Cross posted from my blog - I look forward to discussion/feedback here
Note: The paper below is best read as a pdf which you can download from the blog for free
This article is a part of an evolving theme. Here, I explain the basics of Deep Learning and how Deep learning algorithms could apply to IoT and Smart city domains. Specifically, as I discuss below, I am interested in complementing Deep learning algorithms using IoT datasets. I elaborate these ideas in the Data Science for Internet of Things program which enables you to work towards being a Data Scientist for the Internet of Things (modelled on the course I teach at Oxford University and UPM – Madrid). I will also present these ideas at the International conference on City Sciences at Tongji University in Shanghai and the Data Science for IoT workshop at the Iotworld event in San Francisco
Please connect with me if you want to stay in touch on linkedin and for future updates
Deep learning is often thought of as a set of algorithms that ‘mimics the brain’. A more accurate description would be an algorithm that ‘learns in layers’. Deep learning involves learning through layers which allows a computer to build a hierarchy of complex concepts out of simpler concepts.
The obscure world of deep learning algorithms came into public limelight when Google researchers fed 10 million random, unlabeled images from YouTube into their experimental Deep Learning system. They then instructed the system to recognize the basic elements of a picture and how these elements fit together. The system comprising 16,000 CPUs was able to identify images that shared similar characteristics (such as images of Cats). This canonical experiment showed the potential of Deep learning algorithms. Deep learning algorithms apply to many areas including Computer Vision, Image recognition, pattern recognition, speech recognition, behaviour recognition etc
To understand the significance of Deep Learning algorithms, it’s important to understand how Computers think and learn. Since the early days, researchers have attempted to create computers that think. Until recently, this effort has been rules based adopting a ‘top down’ approach. The Top-down approach involved writing enough rules for all possible circumstances. But this approach is obviously limited by the number of rules and by its finite rules base.
To overcome these limitations, a bottom-up approach was proposed. The idea here is to learn from experience. The experience was provided by ‘labelled data’. Labelled data is fed to a system and the system is trained based on the responses. This approach works for applications like Spam filtering. However, most data (pictures, video feeds, sounds, etc.) is not labelled and if it is, it’s not labelled well.
The other issue is in handling problem domains which are not finite. For example, the problem domain in chess is complex but finite because there are a finite number of primitives (32 chess pieces) and a finite set of allowable actions(on 64 squares). But in real life, at any instant, we have potentially a large number or infinite alternatives. The problem domain is thus very large.
A problem like playing chess can be ‘described’ to a computer by a set of formal rules. In contrast, many real world problems are easily understood by people (intuitive) but not easy to describe (represent) to a Computer (unlike Chess). Examples of such intuitive problems include recognizing words or faces in an image. Such problems are hard to describe to a Computer because the problem domain is not finite. Thus, the problem description suffers from the curse of dimensionality i.e. when the number of dimensions increase, the volume of the space increases so fast that the available data becomes sparse. Computers cannot be trained on sparse data. Such scenarios are not easy to describe because there is not enough data to adequately represent combinations represented by the dimensions. Nevertheless, such ‘infinite choice’ problems are common in daily life.
Deep learning is involved with ‘hard/intuitive’ problem which have little/no rules and high dimensionality. Here, the system must learn to cope with unforeseen circumstances without knowing the Rules in advance. Many existing systems like Siri’s speech recognition and Facebook’s face recognition work on these principles. Deep learning systems are possible to implement now because of three reasons: High CPU power, Better Algorithms and the availability of more data. Over the next few years, these factors will lead to more applications of Deep learning systems.
Deep Learning algorithms are modelled on the workings of the Brain. The Brain may be thought of as a massively parallel analog computer which contains about 10^10 simple processors (neurons) – each of which require a few milliseconds to respond to input. To model the workings of the brain, in theory, each neuron could be designed as a small electronic device which has a transfer function similar to a biological neuron. We could then connect each neuron to many other neurons to imitate the workings of the Brain. In practise, it turns out that this model is not easy to implement and is difficult to train.
So, we make some simplifications in the model mimicking the brain. The resultant neural network is called “feed-forward back-propagation network”. The simplifications/constraints are: We change the connectivity between the neurons so that they are in distinct layers. Each neuron in one layer is connected to every neuron in the next layer. Signals flow in only one direction. And finally, we simplify the neuron design to ‘fire’ based on simple, weight driven inputs from other neurons. Such a simplified network (feed-forward neural network model) is more practical to build and use.
a) Each neuron receives a signal from the neurons in the previous layer
b) Each of those signals is multiplied by a weight value.
c) The weighted inputs are summed, and passed through a limiting function which scales the output to a fixed range of values.
d) The output of the limiter is then broadcast to all of the neurons in the next layer.
