When we offered CS231n (Deep Learning class) at Stanford, we intentionally designed the programming assignments to include explicit calculations involved in backpropagation on the lowest level. The students had to implement the forward and the backward pass of each layer in raw numpy. This is seemingly a perfectly sensible appeal - if you're never going to write backward passes once the class is over, why practice writing them? Are we just torturing the students for our own amusement? Some easy answers could make arguments along the lines of "it's worth knowing what's under the hood as an intellectual curiosity", or perhaps "you might want to improve on the core algorithm later", but there is a much stronger and practical argument, which I wanted to devote a whole post to: The problem with Backpropagation is that it is a leaky abstraction.

Jobs that design automation to manipulate human behavior. That's five classes of jobs that will exist in the future that appears to be safe. Paying someone for a job other than these will likely be illegal. On the other hand, with the exception of the "human safety valve", all these other jobs require high level skills. Jobs of the future need to have an deep understanding of humans as well as machines, and it is in this interaction of man with machine where jobs will exist. I think what few seem to appreciate is that Deep Learning AI is technology that is like human intuition. It is an opposite technology from more classical AI technologies that focused on reasoning. At this time there remains a Semantic Gap. However, humans capabilities are stuck between a rock and a hard place.

Machine Learning, and more so Deep Learning, is so popular now that it is being referred as AI itself. Gladly, your startup just got funded or your new team budget was approved! Now you will be doing Deep Learning as well. You already had fun with Keras, Imagenet, etc. Now here are a few things to consider when getting started for real in your business. I'll illustrate my suggestions with some anecdotes from my time working on self-driving cars at comma.ai with George Hotz early last year. Deep Learning is a data first science.

There is a wide gap between symbolic reasoning and deep learning. In this research, we explore the possibility of using deep learning to improve symbolic reasoning. Briefly, in a reasoning system, a deep feedforward neural network is used to guide rewriting processes after learning from algebraic reasoning examples produced by humans. To enable the neural network to recognise patterns of algebraic expressions with non-deterministic sizes, reduced partial trees are used to represent the expressions. Also, to represent both top-down and bottom-up information of the expressions, a centralisation technique is used to improve the reduced partial trees. Besides, symbolic association vectors and rule application records are used to improve the rewriting processes. Experimental results reveal that the algebraic reasoning examples can be accurately learnt only if the feedforward neural network has enough hidden layers. Also, the centralisation technique, the symbolic association vectors and the rule application records can reduce error rates of reasoning. In particular, the above approaches have led to 4.6% error rate of reasoning on a dataset of linear equations, differentials and integrals.

Using predictive models to identify patterns that can act as biomarkers for different neuropathoglogical conditions is becoming highly prevalent. In this paper, we consider the problem of Autism Spectrum Disorder (ASD) classification. While non-invasive imaging measurements, such as the rest state fMRI, are typically used in this problem, it can be beneficial to incorporate a wide variety of non-imaging features, including personal and socio-cultural traits, into predictive modeling. We propose to employ a graph-based approach for combining both types of feature, where a contextual graph encodes the traits of a larger population while the brain activity patterns are defined as a multivariate function at the nodes of the graph. Since the underlying graph dictates the performance of the resulting predictive models, we explore the use of different graph construction strategies. Furthermore, we develop a bootstrapped version of graph convolutional neural networks (G-CNNs) that utilizes an ensemble of weakly trained G-CNNs to avoid overfitting and also reduce the sensitivity of the models on the choice of graph construction. We demonstrate its effectiveness on the Autism Brain Imaging Data Exchange (ABIDE) dataset and show that the proposed approach outperforms state-of-the-art approaches for this problem.

Teaching machines to accomplish tasks by conversing naturally with humans is challenging. Currently, developing task-oriented dialogue systems requires creating multiple components and typically this involves either a large amount of handcrafting, or acquiring costly labelled datasets to solve a statistical learning problem for each component. In this work we introduce a neural network-based text-in, text-out end-to-end trainable goal-oriented dialogue system along with a new way of collecting dialogue data based on a novel pipe-lined Wizard-of-Oz framework. This approach allows us to develop dialogue systems easily and without making too many assumptions about the task at hand. The results show that the model can converse with human subjects naturally whilst helping them to accomplish tasks in a restaurant search domain.

In recent years there is a growing interest in using deep representations for reinforcement learning. In this paper, we present a methodology and tools to analyze Deep Q-networks (DQNs) in a non-blind matter. Moreover, we propose a new model, the Semi Aggregated Markov Decision Process (SAMDP), and an algorithm that learns it automatically. The SAMDP model allows us to identify spatio-temporal abstractions directly from features and may be used as a sub-goal detector in future work. Using our tools we reveal that the features learned by DQNs aggregate the state space in a hierarchical fashion, explaining its success. Moreover, we are able to understand and describe the policies learned by DQNs for three different Atari2600 games and suggest ways to interpret, debug and optimize deep neural networks in reinforcement learning.

When Sahil Singla joined the social impact startup Farmguide, he was shocked to discover that thousands of rural farmers in India commit suicide every year. When harvests go awry, desperate farmers are forced to borrow from microfinance loan sharks at crippling rates. Unable to pay back these predatory loans, victims kill themselves – often by grisly methods like swallowing pesticides – to escape the threats and violence of their ruthless debt collectors. Singla and his team are tackling this social injustice with one unexpected but powerful tool: deep learning. Recent growth of computational power and structured data sets has allowed deep learning algorithms to achieve extraordinary results.

One good way to frame the question of the limits of Deep Learning is in the context of the Principle of Computational Equivalence by Stephen Wolfram. Wolfram showed that simple cellular automation are able to exhibit complex behaviour that cannot be predicted from initial conditions or the simple rules that specify its incremental behaviour. Certain kinds of cellular automata can exhibit complex behaviour that cannot be reduced to a mathematical model that capture its behaviour in closed form. Wolfram examples of an'irreducible' system that exhibits this complex behaviour are the brain and weather systems. Wolfram classifies these kinds of systems as exhibiting "Universality".

Deep neural networks and Deep Learning are powerful and popular algorithms. And a lot of their success lays in the careful design of the neural network architecture. I wanted to revisit the history of neural network design in the last few years and in the context of Deep Learning. For a more in-depth analysis and comparison of all the networks reported here, please see our recent article. Reporting top-1 one-crop accuracy versus amount of operations required for a single forward pass in multiple popular neural network architectures. It is the year 1994, and this is one of the very first convolutional neural networks, and what propelled the field of Deep Learning.