KDnuggets KDnuggets Top Blogger: An Interview with Adit Deshpande, Deep Learning Aficionado

Read an interview with KDnuggets Top Blogger Adit Deshpande, a deep learning aficionado and masterful blogger, who also just happens to be a second year undergraduate student.
By Matthew Mayo, KDnuggets.

Our next installment of KDnuggets Top Blogger profiles focuses on Adit Deshpande. Adit is a computer science sophomore at UCLA, and is particularly interested in deep learning, something that is evident if you have read his blog posts.

Adit has contributed to KDnuggets three times over the past few months, with an upcoming fourth, and has achieved 2 gold Top Blogger badges for his fantastic and informative pieces.

Adit Deshpande
Aside from his posts on KDnuggets, Adit runs his own blog. You can find him on Twitter @aditdeshpande3. You can also find him on LinkedIn, and view some of his projects on GitHub.

Adit is currently searching for a Summer 2017 machine learning or

What advice would you give other undergraduates looking to tackle advanced machine learning topics?

Learn your math. ASAP. I definitely found this out the hard way. Getting a conceptual idea of how CNNs/RNNs work is great and if you want to just use popular deep learning frameworks for side projects and whatnot, then staying on that level of abstraction is probably good enough. However, if you really want to pursue a career in ML, learning the intricacies of backprop, other optimization techniques, and ML algorithms will be time well spent.
That being said, the math behind some ML algorithms definitely goes over my head a lot of the time, so it can be a tough road without the requisite courses. However, in order to first learn those conceptual ideas, I’d recommend Stanford CS 231N for CNNs, Stanford CS 224D for NLP, David Silver’s RL course for Reinforcement Learning, and Andrew’s Ng’s Coursera course for more of a general view of ML.

What is something you have read recently, whether or not related to machine learning, that you have found particularly interesting?

Don’t think this counts as something I read, but there was a really interesting talk given by Amnon Shashau (given at CVPR 2016) on how machine learning fits with self driving cars. I really enjoyed this presentation because I felt that Shashau really simplified this complex ML system into 3 main components, sensing, mapping, and the driving policy. I won’t go too far into these, but I thought it was very interesting to see all of the different factors and considerations that an ML based self driving system needs to take into account in order to be successful.

Where do you see deep learning taking artificial intelligence? Are we on the road to the "master algorithm" or general AI, or is this just another of many developments before we get to that point (assuming that is a reachable destination, of course)?

I think that deep learning is definitely a development that could prove to be a critical component of what’s necessary for the “master algorithm”. From the definition of deep learning itself, it is the idea of being able to represent the world in a nested hierarchy of concepts, where each concept is defined relative to simpler concepts (e.g The early layers of a CNN look for simple shapes and edges while the later layers look for more high level and abstract concepts such as faces of dogs). I think this really fits perfectly with the idea of a machine understanding the world, having intelligence, and becoming this general purpose AI.

In my mind, one really concerning obstacle that deep learning faces is this problem of interpretability. If we can’t probe a neural network to see why it made the classification it made, then can that still be labeled as intelligence (this term “intelligence” is definitely up for debate on what it actually entails)? Some may argue that you don’t necessarily need to know how/why a system works in order to use it (e.g Don’t need to understand rigid body physics to use an axe). People may also argue that humans themselves aren’t perfect at explaining their actions or decisions (e.g “I made decision X because it just felt right”).However, I think that lack of interpretability makes it a lot tougher for fields such as medicine, insurance, and law to accept the black box of deep learning.

Thanks for your time, Adit. I genuinely look forward to your future blog posts.

Adit Deshpande recent KDnuggets posts include:

  • Gold Blog
    9 Key Deep Learning Papers, Explained
    - 20 Sep 2016
    If you are interested in understanding the current state of deep learning, this post outlines and thoroughly summarizes 9 of the most influential contemporary papers in the field.
  • A Beginner’s Guide To Understanding Convolutional Neural Networks Part 2 - 08 Sep 2016
    This is the second part of a thorough introductory treatment of convolutional neural networks. Have a look after reading the first part.
  • Gold Blog
    A Beginner’s Guide To Understanding Convolutional Neural Networks Part 1
    - 06 Sep 2016
    This is the second part of a thorough introductory treatment of convolutional neural networks. Have a look after reading the first part.