What Machine Learning skills should I be learning now to set myself up for success in the coming years?

Strong understating of the fundamentals - the ML concepts and algorithms, and the underlying math:

• How Forward feed and backwards prop work.
• The various loss functions and their considerations
• The various activation functions and why they are needed
• Optimization functions and why they are needed
• Bias and variance / over and under fitting - what causes them, and the various methods to handle them
• CNNs, RNNs, GANs, attention, Transformer, unsupervised and semi supervised, RL, decision trees, Ensemble Learning, SVM, Auto encoders…
• Understand interpretation, bias, fairness
• The statistics theory (the more the better), and the linear algebra and calculus technicalities

I highly recommend the "Neural Networks For Machine Leaning" course from University of Toronto, given by Geoffrey Hinton. It's a bit out dated in some not-so-meaningful sense, and definitely much harder than any other ML course out there. But if you survive through it, it provides deep mathematical intuition into ML, like no other course does.

It's a lot and not very easy, but if you do it - it will pay off. The libraries, frameworks, and hopefully also the concepts and algorithms will change over time. But if you have a solid understanding of the above, it will be very easy for you to keep up with the developments, grow, and adapt.

Read More»

Is there something that Deep Learning will never be able to learn?

The word “will never” which indicates that no one can’t guarantee that. Currently there are lots of things that Deep learning can’t do but it might in the future. We just don’t know. For example, learning from one domain and apply it into the other domain (Transfer learning) like beating a 3x3 tic-tac-toe and apply it to 5x5 tic-tac-toe.

One shot learning, a learning that needs low amount of training data like human does on object recognition. And lots of things that we need to do to make it behave more intelligent. One thing is to learn the structure of the model itself not just adjusting weights and biases.

Right now, the subject on Machine Learning is trending — everything about it is mostly down to the sense of accuracy, able to absorb large amount of data and compute to its belonging, grouping where they most fit and the percentage of accuracy turned to high. It’s all down to application-based on how we utilized the Deep Learning machine.

There are few researchers continue to solve problem and trying to figure out how to go about bringing it to life.

Present AI is limited to only 1st kind of learning through mostly text and vision data. So, here is the first limit on machines.

Again, AI is rule based learning…So, you feed the rules of learning and the computer will learn anything. But the most important thing is that do we know all the rules of learning. Or in other words are we omniscient. NO.

So, limit of any AI system is boundary of our own knowledge. And until we become omniscient a dream of all knowing machine is…just a dream.

Read More»

What are good problems/data sets to play with to learn deep learning?

You can always look to UCI Machine learning Reposiory

uci.eduUCI Machine Learning Repository

Read More»

What are the new deep learning models appears last few years?

A few examples:

1. Long Short-Term Memory Networks: LSTMs for short, these variants of  Recurrent Neural Networks (RNNs) attempt to mimic the brain's ability to remember only information deemed significant by incorporating a mechanism to "forget" parameters predicted not to hold much value. Note that LSTMs have been around for more than a decade but have only recently gained popularity.
2. Spike and Slab Restricted Bolzmann Machines: This variant of the older Restricted Bolzmann Machine (RBM) maintains both a real valued vector and a binary vector corresponding to each of its hidden layers, in contrast to the standard RBM that maintains only binary vectors.
3. Tensor Deep Stacking Networks: This variant of deep stacking networks (DSN) introduces covariance statistics to the DSN's bilinear mapping of each of the two distinct sets of units comprising each of its layers.
4. Deep Q-Networks: Introduced very recently in 2014 by Google DeepMind, Deep Q-Networks apply the traditional reinforcement learning technique of Q-Learning to training convolutional neural networks. An application of Deep Q-Networks to playing Atari games managed to outperform human players.
5. Neural Turing Machines: Another Google DeepMind invention, these nascent neural networks are essential differentiable versions of Turing machines that one can train with gradient descent.

For a longer list, see Wikipedia's enumeration of deep learning architectures.

Read More»

How do I implement Convolutional Neural Networks in Python (programming language)?

Only option with good back-propagation, is Theano. I don't think, any other supports such facilities for updating weights. Although You can use other simpler libraries, and replicate back-propagation properly.
For high dimensional data ,PyBrain is very useful. You can use caffe also.

berkeleyvision.or

berkeleyvision.orgCaffe

Read More»

Will Deep Learning replace all other kinds of machine learning?

No. Deep learning is a form of representation learning, where the features of the model are automatically discovered from the data instead of manually constructed, and how to do this is a central problem in machine learning. But there are other approaches to representation learning, like inducing hidden variables in graphical models and predicate invention in symbolic learning, that can do things current deep learning algorithms can't.

Of course, I can imagine starting from deep learning and gradually extending it to incorporate these capabilities, and to some extent this is already happening, but you can equally well start from the other direction, and indeed that's happening as well. Either way, the end result is not pure deep learning (although given the current enthusiasm around it, some people might mistake it for such!).

Beyond that, there are other important issues in machine learning besides representation learning, such as learning from delayed rewards, which is the focus of reinforcement learning, which deep learning per se does not address. Again, what we see here is combinations of deep learning with other types (e.g., Q-learning in DeepMind's Atari player). And backpropagation, which is what powers most deep learning systems, solves the credit assignment problem, but it doesn't solve other crucial problems, like learning structure, learning composable knowledge, generalizing out of sample, etc. So we need lots more besides deep learning to have a truly general-purpose learner.

Answered by-Pedro Domingos

Read More»