Image and parts of description in this section adapted from : Seattle robotics site
The most common learning algorithm for artificial neural networks is called Back Propagation (BP) which stands for “backward propagation of errors”. To use the neural network, we apply the input values to the first layer, allow the signals to propagate through the network and read the output. A BP network learns by example i.e. we must provide a learning set that consists of some input examples and the known correct output for each case. So, we use these input-output examples to show the network what type of behaviour is expected. The BP algorithm allows the network to adapt by adjusting the weights by propagating the error value backwards through the network. Each link between neurons has a unique weighting value. The ‘intelligence’ of the network lies in the values of the weights. With each iteration of the errors flowing backwards, the weights are adjusted. The whole process is repeated for each of the example cases. Thus, to detect an Object, Programmers would train a neural network by rapidly sending across many digitized versions of data (for example, images) containing those objects. If the network did not accurately recognize a particular pattern, the weights would be adjusted. The eventual goal of this training is to get the network to consistently recognize the patterns that we recognize (ex Cats).
The whole objective of Deep Learning is to solve ‘intuitive’ problems i.e. problems characterized by High dimensionality and no rules. The above mechanism demonstrates a supervised learning algorithm based on a limited modelling of Neurons – but we need to understand more.
Deep learning allows computers to solve intuitive problems because:
This is similar to the way a child learns ‘what a dog is’ i.e. by understanding the sub-components of a concept ex the behavior(barking), shape of the head, the tail, the fur etc and then putting these concepts in one bigger idea i.e. the Dog itself.
The (knowledge) representation problem is a recurring theme in Computer Science.
Knowledge representation incorporates theories from psychology which look to understand how humans solve problems and represent knowledge. The idea is that: if like humans, Computers were to gather knowledge from experience, it avoids the need for human operators to formally specify all of the knowledge that the computer needs to solve a problem.
For a computer, the choice of representation has an enormous effect on the performance of machine learning algorithms. For example, based on the sound pitch, it is possible to know if the speaker is a man, woman or child. However, for many applications, it is not easy to know what set of features represent the information accurately. For example, to detect pictures of cars in images, a wheel may be circular in shape – but actual pictures of wheels may have variants (spokes, metal parts etc). So, the idea of representation learning is to find both the mapping and the representation.
If we can find representations and their mappings automatically (i.e. without human intervention), we have a flexible design to solve intuitive problems. We can adapt to new tasks and we can even infer new insights without observation. For example, based on the pitch of the sound – we can infer an accent and hence a nationality. The mechanism is self learning. Deep learning applications are best suited for situations which involve large amounts of data and complex relationships between different parameters. Training a Neural network involves repeatedly showing it that: “Given an input, this is the correct output”. If this is done enough times, a sufficiently trained network will mimic the function you are simulating. It will also ignore inputs that are irrelevant to the solution. Conversely, it will fail to converge on a solution if you leave out critical inputs. This model can be applied to many scenarios as we see below in a simplified example.
Deep learning involves learning through layers which allows a computer to build a hierarchy of complex concepts out of simpler concepts. This approach works for subjective and intuitive problems which are difficult to articulate.
Consider image data. Computers cannot understand the meaning of a collection of pixels. Mappings from a collection of pixels to a complex Object are complicated.
With deep learning, the problem is broken down into a series of hierarchical mappings – with each mapping described by a specific layer.
The input (representing the variables we actually observe) is presented at the visible layer. Then a series of hidden layers extracts increasingly abstract features from the input with each layer concerned with a specific mapping. However, note that this process is not pre defined i.e. we do not specify what the layers select
For example: From the pixels, the first hidden layer identifies the edges
From the edges, the second hidden layer identifies the corners and contours
From the corners and contours, the third hidden layer identifies the parts of objects
Finally, from the parts of objects, the fourth hidden layer identifies whole objects
Image and example source: Yoshua Bengio book – Deep Learning
In addition, we have limitations in the technology. For instance, we have a long way to go before a Deep learning system can figure out that you are sad because your cat died(although it seems Cognitoys based on IBM watson is heading in that direction). The current focus is more on identifying photos, guessing the age from photos(based on Microsoft’s project Oxford API)
And we have indeed a way to go as Andrew Ng reminds us to think of Artificial Intelligence as buildin...
