Movements on automating complex workflows

Part 1 (2017)
1

Hello all,

Hopefully this is not too off-topic for this class forum. If it is, I’m okay with deleting this thread. I would like to hear some opinions on this.

After reading some articles & lectures of the Deep Learning topic, I feel that in general DL (+ other ML techs) allows us to:

  1. Understand complex stuffs like sentences, pictures, audio clips, etc & extract the labels.
  2. Predict things based on collected statistics.
  3. Able to program mechanisms that are way too complex to model such as adaptive complex things, without coding the instructions manually. (self-driving, etc)

Number 1 allows us to automate the “sensing” processes which previously were not feasible as it is too complex & rigid. It is a good thing since it opens up many new possibilities.

Number 2 arguably has been done since many years ago. We just use way more parameters nowadays & ML DL might find new unnoticed correlations.

I would say number 3 is probably the most impactful one since the impact feels exponential. We can automate & do many things that were not feasible before. With it, we can program without having to lay out the exact sets of instructions (this feels linear).

Somehow, lots of online classes and articles about the achievements around DL & ML I have read were mostly about labelling things with accuracies better than humans (animals, MRI pics, etc). Also, Kaggle are mostly about classifying.

Why is it that we don’t see much movements on number 3? Other than self-driving cars, I am not aware of other complex workflows being automated (laundry? dust cleaning? security? real time coaching? resistance exoskeleton for sport trainings at home? etc)

If it turns out that it’s just me missing the information, please do let me know.

1 Like
2

I don’t know, but if I had to guess it’s that not enough people know how to do number 1 and number 2. As more people know how, there will be more startups and product teams tackling the number 3 challenges. I bet this MOOC will be the cause of many startups and product teams that will help the world, in a number 3 way.

To elaborate from an economic viewpoint, these number 3 things cost more. They often involve hardware and safety precautions. As more people know how to do number 1 and number 2, the cost of paying them will go down (and the development time will go down, because they’ll be faster), freeing up resources to spend on other parts of the product development process (and other processes like distribution, marketing/sales, fundraising, etc.).

I think Jeremy mentioned the lack of tools for deep learning in either the lesson 1 overview or a keynote. Imagine if there were no Keras and we had to use TensorFlow’s low-level API. I bet there are many things in the deep learning process that we will look back on as low-level APIs in a few years.

Two things you might be interested in:

  1. An attempt at automating cooking: http://www.moley.com/

  2. One of OpenAI’s goals is to create a household robot.

2 Likes
3

Suggesting actions is quite a bit harder than making predictions. Generally this is done with reinforcement learning. Here’s some applications: https://www.quora.com/What-are-some-practical-applications-of-reinforcement-learning

1 Like