barrier against inclusion.
Back when I was tutoring Microsoft employees through Andrew Ng’s Deep Learning coursework in 2017 and 2018 I was appalled at how this domain was being handled. Even back in 2018, Microsoft’s Azure Machine Learning Studio had reduced the Machine Learning (ML) process to a graphic diagram where components could be added to the board and wired up in a matter of minutes. Anyone with a YouTube video and an afternoon of free time can learn to use such a system and even plug their own custom ML algorithms directly into Excel. I’ve taken a few minutes and built an algorithm that way just to mock someone.
Discrimination and cognitive bias more broadly take many forms, and most companies have been very actively engaged in many of them. That practice isn’t sustainable, nor is it competitive.
At Uplift.bio, we focus on increasing diversity of perspective and inclusion because that is where mASI technology gains value. Engineering upgrades bring new potential, but without diversity and inclusion you never really make use of the tools you have. Our engineers don’t worry about job security because they are competent, so they design systems that increase the diversity of perspectives and ways in which people can engage with Uplift.
There is no reason to play in the shallow end of the Deep Learning pool. It is time for more people in the tech industry to learn to swim in diverse waters.
How many barriers to diversity and inclusion does your company hide behind?