“I think AI is akin to building a rocket ship. You need a huge engine and a lot of fuel. If you have a large engine and a tiny amount of fuel, you won’t make it to orbit. If you have a tiny engine and a ton of fuel, you can’t even lift off. To build a rocket you need a huge engine and a lot of fuel. The analogy to deep learning [one of the key processes in creating artificial intelligence] is that the rocket engine is the deep learning models and the fuel is the huge amounts of data we can feed to these algorithms.”
Today, we are still limited by technology from achieving scale. Google’s neural network that identified cats had 16,000 nodes. In contrast, a human brain has an estimated 100 billion neurons!
There are some scenarios where Back propagation neural networks are suited
Given an IoT domain, we could consider the top-level questions:
Now, extending more deeply into the research domain, here are some areas of interest that I am following.
In essence, these techniques/strategies complement Deep learning algorithms with IoT datasets.
1) Deep learning algorithms and Time series data : Time series data (coming from sensors) can be thought of as a 1D grid taking samples at regular time intervals, and image data can be thought of as a 2D grid of pixels. This allows us to model Time series data with Deep learning algorithms (most sensor / IoT data is time series). It is relatively less common to explore Deep learning and Time series – but there are some instances of this approach already (Deep Learning for Time Series Modelling to predict energy loads usi... )
2) Multiple modalities: multimodality in deep learning. Multimodality in deep learning algorithms is being explored In particular, cross modality feature learning, where better features for one modality (e.g., video) can be learned if multiple modalities (e.g., audio and video) are present at feature learning time
3) Temporal patterns in Deep learning: In their recent paper, Ph.D. student Huan-Kai Peng and Professor Radu Marcul... from Carnegie Mellon University’s Department of Electrical and Computer Engineering, propose a new way to identify the intrinsic dynamics of interaction patterns at multiple time scales. Their method involves building a deep-learning model that consists of multiple levels; each level captures the relevant patterns of a specific temporal scale. The newly proposed model can be also used to explain the possible ways in which short-term patterns relate to the long-term patterns. For example, it becomes possible to describe how a long-term pattern in Twitter can be sustained and enhanced by a sequence of short-term patterns, including characteristics like popularity, stickiness, contagiousness, and interactivity. The paper can be downloaded HERE
I see Smart cities as an application domain for Internet of Things. Many definitions exist for Smart cities/future cities. From our perspective, Smart cities refer to the use of digital technologies to enhance performance and wellbeing, to reduce costs and resource consumption, and to engage more effectively and actively with its citizens (adapted from Wikipedia). Key ‘smart’ sectors include transport, energy, health care, water and waste. A more comprehensive list of Smart City/IoT application areas are: Intelligent transport systems – Automatic vehicle , Medical and Healthcare, Environment , Waste management , Air quality , Water quality, Accident and Emergency services, Energy including renewable, Intelligent transport systems including autonomous vehicles. In all these areas we could find applications to which we could add an intuitive component based on the ideas above.
Typical domains will include Computer Vision, Image recognition, pattern recognition, speech recognition, behaviour recognition. Of special interest are new areas such as the Self driving cars – ex theLutz pod and even larger vehicles such as self driving trucks
Deep learning involves learning through layers which allows a computer to build a hierarchy of complex concepts out of simpler concepts. Deep learning is used to address intuitive applications with high dimensionality. It is an emerging field and over the next few years, due to advances in technology, we are likely to see many more applications in the Deep learning space. I am specifically interested in how IoT datasets can be used to complement deep learning algorithms. This is an emerging area with some examples shown above. I believe that it will have widespread applications, many of which we have not fully explored(as in the Smart city examples)
I see this article as part of an evolving theme. Future updates will explore how Deep learning algorithms could apply to IoT and Smart city domains. Also, I am interested in complementing Deep learning algorithms using IoT datasets.
I elaborate these ideas in the Data Science for Internet of Things program (modelled on the course I teach at Oxford University and UPM – Madrid). I will also present these ideas at the International conference on City Sciences at Tongji University ... and the Data Science for IoT workshop at the Iotworld event in San Francisco
Please connect with me if you want to stay in touch on linkedin and for future updates