I’m consulting for a company that the had a crucial team member taking “extended vacations”. Ouch!
That made me think of how it was super important not to have “single point of failures” your team. Meaning, almost always ensure that at least two team members know how to handle a given piece of software or architechture. (Or at least have good documentation around it)
While talking about this in a team call, I made the analogy with the “Dropout” concept in Machine Learning.
“Dropout” is an ML technique that implies randomly shutting of neurons of your network, so the network doesn’t become too reliant on some “unique” neurons.
In that line one could make the case that it’d be good for a company to have a policy of employees being able to take PTO when they wanted, as long as they ensured the company can still continue to function without them overseeing the process.
That analogy above made me think if there were any other analogies that could be made between ML techniques and a company.
If we envision a startup as a neural network we could then propose:
Batch Normalization
In the Machine Learning world Batch Normalization means normalizing the input data before it’s processed by the network, so it has norm 0 and variance of 1.
In a company’s organization it would mean having a well-defined and consistent process for evaluating and onboarding new employees, so they can quickly become productive members of the team.
It would also mean having a polished hiring process that would reject super intelligent but emotionally unstable employees (high variance) that only add chaos to your organization.
Early stopping
In the Machine Learning world, Early stopping means stopping the training process before it’s completed to avoid overfitting.
In a company’s organization would mean being able to recognize when a project or initiative is good enough to be able to ship it.
I’d say it’d be the perfect analogy of launching fast an MVP in order to collect feedback, instead of trying to polish all the little details before testing it out with your client base.
L1/L2 regularization
in the Machine Learning world, L1/L2 regularization means adding a penalty term to the cost function to prevent single neurons from taking too much weight. This is similar to dropout in the sense that it tries to avoid overfitting.
I’d say the analogy to this in a company’s organization would be to prevent certain team members to dominate the zoom meetings by talking too much. Instead let all team members chat and express their opinions on the subject matter before taking a decision.
Momentum optimization
In the Machine Learning world, Momentum optimization means taking into account the past gradients when updating the model’s weights.
In a company’s organization would mean keeping track of what worked and what didn’t of past projects. Maybe keep a shared team notebook where everyone writes their learnings.
In this way as the company is growing (a.k.a the network is “learning”) it can keep be better informed on which “direction” gives the best benefits
Residual connections
In the Machine Learning world Residual connections mean adding the output from a previous layer to the output of the current layer in order to become the input of the next layer.
In a company’s organization this would mean fostering a culture of transparency and communication, so that everyone is informed and can build on the work of others.
In the most strict sense, I’d allow skip connections between layers of management. Meaning that if a single employee wants to notify the CEO about some internal issue, he should be in its right to do so
Pre-training and fine-tuning
In the Machine Learning world this would mean using a pre-trained model as a starting point and then fine-tuning (i.e chainging a little number of parameters here and there) it for a specific task.
In a company’s organization it would mean taking proven best practices and models from other successful companies and adapting them to fit the unique value prop of your org.
I think I could go on forever, stretching other ML topis to make them match organizational best practices.
It’s fun to play around with this: mapping of a concept from one field to another is a fun game to play and a good excercise.
I wonder what this would look like in embedding space… matching a vector in some given cluster to another vector cluster but in a similar position?
I don’t know really! Tell me what you think and if you can come up with other useful / fun analogies!